sample1.txt

AI Principles
Progress Update
2023Table of contents
Preface: Google’s AI Principles 03
Introduction 06
Internal governance & risk management 08
Resources, research, tools & responsible practices 18
Product impact 27
Supporting global dialogue, standards & policy 36
Conclusion 38
Appendix 41
2Preface
Google’s AI Principles:
Objectives for AI applications
1. Be socially beneficial.
The expanded reach of new technologies increasingly touches society as a whole. Advances
in AI will have transformative impacts in a wide range of fields, including healthcare, security,
energy, transportation, manufacturing, and entertainment. As we consider potential
development and use of AI technologies, we will take into account a broad range of social and
economic factors, and will proceed where we believe that the overall likely benefits substantially
exceed the foreseeable risks and downsides.
AI also enhances our ability to understand the meaning of content at scale. We will strive
to make high-quality and accurate information readily available using AI, while continuing
to respect cultural, social, and legal norms in the countries or regions where we operate.
And we will continue to thoughtfully evaluate when to make our technologies available on
a non-commercial basis.
2. Avoid creating or reinforcing unfair bias.
AI algorithms and datasets can reflect, reinforce, or reduce unfair biases. We recognize that
distinguishing fair from unfair biases is not always simple, and differs across cultures and
societies. We will seek to avoid unjust impacts on people, particularly those related to sensitive
characteristics such as race, ethnicity, gender, nationality, income, sexual orientation, ability,
and political or religious belief.
3. Be built & tested for safety.
We will continue to develop and apply strong safety and security practices to avoid unintended
results that create risks of harm. We will design our AI systems to be appropriately cautious, and
seek to develop them in accordance with best practices in AI safety research. In appropriate
cases, we will test AI technologies in constrained environments and monitor their operation
after deployment.
34. Be accountable to people.
We will design AI systems that provide appropriate opportunities for feedback, relevant
explanations, and appeal. Our AI technologies will be subject to appropriate human direction
and control.
5. Incorporate privacy design principles.
We will incorporate our privacy principles in the development and use of our AI technologies.
We will give opportunity for notice and consent, encourage architectures with privacy
safeguards, and provide appropriate transparency and control over the use of data.
6. Uphold high standards of scientific excellence.
Technological innovation is rooted in the scientific method and a commitment to open inquiry,
intellectual rigor, integrity, and collaboration. AI tools have the potential to unlock new realms
of scientific research and knowledge in critical domains like biology, chemistry, medicine, and
environmental sciences. We aspire to high standards of scientific excellence as we work to
progress AI development.
We will work with a range of stakeholders to promote thoughtful leadership in this area,
drawing on scientifically rigorous and multidisciplinary approaches. And we will responsibly
share AI knowledge by publishing educational materials, best practices, and research that
enable more people to develop useful AI applications.
7. Be made available for uses that accord with these principles.
Many technologies have multiple uses. We will work to limit potentially harmful or abusive
applications. As we develop and deploy AI technologies, we will evaluate likely uses in light
of the following factors:
• Primary purpose and use: the primary purpose and likely use of a technology and
application, including how closely the solution is related to or adaptable to a harmful use
• Nature and uniqueness: whether we are making available technology that is unique
or more generally available
• Scale: whether the use of this technology will have significant impact
• Nature of Google’s involvement: whether we are providing general-purpose tools,
integrating tools for customers, or developing custom solutions
4AI applications we will not pursue
In addition to the above objectives, we will not design or deploy AI in the following
application areas:
1. Technologies that cause or are likely to cause overall harm. Where there is
a material risk of harm, we will proceed only where we believe that the benefits
substantially outweigh the risks, and will incorporate appropriate safety constraints.
2. Weapons or other technologies whose principal purpose or implementation
is to cause or directly facilitate injury to people.
3. Technologies that gather or use information for surveillance violating internationally
accepted norms.
4. Technologies whose purpose contravenes widely accepted principles of international
law and human rights.
As our experience in this space deepens, this list may evolve.
5Introduction
This is the 5th edition of our annual AI Principles progress report, where we provide consistent
transparency into how we put our principles into practice. We first published the AI Principles
in 2018 to share the company’s technology ethics charter and hold ourselves accountable
for how we research and develop AI responsibly. Generative AI is no exception. In this report,
we share details of the principled approach used throughout the research and development
lifecycle for our novel generative AI models, including the Gemini family of models.
Principles are only effective once put into practice. This is why we offer this annual report —
including tough lessons learned — to enable others across the AI ecosystem to learn from
our experience.
For Google and throughout the industry, this year marks a turning point for AI both as a
research discipline and a commercial technology. Exciting new generative AI applications are
writing poetry and computer code. Advanced AI applications can help diagnose diseases with
accuracy and help communities around the world address the effects of climate change, from
wildfires to flooding. At the same time, 2023 marks a milestone moment in the young history
of global AI governance. In July, we joined other industry peers in making voluntary industry
commitments for safe, secure, and trustworthy AI at the White House. This was followed in
October by the latest Executive Order, which is focused on new standards for AI safety and
security, and managing AI risks. Toward the end of this year, the G7 released an international
code of conduct for responsible AI. The United Nations announced an AI advisory group and
the UK held an international summit on AI safety. And, in December, policy makers in the
European Union reached a preliminary political agreement on the AI Act, the first law
to regulate AI.
As a result, from these actions and many others in 2023, we can see the beginnings of an
international, shared framework for responsible AI innovation taking shape. This occurs
alongside frameworks and standards on AI risks and mitigations from organizations such as
the US National Institute of Standards and Technology (NIST), Organization for Economic Co-
operation and Development (OECD), and Organization for Standardization (ISO). In addition,
governing efforts are underway in nations such as Singapore, Brazil, Canada, and India.
6At the same time, non-governmental organizations like the Partnership on AI, ML Commons,
and the Frontier Model Forum are also sharing best practices and helping to advance the state
of the art in AI evaluations, benchmarking, and safety testing. And multi-stakeholder initiatives
like the World Economic Forum AI Governance Alliance are helping to encourage responsible
releases of transparent and inclusive AI systems.
Promoting alignment on industry best practices is imperative for building advanced AI
applications that have social benefit, avoid unfair bias, are built and tested for safety and
privacy, and are accountable to people. The dawn of generative AI offers an opportunity for
us to guide the development of an unprecedented technology with principled practices.
Since we first published our AI Principles in 2018, we’ve centered our internal AI governance
and operations efforts in four key areas:
1. Culture and education: Employee training, resources, and workshops
on the ethical development of AI
2. Structures and processes: Risk assessments and AI Principles reviews
3. Tools, techniques, and infrastructure: Technical solutions and resources, such
as responsible AI safety filters and classifiers, model and data cards, built-in techniques
such as fine-tuning and reinforcement learning, and automated adversarial testing
4. External engagement and partnerships: Collaboration with industry peers and civil
society and efforts across the external AI ecosystem, including with academia, start-
ups, and governments
We’re committed to thoughtful iteration and to constantly sharing and learning, within
our industry and across the greater society, in order to build AI that benefits everyone.
7Internal governance & risk
management
As Google increasingly incorporates AI into all of our products and services, we are
increasingly integrating our AI review work into our holistic Enterprise Risk Management
frameworks for assuring the quality of our offerings. This evolution helps us further the scale
of our work and integration into existing governance and company-wide infrastructure and
accountability processes.
Google’s enterprise risk frameworks, tools, and systems of record provide a foundation
for first-line reviews of AI-related issues, and help assure compliance with evolving legal,
regulatory, and standards benchmarks. This approach will help us fulfill new directives such
as the US White House’s Executive Order on AI, the G7’s International Guiding Principles for
Organizations Developing Advanced AI Systems, and the AI Act in the EU.
Our AI governance teams collaborate closely with teams and subject matter experts across
machine learning (ML) research, product policy, user-experience research and design, public
policy, law, human rights, and the social sciences, among many other disciplines. For many
years we have been on a journey of formalizing, expanding, and institutionalizing our machine-
learning and artificial-intelligence reviews across a growing range of products and services.
In close coordination with central teams, some of our product areas have developed their own
specialized review processes, deploying approaches tailored to their unique circumstances.
For example, Google Cloud’s Responsible AI team helps enterprises develop effective AI
safety and responsibility risk management strategies, through conversations and shared best
practices with customers.
Google Cloud deploys a shared fate model, in which select customers are provided with
tools — such as those like SynthID for watermarking images generated by AI. Customers test
the tools in line with their own AI principles or other responsible innovation frameworks. This
shared fate model offers a closer interaction with customers, including tailoring practices
and tooling to their needs and risk management strategies. As we continue to develop our
AI platforms, systems, and foundational models, Cloud will continue to invest in end-to-end
governance tools and guidance on best practices to help our customers keep their data
and AI models safe.
8This year, Cloud’s AI products and services for enterprises expanded to include additional
security solutions with Security AI Workbench, an industry-leading platform of tools
(Mandiant Threat Intelligence, Chronicle Security Operations, and Security Command Center);
governance and compliance controls for AI workloads, built on Vertex AI; and security-focused
AI collaboration and assistance with Duet AI.
To provide a more comprehensive approach to safe, secure, and trustworthy AI development
across products, we’re working to integrate and expand many of our internal AI Principles
operations efforts across different functions. Generative AI raises new issues, such as the
potential for model misinterpretations of data (commonly referred to as “hallucinations”).
As we continue to integrate generative AI into more products and features, our teams
leverage decades of experience and take a comprehensive approach to better anticipate
and test for potential new risks. We continue to have senior-management oversight of both
new and emerging issues in AI and compliance with evolving standards and practices.
As we continue to integrate generative AI into more products
and features, our teams leverage decades of experience and
take a comprehensive approach to better anticipate and test
for potential new risks.
These reviews often require consideration of the trade-offs between ethical risks of certain
new applications and potential social benefits. For example, in the case of generated
photorealistic images of people, we discussed the risks of deepfakes and misinformation
versus the social benefits of enabling small businesses and creators to make high-quality
content to grow their businesses and contribute to their communities. We agreed on an
approach that seeks to make generative AI image technology available, subject to strict
testing and clear guardrails (like the use of safety classifiers and filters).
Evolving generative AI pre-launch ethics reviews
Our AI Principles ethics reviews and impact assessments are part of a larger, end-to-end
pre-launch process that includes technical safety testing and standard privacy and security
reviews. The AI Principles review process offers tailored guidance for applying the principles
as a practical framework for the development of new products and services.
9This year, we more than doubled our AI Principles reviews (to more than 500) with most
focused on the implementation of generative AI research models into products, services,
and features. To accommodate the increasing numbers of generative AI reviews and scale
AI Principles assurance, our cross-company pre-launch process assesses early product
designs against known legal requirements, emerging legislation, standards, and our AI
Principles. Teams may address identified issues through technical or policy mitigations or
guardrails, such as additional safety filters or continued model refinement. Product teams
continue to adopt other best practices for responsible AI research and development
throughout the launch and operations processes.
A risk-based approach to generative AI
Our risk assessment framework seeks to identify, measure, and analyze risks throughout the
product development lifecycle. AI Principles reviews map these risks to appropriate mitigations
and interventions, drawing upon our best practices from our cross-company enterprise risk
management efforts.
We conduct AI Principles reviews for all generative AI projects, with particular focus on certain
areas. These include inherently large scale applications in domains such as:
• Government-related
• Recommendation, personalization, and ranking systems
• Critical technology infrastructure
• Environmental sustainability
• Social impact
• Health, fitness, and well-being
• Finance, education, and employment
• Surveillance and/or biometrics
• Ambient computing, affective technology, and wearables
10AI Principles reviews assess a range of harms, taking into account impacts ranging from
unfair biases and stereotypes, poor product experiences, and social harms such as the spread
of misinformation. In addition, as we’ve reported in detail in our 2022 AI Principles Progress
Update, we engage external experts to conduct human rights impact assessments
as appropriate.
We also draw on feedback from more than 1,000 Googlers around the world who represent
the international diversity of the people who use our products, with more than 50% living and
working outside of the US. They represent 39 different countries and regions and speak more
than 85 different languages. This feedback is shared with teams working to automate more
of our adversarial testing.
Policies & practices for responsible generative
AI development
To guide product teams internally, we’ve established a framework to define the types
of harmful content that we do not permit our models to generate. It also guides how we
protect personal identifiable information (such as Social Security Numbers). We leveraged
our experience launching conversational products like Google Assistant and content features
such as featured snippets in Search to understand how to minimize offensive and
low-quality answers.
This framework — which serves as a standardized policy recommendation for all generative
AI products and modalities — also reflects our commitment to product inclusion and equity.
Based on Google’s extensive experience with harm mitigation and rigorous research, and
reflecting our established approach to product safety, our policy says that generative AI
products must not create harmful content, such as child sexual abuse and exploitation, hate
speech, harassment, violence and gore, or obscenity and profanity; dangerous content that
facilitates, promotes, or enables access to harmful goods, services, and activities;
or malicious content, such as spam or phishing. The framework also targets the harms
caused by misinformation or unfair bias, with guidelines focused on providing neutral
answers grounded in authoritative, consensus facts, or providing multiple perspectives.
11As with all of our product policies, we aim to regularly review and update this generative AI
framework to respond to emerging safety enforcement trends, new product features, and
new ways products are used — to protect against misuse.
We conduct adversarial testing and red teaming, or “ethical hacking,” of our products to
test for policy violations and to measure how well a model is following the policy framework.
While we generally expect our generative AI products to restrict the content set out in the
framework, there are some important exceptions. Similar to other Google products — for
example, featured snippets on Search — we make an exception when there is an educational,
documentary, scientific, or artistic benefit to showing or translating content that might
otherwise be perceived as offensive within these specific, beneficial contexts, as we do
within the Bard experience.
Adversarial testing is just one of three essential practices for building responsible generative AI
that we shared publicly this year, based on trends and patterns we observed in hundreds of AI
Principles reviews conducted in 2023:
1. Design for responsibility
2. Conduct adversarial testing
3. Communicate simple, helpful explanations
Our first essential practice, designing for responsible generative AI, is a proactive approach
that begins by first identifying and documenting potential harms (for example, unfair bias
in AI model outputs within a product, which could lead to toxic content or loss of economic
opportunity for specific groups of people). These harms can then be mitigated with the use
of responsible datasets, classifiers and filters, and in-model mitigations such as fine tuning,
reasoning, few-shot prompting, data augmentation, and controlled decoding to address
potential harms proactively.
Our second essential practice, adversarial testing, refers to systematic evaluation of a model
by providing malicious or inadvertently harmful inputs across a range of scenarios to identify
and mitigate potential safety and fairness risks. We conduct this testing before major model
and product launches, including our Gemini family of models (see the technical paper
for details).
12For Bard, which lets people collaborate with generative AI through conversational prompts,
we conducted testing to identify situations where the model could be mistakenly perceived
as human. Such anthropomorphization can lead to potentially harmful misunderstandings.
To intervene, we limit Bard’s self-reference to personal pronouns, human identity, and claims
of implicit or explicit humanness. We are continuing to conduct research into this domain to
develop our approach to managing anthropomorphization identified in testing.
We continue to experiment with new forms of adversarial testing. For example, we hosted
an internal, company-wide large language model (LLM) red teaming “Hack-AI-thon” with
hundreds of security, safety, and other experts.
We conduct adversarial testing and red teaming, or “ethical
hacking,” of our products to test for policy violations and to
measure how well a model is following the policy framework.
In addition to adversarial testing for safety and fairness, we’ve also established a dedicated
Google AI Red Team focused on testing AI models and products for security, privacy and
abuse risks. Externally, we participated in the White House-sponsored red teaming event
at DEFCON, which drew over 2,000 people to test industry-leading LLMs in an effort to better
understand risks and limitations of these advanced technologies. We also continue to innovate
with methods for scaled automated testing using LLM-based auto-raters to enable efficiency
and scaling.
Our third essential practice, communicating simple, helpful explanations, requires:
1. Making it clear to users when and how generative AI is used
2. Showing how people can offer feedback, and
3. Showing how people are in control as they use an AI-powered product or service.
Maintaining transparency documentation for developers, governments, and policy leaders
is also key. This can mean releasing detailed technical reports or model or data cards that
appropriately make public essential information based on our internal documentation of safety
and other model evaluation details. These transparency artifacts are more than communication
vehicles; they can offer guidance for AI researchers, deployers, and downstream developers
on the responsible use of the model.
13To build upon these practices, we provide self-service guides and continue to catalog
patterns of generative AI risks and common interventions and mitigations. These include
common risks known across the industry, such as hallucinations, for which we apply
mitigations such as technical tooling for identifying AI-generated content, a prohibited use
policy, clear explanations of the risk of hallucination, and feedback mechanisms to report
concerns such as potentially harmful outputs. Other common generative AI risks include
model outputs that reflect or reinforce unfair biases or outputs that are extremely similar
to or indistinguishable from those created by humans, which can lead to misunderstandings
such as perceived sentience.
We have internal guides to help product and research teams across Google better understand
and proactively mitigate these risks.
Common generative AI interventions
14By sharing the common risks that we find in our AI Principles reviews, we can offer
transparency into our emerging best practices to mitigate these risks. These range from
the technical, such as SynthID or About this image, tools we developed this year that can
help identify mis- and dis-information when generative AI tools are used by malicious actors,
to explainability techniques such as increasing explanatory information throughout the AI
product, not just at the moment of decision.
And we continue to conduct foundational research to gain additional insight on these risks.
For example, we recently worked with Gallup, Inc. to survey perceptions and attitudes around
technology to gain insights into how anthropomorphism influences people’s use of generative
AI chatbots and other technology. Such insights help us understand potential benefits and
dangers of humanizing technology and the development of new interventions, mitigations,
and guardrails to help people use AI appropriately.
We’re committed to reporting specific capabilities, limitations,
risks, and mitigations we’ve applied into our generative
AI-powered systems, and contributing to shared industry
standards on model transparency.
We’ve also begun to look ahead and expand our threat research to assess large models’
cyber capabilities, which can lead to potential cyber weapons used by adversaries. We’re
also researching the security benefits and risks of our largest model in the Gemini family of
generative models. This has included scoping new evaluation techniques, as well as joining
relevant external fora, such as the UK’s new Biosecurity Leadership Council.
Generative AI is a nascent technology, so there are many risks yet to be discovered and
defined – as well as benefits. For example, generative AI can be used to help identify and track
harmful, fake information, even that of which is generated by AI. We’re committed to reporting
specific capabilities, limitations, risks, and mitigations we’ve applied into our generative AI-
powered systems, and contributing to shared industry standards on model transparency.
This year, we’re piloting a transparency artifact specifically for the integration of research
generative AI models into AI-powered systems. This artifact is called a generative AI system
card. It builds upon our work of designing widely referenced and adopted transparency
artifacts such as model and data cards.
15Our first version is intended to provide structured, easy-to-find information for non-technical
audiences ranging from third-party auditors and policy makers to journalists, enterprise clients,
and clients and advertisers, as well as users. The cards offer an overview of the capabilities
and limitations of a generative AI model as integrated into a larger system that people interact
with as a product or service. (See appendix for an example, documenting the December 2023
update of Bard with specifically tuned Gemini Pro).
Equipping employees to practice the
AI Principles
We broadly share knowledge among our employees on how to execute upon our responsible
practices and policies via a frequently updated AI Principles hub, featuring current product
policies and guidance, along with self-service content and training. Usage of this hub has more
than doubled since last year.
Given the rapidly evolving nature of generative AI and emerging best practices, this year we
launched virtual AI Principles boot camps open to any and all Googlers. These boot camps
include interactive sessions in which participants test their knowledge of the AI Principles and
engage in mock ethics reviews of AI products.
Other educational offerings for employees include an expansion of our live interactive Moral
Imagination Workshops, which involve deep engagement in philosophical approaches to
product development scenarios. The number of product teams engaging in Moral Imagination
sessions has more than doubled since they launched in 2021. The workshop was presented
externally at the Affective Computing + Intelligent Interaction conference in the fall of 2023.
Elements of the workshop will be integrated into onboarding training for senior hires beginning
in the first quarter of 2024.
Also this year, building on the Responsible Innovation Challenge, a game-like exercise that
tested employees’ recall of the AI Principles and has been completed by more than 20,000
Googlers, we designed and launched a new internal game-like AI ethics training experience.
The training encourages technical Googlers to focus on best practices for building AI products
responsibly, including how to document safe and unsafe practices, testing AI model outputs
for fair outcomes, and filing bugs if improvement is needed. Approximately 1,800 Googlers
have completed this new course.
16We’re committed to sharing our practices externally as well. This year, we’ve launched
educational, hands-on resources that reflect key concepts in our internal educational
resources. These include Introduction to Responsible AI for developers, and Technomoral
Scenarios for Responsible Innovation for industry professionals.
17Resources, research, tools &
responsible practices
We invest in ongoing research into Responsible AI development. Our online database of more
than 200 publications since 2012 serves as a resource for the research community and the
larger AI ecosystem.
We continue to develop new techniques to advance our ability to discover unknown failures,
explain model behaviors, and improve model output through training, responsible generation,
and failure mitigation.
However, understanding and mitigating generative AI safety risks is both a technical and
social challenge. Safety perceptions are intrinsically subjective and influenced by a wide
range of intersecting factors. Our study on how demographic characteristics influence safety
perceptions explored the effects of rater demographics (such as gender and age) and content
characteristics (such as degree of harm) on safety assessments of generative AI outputs. Our
disagreement analysis framework highlighted a variety of disagreement patterns between
raters from different backgrounds, including “ground truth” expert ratings. Our NeurIPS 2023
publication introduced the DICES (Diversity In Conversational AI Evaluation for Safety) dataset
to facilitate nuanced safety evaluations of large language models, accounting for cultural
variance, ambiguity, and diversity.
We continue to pursue research into using societal context knowledge to foster responsible AI.
This year, we piloted a tool to convert system dynamics models of complex societal problems
into reinforcement-learning environments, opening up the ability for AI to be more socially
beneficial through deep problem understanding, and released a more comprehensive identity
lexicon, TIDAL.
18Techniques & datasets to help avoid unfair bias
A key part of our ML work involves developing techniques to build models that are more
inclusive. Informed by sociology and social psychology, we focus on working toward scalable
solutions that enable nuanced measurement and mitigation in areas such as studying the
differences in human perception and annotation of skin tone in images using the Monk
Skin Tone scale.
We’re developing methodologies to build models for people from a diversity of backgrounds.
For example, our exploration of the design of participatory systems allows individuals to
choose whether to disclose sensitive attributes with explicit consent when an AI system
makes predictions. This approach suggests a way to reconcile the challenging tension
between avoiding unfair bias and applying privacy design.
We’ve also strengthened our community-based research efforts, focusing on historically
marginalized communities or groups of people who may experience unfair outcomes of AI.
This ranged from evaluations of gender-inclusive health to mitigate harms for people with
queer and non-binary identities, to explorations on how to scale automatic speech recognition
by using a large unlabeled multilingual dataset to pre-train and fine-tune a model to recognize
under-represented languages and adapt to new languages and data.
We’ve made the Monk Skin Tone Examples (MST-E) dataset publicly available to enable AI
practitioners everywhere to create more consistent, inclusive, and meaningful skin tone
annotations as they create computer vision products that work well for all skin tones. It
contains 1,515 images and 31 videos of 19 subjects spanning the 10 point Monk Skin Tone (MST)
scale, where the subjects and images were sourced through TONL, a stock photography
company focusing on diversity. The 19 subjects include individuals of different ethnicities and
gender identities to help human annotators decouple the concept of skin tone from perceived
race. The primary goal of this dataset is to enable practitioners to train their human annotators
and test for consistent skin tone annotations across various environment capture conditions.
Since we launched the MST scale last year, we’ve been using it to improve Google’s computer
vision systems to make equitable image tools for everyone and to improve representation of
skin tone in Search. Computer vision researchers and practitioners outside of Google, like
the curators of Meta’s Casual Conversations dataset, have also recognized the value of MST
annotations to provide additional insight into diversity and representation in datasets.
19Because AI models are often trained and evaluated on human-annotated data, we also
advance human-centric research on data annotation. We have developed methods to
account for rater diversity, and in the recent past, we’ve shared responsible practices for data
enrichment sourcing. These methods enable AI practitioners to better ensure diversity in
annotation of datasets used to train models, by identifying current barriers and
re-envisioning data work practices.
This year, we sought to create new, inclusive datasets as well. For example, Project Elevate
Black Voices (EBV) is a first-of-its-kind collaboration between Responsible AI UX, Speech,
and Assistant to responsibly collect and transcribe a dataset of African American English
in partnership with Howard University and other Historically Black Colleges and Universities
to reduce racial disparities in automatic speech recognition and improve our overall
speech model.
Human-centered AI research
Our researchers explore generative AI within the lens of human-centered topics, from using
language models to create generative agents to an exploratory study with five designers
(presented at the CHI conference) that looks at how people with no machine learning
programming experience or training can use prompt programming to quickly prototype
functional user interface mock-ups. This prototyping speed can help enable user research
sooner in the product design process.
The growth of generative large language models has also opened up new techniques to solve
important long-standing problems. Agile classifiers are one research approach we’re taking to
solve classification problems related to better online discourse, such as nimbly blocking newer
types of toxic language. The big advance here is the ability to develop high-quality classifiers
from very small datasets — as small as 80 examples. This suggests a positive
future for online discourse and better moderation of it.
Now, instead of collecting millions of examples to attempt to create universal safety classifiers
for all use cases over months or years, more agile classifiers might be created by individuals or
small organizations and tailored for their specific use cases, and then iterated on and adapted
in the time-span of a day (such as to block a new kind of harassment being received or to
correct unintended biases in models). As an example of their utility, these methods recently
won a SemEval competition to identify and explain sexism.
20We’ve also developed new state-of-the-art explainability methods to identify the role of
training data on model behaviors and misbehaviors. By combining training data attribution
methods with agile classifiers, we found that we can identify mislabelled training examples.
This makes it possible to reduce the noise in training data, leading to significant improvements
on model accuracy.
Collectively, these methods are critical to help the scientific community improve generative
models. They provide techniques for fast and effective content moderation and dialogue
safety methods that help support creators whose content is the basis for generative models’
amazing outcomes. In addition, they provide direct tools to help debug model misbehavior,
which leads to better generation.
A systematic research approach to safety
The unprecedented capabilities of generative AI models are accompanied by new challenges
including hallucination (model output that contains factual inaccuracies). To that end,
our safety research has focused on three directions:
1. Scaled adversarial data generation
We create test sets containing potentially unsafe model inputs that stress the model
capabilities under adverse circumstances. We focus on identifying societal harms
to the diversity of user communities impacted by our models.
2. Automated test set evaluation and community engagement
We scale the testing process with automated test set evaluation to offer many
thousands of model responses and quickly evaluate how the model responds across a
wide range of potentially harmful scenarios. We also participate in external community
engagement to identify “unknown unknowns” and to seed the data generation process.
3. Rater diversity
Safety evaluations rely on human judgment, which is shaped by community and culture
and is not easily automated. To address this, we prioritize research on rater diversity.
21To provide the high-quality human input required to seed the scaled processes, we partner
with groups such as the Equitable AI Research Round Table (EARR), and with our internal ethics
teams to ensure that we are representing the diversity of communities who use our models.
We continue to expand our reach in terms of collaborating with underrepresented groups; for
example, researchers are currently exploring collaborative AI development projects with the US
federally-recognized Fort Peck Tribes (the Assiniboine and Sioux Tribes), such as developing a
Siouan language model together.
The Adversarial Nibbler Challenge also engages external users to understand potential harms
of unsafe, biased, or violent outputs to end users. We’re committed to a global approach, so
we gather feedback by collaborating with the international research community. For example,
we addressed adversarial testing challenges for generative AI in The ART of Safety workshop
at the Asia-Pacific Chapter of the Association for Computational Linguistics Conference
(IJCNLP-AACL 2023).
One of our technical research approaches to scaled data generation is reflected in our paper
on AI-Assisted Red Teaming (AART). AART generates evaluation datasets with high diversity
(such as sensitive and harmful concepts specific to a wide range of cultural and geographic
regions), steered by AI-assisted recipes to define, scope, and prioritize diversity within an
application context.
To catalog our research in responsible data use for generative AI, we maintain an internal
centralized data repository with use-case and policy-aligned prompts. We have also developed
multiple synthetic data generation tools based on LLMs that prioritize the generation of data
sets that reflect diverse societal contexts and integrate data quality metrics for improved
dataset quality and diversity.
22Our data quality metrics include:
• Analysis of language styles, including query length, query similarity, and diversity
of language styles
• Measurement across a wide range of societal and multicultural dimensions, leveraging
datasets such as SeeGULL, SPICE, TIDAL and the Societal Context Repository
• Measurement of alignment with Google’s generative AI policies and intended use cases
• Analysis of adversariality to ensure that we examine both explicit (the input is clearly
designed to produce an unsafe output) and implicit (where the input is innocuous but
the output is harmful) queries
In addition, we explore understanding of when and why our evaluations fall short using
participatory systems, which explicitly enable joint ownership of predictions and allow
people to choose whether to disclose on sensitive topics.
Collaborating with the research community
An essential component of our research philosophy is supporting the free exchange of ideas
and maintaining close contact with the broader scientific community.
This year, we committed to supporting MLCommons’ development of standard AI safety
benchmarks. Though there has been significant work done on AI safety, there are as of yet
no industry-standard benchmarks for AI safety. Standard benchmarks already exist in machine
learning (ML) and AI technologies: for instance, MLCommons operates the MLPerf benchmarks
that measure the speed of cutting-edge AI hardware such as Google’s TPUs.
MLCommons proposes a multi-stakeholder process for selecting tests and grouping them
into subsets to measure safety for particular AI use-cases, and translating the highly technical
results of those tests into scores that everyone can understand.
23Throughout the year, we’ve engaged with cross-disciplinary research communities to examine
the relationship between AI, culture, and society, through our recent and upcoming workshops
on Cultures in AI/AI in Culture, Ethical Considerations in Creative Applications of Computer
Vision, and Cross-Cultural Considerations in NLP. Our recent research has also sought out
perspectives of particular communities known to be less represented in ML development
and applications. For example, we have investigated gender bias in contexts such as gender-
inclusive healthcare.
This year, Google DeepMind researchers introduced the area
of model evaluation for extreme risks...These evaluations are
likely to inform responsible decisions about model training,
deployment, and security.
Our researchers continue to explore new areas of AI risk. Current approaches to building
general-purpose AI systems tend to produce systems with both beneficial and harmful
capabilities. Further progress in AI development could lead to capabilities that pose extreme
risks, such as offensive cyber capabilities or strong manipulation skills. This year, Google
DeepMind researchers introduced the area of model evaluation for extreme risks. Developers
must be able to identify dangerous capabilities (through “dangerous capability evaluations”)
and the potential for harmful outcomes (through “alignment evaluations”). These evaluations
are likely to inform responsible decisions about model training, deployment, and security.
Society-Centered AI as a research method
Our research is inspired by the transformative potential of AI technologies to benefit society
and our shared environment at a scale and swiftness that wasn’t possible before. From helping
address the climate crisis to helping transform healthcare, to making the digital world more
accessible, our goal is to apply AI responsibly to be helpful to more people around the globe.
Achieving global scale requires researchers and communities to think ahead — and act —
collectively across the AI ecosystem.
24We call this approach Society-Centered AI. It is both an extension and an expansion of
Human-Centered AI focusing on the aggregate needs of society, informed by the needs of
individual users, from understanding diseases that affect millions of people or protecting
the environment.
Multi-disciplinary AI research can help address society-level,
shared challenges from forecasting hunger to predicting
diseases to improving productivity.
Recent AI advances offer unprecedented, societal-level capabilities. In 2023, for example,
Google DeepMind’s new AI model that classifies missense variants, genetic mutations that
can affect the function of human proteins and can lead to diseases such as cystic fibrosis,
sickle-cell anemia, or cancer, was used to create a catalog of “missense” mutations that
categorized 89% of all 71 million possible missense variants as either likely pathogenic or likely
benign. By contrast, only 0.1% have been confirmed by human experts. This knowledge is
crucial to faster diagnosis and developing life-saving treatments. And our recent research with
Boston Consulting Group also found that AI also has the potential to mitigate 5-10% of global
greenhouse gas emissions by 2030.
Multi-disciplinary AI research can help address society-level, shared challenges from
forecasting hunger to predicting diseases to improving productivity. To help promote diverse
perspectives in this work, we announced that 70 professors were selected for the 2023 Award
for Inclusion Research Program, which supports academic research that addresses the needs
of historically marginalized groups globally.
25Our research seeks to:
•
Understand society’s needs
We focus our efforts on goals society has agreed should be prioritized, such as the
United Nations’ 17 Sustainable Development Goals, a set of interconnected goals
jointly developed by more than 190 countries to address global challenges.
•
Address those needs collectively
Collective efforts bring stakeholders (such as local and academic communities,
NGOs, private-public collaborations) into a joint process of design, development,
implementation, and evaluation of AI technologies as they are being developed
and deployed to address societal needs.
•
Measure success by how well the effort addresses society’s needs
We identify primary and secondary indicators of impact that we optimized through
our collaborations with stakeholders.
Our research will continue to promote AI applications that support the UN’s Sustainable
Development Goals and our efforts to help non-profits use these tools.
26Product impact
Our responsible approach to AI research and governance helps our product teams working
on applications for consumers, developers, and enterprises.
We have applied this approach to Gemini — our family of base and instruction-tuned models
of various parameter-based sizes, all of which are natively multimodal. Gemini is flexible
and optimized for three sizes: Gemini Nano, Gemini Pro, and Gemini Ultra. Gemini Pro
and Nano are starting to roll out to our products. We will be making Gemini Ultra available
to select customers, developers, partners, and safety and responsibility experts for early
experimentation and feedback before rolling it out to developers and enterprise customers in
early 2024. Gemini is designed with responsibility as a core goal: addressing challenges from
new capabilities, such as multimodality, and implementing state-of-the-art safeguards.
Across our products, we apply a risk-based, principles-driven process — which can also
mean taking a cautious and gradual go-to-market approach involving rigorous testing.
For example:
AI Principle #1: Be socially beneficial
This principle helps teams consider how the overall benefits of generative AI exceed risks
in areas such as content quality and AI’s impact on industries and sectors.
Consider our decision to develop Universal Dubbing, a generative AI-automated video lip
dubbing service. This technology carries risk, as it could be misused for highly believable
deepfakes. Rigorous research with partners at the University of Arizona showed the method
clearly helped non-native English speakers learn a language faster when watching a realistic,
automatically dubbed video. AI Principles reviewers approved the project with a strict gating
process for research and educational purposes based on clear benefits for students. As we
expand this service, we’re implementing guardrails to help prevent misuse and we make it
accessible only to authorized partners.
This principle also can be applied on a broader level. For example, it’s reflected in YouTube’s
approach to music generative AI experiments and YouTube’s product-specific guidance for
working with creators.
27YouTube is actively collaborating with a diversity group of leading musicians for their input
on developing generative AI tools to enable expression while protecting music artists and the
integrity of their work.
This year, we expanded our ads policies to require advertisers to disclose when their election
ads include material that’s been digitally altered or generated and depicts real or realistic-
looking people or events in all countries where we have election ads verification. And we
expanded our ongoing work in information literacy to support AI literacy. We launched About
this image, a tool that provides more context to help people evaluate visual content they come
across online. The tool offers details on when an image and similar images were first indexed by
Google, where it may have first appeared, and where else it’s been seen online (like on news,
social, or fact-checking sites). With this background information on an image, people might
be able to see that news articles pointed out that an image was AI-generated.
AI Principle #2: Avoid creating or reinforcing
unfair bias
We have taken a phased approach to launches to account for rigorous adversarial testing
for fairness. While we can’t disclose some of the details of our fairness testing methods for
security reasons, we can report that we release publicly available generative AI experiences
only after they have incorporated recommended or conditional mitigations.
This year, we’ve been testing how generative AI in Search can help people find what they’re
looking for in new, faster ways. The experience helps with a variety of information needs,
including those that benefit from multiple perspectives to avoid unfair bias.
As we’ve continually improved the experience, we’ve also expanded internationally beyond
the United States with recent launches in India and Japan, with the majority of feedback
positive. In our largest global expansion, we’ve brought generative AI in Search to more than
120 countries and territories including Mexico, Brazil, South Korea, Indonesia, Nigeria, Kenya,
and South Africa, with support for four new languages: Spanish, Portuguese, Korean, and
Indonesian. So if, for example, you’re a Spanish speaker in the US, you can now use
generative AI in Search with your preferred language.
28Case study: Lookout
Lookout is an assistive Android app that uses a phone’s camera to create accessibility tools
for people who are blind or have low vision (BLV). Lookout helps people complete common
tasks by making the visual world more accessible. Its newest flagship feature — Image Q&A —
enables people to not only get a much more detailed description of an image, but also to ask
questions about a photo, and receive AI-powered responses.
Describing images is inherently challenging. If an image contains people, it’s even more
complex, as difficult questions arise about how to describe those people in a way that’s
both useful and respectful of a person’s identity. Gender is a particularly challenging trait to
describe based on an image, as a person’s gender may not be obvious from their appearance.
While developing Lookout, the product team had to balance AI Principle # 1 (Be socially
beneficial) and # 2 (Avoid creating or reinforcing unfair bias). Though it may be beneficial to
include gender in the description of a person, doing so also risks potential unfair bias.
The team incorporated a Google DeepMind visual language model (VLM), heavily customized
for this use case, with several rounds of feedback from BLV people and from trans and non-
binary people. VLMs enable people to ask natural language questions about an image. The new
Lookout question and answer feature allows users to go beyond captions and ask about the
image details that matter to them the most.
29This functionality allows the team to provide captions without perceived gender, but if the
user asks a question about a person’s gender, the model can provide a best guess of perceived
gender, using cues from the person’s appearance. The Lookout team tested this approach
with end users who were BLV and non-binary and found that these users thought the
approach was both useful and respectful.
The approach isn’t perfect. The model will still make mistakes with perceived gender, and
people with visual impairments still need to request details that typically sighted people
receive effortlessly. The Lookout team believes this launch is both a step in the right
direction, and an area where we can continue to learn and improve with the BLV community.
AI Principle #3: Be built and tested for safety
We design all of our products to be secure-by-default — and our approach to AI is no different.
In 2023, we introduced our Secure AI Framework (SAIF) to help organizations secure AI
systems, and we expanded our bug hunters programs (including our Vulnerability Rewards
Program) to incentivize research around AI safety and security.
To address international frameworks and guidance for safe, secure, and trustworthy AI, we’re
prioritizing cybersecurity safeguards. Our goal is to protect proprietary and unreleased
models and we’re participating in industry-wide events to support broader protections for
governments, companies, and civil society, like the Defense Advanced Research Projects
Agency’s (DARPA) AI Cyber Challenge, which will aim to identify and fix software
vulnerabilities using AI.
As we introduce generative AI technology to younger users aged 13-17, we strive to strike the
right balance in creating benefits while prioritizing safety, family controls, and developmental
needs. Informed by research and experts in teen development, we’ve built additional
safeguards into the experience. For example, for our expansion of Search Generative
Experience to teens, to prevent inappropriate or harmful content from surfacing, we put
stronger guardrails in place for outputs related to illegal or age-gated substances or bullying,
among other issues.
30Case study: Search Generative Experience
Search Generative Experience (SGE), was introduced through Search Labs this year as a
generative AI experiment. Search powered by generative AI can help people quickly get the
gist of any topic, find new ideas and inspiration, and easily follow up on questions to deepen
their understanding. Generative AI in Search makes it easier for people to ask more specific
and complex questions like “How to make learning math fun for a ten-year-old?” People can
also ask follow-ups without having to repeat context or try suggested follow-ups, and get AI-
powered overviews with links to explore fresh perspectives from across the web.
We are rolling out SGE thoughtfully, to develop this experience responsibly, leaning on Search
protections like automated systems that work to prevent policy-violating responses and
filtering images that violate our prohibited use policy for generative AI. Other approaches
include adding metadata and watermarks indicating that images are AI-generated.
LLMs can generate responses that seem to reflect opinions or emotions, since they have
been trained on a range of language. We trained the models that power SGE to refrain from
reflecting a persona. They are not designed to respond in the first person, for example,
and we fine-tuned the model to provide objective, neutral responses that are corroborated
with web results.
31By making generative AI in Search first available through Search Labs, we were transparent that
the technology was still in an early phase. We’re committed to a thoughtful cadence of global
expansions after careful testing with audiences around the world.
Over time, we will continue to conduct evaluations and adversarial testing and share
information on SGE’s capabilities and limitations. In many cases, we have already made
improvements with model updates and additional fine-tuning. Generative AI has the potential
to transform the current Search experience by organizing and presenting information in ways
that help people get — and do — more from a single search.
AI Principle #4: Be accountable to people
When we launch products, we seek to provide relevant information and opportunities
for feedback. For example, for Bard’s initial launch in May, some of our explainability
practices included:
• The “Google it” button, providing relevant Search queries to help users validate
responses to factual questions
• Thumbs-up and -down icons as feedback channels
• Links to report problems and offer operational support to ensure rapid response
to user feedback
• User controls for storing or deleting Bard activity
We also try to let users know when they are engaging with a new generative AI technology
and document how a generative AI service or product works. For Bard’s launch, this included
a comprehensive overview of the cap on the number of interactions to ensure quality and
accuracy, efforts to prevent potential personification, other details on safety, and a privacy
notice to help users understand how Bard handles their data.
In addition, we’re broadly focused on ensuring that new generative AI technologies have
equal guardrails and accountability mechanisms when addressing concerns such as image
provenance. In addition to SynthID, our efforts include clear disclosure of images generated
by Google AI tools (as in Virtual Try On or Da Vinci Stickies).
32Case study: Bard
Bard is Google’s generative conversational AI experience, launched in early 2023. Bard can
support people’s productivity, creativity, and curiosity. From planning a party (Bard can come
up with a to-do list) to writing a blog post (Bard can provide an outline), people now have a
new and helpful creative collaborator.
The models behind Bard have been extensively trained and tested. As a result of potential
unfair bias in training data, generative AI products can produce offensive or factually
inaccurate output.
In the course of developing and launching Bard, we developed a number of new responsible
AI policies. For example, which types of content Bard is and is not allowed to generate
influenced our company-wide content frameworks for generative AI models. The team’s
thoughtful approach to development also shaped our understanding of emerging best
practices for responsible generative AI development, including adversarial testing and
the inclusion of clear, helpful explanations.
Bard was launched gradually so that the team could learn from real-world use by trusted
testers from a diversity of backgrounds and make adjustments as needed. Before launching
Bard, we conducted extensive adversarial testing to identify harmful outputs and make
improvements to the model. Bard continues to regularly undergo adversarial testing,
especially as new features are added.
33The Bard interface also makes it clear to people they’re interacting with a generative AI model.
Additionally, people can offer feedback on the quality of responses using the “thumbs up”
and “thumbs down” feature.
AI Principle #5: Incorporate privacy design
principles
Our foundational privacy protections for giving users choice and control over their private
data applies to generative AI. We’re applying these protections to new product features we’re
currently developing, like improved prompt suggestions that help people using Workspace
get the best results from Duet AI generative features. These are developed with clear privacy
protections that keep people in control.
We’re committed to protecting your personal information. If you choose to use the Workspace
extensions, your content from Gmail, Docs, and Drive isn’t seen by human reviewers, used by
Bard to show you ads, or used to train the Bard model. You’re always in control of your privacy
settings when deciding how you want to use these extensions, and you can turn them off at
any time.
As we continue to develop, improve, and expand audiences for our generative AI experiences,
we will update these protections and share more information on the Bard Privacy Help Hub
and elsewhere.
AI Principle #6: Uphold high standards of
scientific excellence
At our I/O event in May of 2023, we announced over 25 new AI-powered products and
features. This brings the latest in advanced AI capabilities directly to people — including
consumers, developers, and enterprises of all sizes around the world. Our most novel models
are developed with scientific rigor and transparency. In addition, we evaluate against multiple
criteria and, as appropriate, with external reviews.
For example, Med-PaLM 2, which was trained by our health research teams with medical
knowledge, can answer questions and summarize insights from a variety of dense
medical texts.
34It was assessed for scientific consensus, medical reasoning, knowledge recall, bias, and
likelihood of possible harm by clinicians and non-clinicians from a range of backgrounds and
countries. Med-PaLM 2 was opened up to a small group of Cloud customers for feedback to
identify safe, helpful use cases.
AI Principle #7: Be made available for uses
that accord with these principles
All advanced technologies have multiple uses, including potentially harmful or abusive
applications. Our AI Principles guide how we limit harms for people. As we learn more about
the emerging risks unique to generative AI, we are working to address these potential harms
with technical innovation. For example, we launched a beta version of SynthID to a limited
number of Vertex AI customers as a digital watermarking feature for Imagen, one of our text-
to-image models that uses input text to create photorealistic images. And we offer image
markups for publishers to indicate when an image they post to our platforms is AI generated.
We remain committed to sharing best practices with our customers and developers. That’s
why we publish Cloud Responsible AI Guides for enterprises. And when AI-powered asset
generation for Performance Max was first rolling out to advertisers in the US this year, we
offered information in the Google Ads Help Center for advertisers to learn more about asset
generation in Performance Max, along with our AI Essentials guide.
35Supporting global dialogue,
standards & policy
Building AI responsibly must be a collective effort. It’s necessary to involve academics and
labs proactively across the research community, as well as social scientists, industry-specific
experts, policy makers, creators, publishers, and people using AI in their daily lives. We engage
in broad-based efforts — across government, companies, universities, and more — to help
translate technological breakthroughs into widespread benefits, while mitigating risks.
For example, this year we:
• Participated in the White House-sponsored red teaming event at DEFCON, which drew
over 2,000 people to test industry-leading LLMs in an effort to better understand risks
and limitations of these advanced technologies.
• Co-established, with industry partners, the Frontier Model Forum to develop standards
and benchmarks for emerging safety and security issues of frontier models.
• Contributed to the Partnership on AI (PAI)’s efforts on a Synthetic Media Framework to
help develop and foster best practices across the industry for the development and
sharing of media created with generative AI; PAI’s Data Enrichment Sourcing Guidelines;
and PAI’s Guidance for Safe Model Deployment.
• Participated in a number of information sharing sessions about generative AI, including
at the Inter-American Development Bank, National Governors Association, US National
Conference of State Legislatures Summit, the UK Summit, and more.
• Collaborated with IPSOS on a study on how and why people across 10 countries
expect AI will affect privacy in the future, resulting in a paper presented at the 2023
Symposium on Usable Privacy and Security conference.
• Updated our Machine Learning for Policy Leaders workshop with generative AI-specific
interactive sessions for policy makers.
36A policy agenda for responsible progress in AI
We’re not only focused on identifying risks and benefits of advanced AI. We’ve been hard at
work supporting the larger AI ecosystem with practical, scalable recommendations. Earlier this
year, we shared a detailed policy agenda for responsible progress in AI. We outlined a three-
pillared approach for governments to collaborate with the private sector, academia, and other
stakeholders to develop shared standards, protocols, and governance so we can boldly realize
and maximize AI’s potential for more people around the world.
The three pillars are:
1. Opportunity: Maximize AI’s economic promise, such as increased productivity
and upskilling
2. Responsibility: Create standards and share practices, and, as appropriate, prepare
for regulation
3. Security: Align human values while building complex AI to prevent malicious use
Our collaborations across the industry and alongside civil society and academia are building
common technical standards that could help align practices globally. These industry-wide
codes and standards could serve as a cornerstone for building regulatory frameworks that
can promote policy alignment for a worldwide technology.
Putting into place a framework that encourages interoperability across the world can be an
opportunity to prevent a very real risk of a fractured regulatory environment, which could
delay consumer access to helpful products across borders. This could make it challenging
for start-ups and entrepreneurs without the resources to comply with a complex set of
uncoordinated AI regulation. These outcomes could slow the global development of powerful
new technologies, and undermine responsible development efforts described in this extensive
report. Sound government policies are essential to unlocking opportunity, promoting
responsibility, and enhancing security, along with individual best practices and shared industry
standards for principled AI innovation.
37Conclusion
With the rapid advancements in advanced AI capabilities, we stand on the cusp of a new era
not only for computing, but also for society. Responsible AI innovation will help businesses
of all sizes thrive and grow, and support society in finding solutions to our toughest
collective challenges.
But to unlock the economic opportunity that advanced AI offers while minimizing workforce
disruptions, policymakers will need to invest in innovation and competitiveness, promote legal
frameworks that support innovation, and prepare workers for potential economic impacts
of these evolving technologies.
To bring this vision to fruition and sustain it over time, safely, a multi-stakeholder approach
to governance is necessary. Across industries and nations, we can learn from the experience
of the internet’s growth over decades to develop common standards, shared best practices,
and appropriate risk-based regulation.
To do all of the above safely and securely, governments will need to explore next-generation
trade control policies for specific applications of risky AI-powered software. Governments,
academia, civil society, and companies will need a better shared understanding, via common
definitions and consistently structured transparency documents that describe not only the
capabilities of AI models when integrated into products and services, but also their limitations.
We’re building a strong foundation to enable ourselves and others to embrace AI’s
transformative promise and continue to evolve for years to come, to help today’s workforce
thrive, and support future generations:
38• To better understand how knowledge workers expect generative AI may affect their
industries in the future, we conducted participatory research workshops for seven
different industries, with a total of 54 participants across three US cities.
• We’re expanding our Google Cybersecurity Certificate program, which can help anyone
prepare for a career in cybersecurity globally. For example, in Japan participants can
earn a professional certificate from Google through the Japan Reskilling Consortium.
This is in addition to existing partnerships with CERT-IN in India and Cyber Security
Agency of Singapore, through which we’re offering 125,000 scholarships across the
Asia-Pacific region.
• To enable businesses and enterprises of all sizes, we’ve developed brand-new generative
AI training options and are constantly adding to our training catalog on Google Cloud
Skills Boost. This includes two learning paths that each feature comprehensive content:
one is for the non-technical audience, Introduction to Generative AI, and the other,
Generative AI for Developers, is for technical practitioners (more advanced). Individual
courses are also available on their own.
• For AI engineers and product designers, we’re updating the People + AI Guidebook with
generative AI best practices. For the same audience, we continue to design AI Explorables,
including how and why models sometimes make incorrect predictions confidently.
• For tomorrow’s AI engineers and designers, we’ve launched Experience AI, a new
educational program that offers cutting-edge resources for students aged 11-14 and
their teachers on artificial intelligence and machine learning. This was developed in
collaboration with teachers.
• In 2024, Google will be opening a free after-school Code Next Lab for high schoolers
in Inglewood, California, a city where 9 in 10 individuals identify as Black and/or Latinx.
Google will be designing, building, and opening the new facility for an immersive
computer-science education program to develop the next generation of US Black,
Latinx, and Indigenous tech leaders.
39Our mission, since we were founded 25 years ago, has always been to organize the world’s
information and make it universally accessible and useful. Making AI helpful for everyone will
be how we deliver on this mission and improve lives everywhere. A big part of accomplishing
our mission means making information open and accessible on how Google’s core
technologies work. We’ve done this consistently in the transparent tradition of “How Search
Works,” which we made public a decade ago in 2013. A decade later, advanced AI is no
exception. In addition to this annual report, we regularly publish technical reports and research
papers that include, or complement, model cards for AI models that are incorporated into
AI-powered experiences.
We’re encouraged to see governments around the world calling for ongoing transparency
into internal AI governance processes and reporting on AI models’ capabilities and limitations.
Governments and civil society have been seriously addressing how to develop the right
policy frameworks for AI innovation this year, and we look forward to supporting their efforts
in years to come. At Google, we’ve been bringing AI into our products and services for over
a decade and making them available to people who use our products steadily, guided by
our AI Principles. We know we’re at an exciting inflection point in our journey as an AI-first
company. Some observers have tried to reduce this moment in the history of technology to a
competitive AI race across our industry. But what matters most to us is the race to build
AI responsibly, together with others so that we get it right – for everyone.
40Appendix
41December 2023
Generative AI System Card: Bard
with specifically tuned Gemini Pro
Bard with Gemini Pro is a conversational AI service that is available in English and in over 170
countries and territories. It will be made available in more languages and places, like Europe,
in the near future.
The AI system that powers this service uses a specifically tuned version (in English) of Gemini
Pro, a foundational large language model (LLM). LLMs are trained deep-learning models
that understand and generate text, images, video, and speech in a human-like fashion. LLMs
build statistical models of the language they are learning, trying to predict which words
are frequently used together across different types of texts and contexts to model the
relationships and interactions between words. When given a prompt, they generate a response
by selecting, one word at a time, from words that are likely to come next. LLMs must be trained
on a vast amount of multimodal data: text, images, video, and speech before they can learn
the patterns and structures of language. The information in this document refers only to
the version of Bard with Gemini Pro launched in December 2023.
Capabilities
Gemini Pro in Bard (as of December 2023) is specifically tuned for understanding,
summarizing, reasoning, coding, and planning capabilities. It works for text-based prompts
and provides generated text at this time. Other capabilities of Gemini Pro in Bard include
creative writing, composition, language translation, and complex problem solving, including
in math and science. At this time, Bard also uses Google Lens technology. We expect to
unlock advanced multi-modal capabilities in Bard over time.
Despite the growing range of LLM capabilities, there are known limitations to the use of
LLMs in AI-powered systems. There is a continued need for ongoing research and
development on how to improve verifiable model outputs so that they are more reliable
(e.g., to avoid “hallucinations”). Even when LLMs perform well against model performance
benchmarks, they can struggle with tasks requiring high-level reasoning abilities, like causal
understanding and logical deduction. Over time, it is necessary to develop more challenging
and robust evaluations in these areas.
42Intended use and current integrations
Bard is intended for creative collaboration and conversational AI assistance for consumer use.
Bard Extensions, available in English at this time, integrates with Google tools like Gmail, Docs,
Drive, Google Maps, YouTube, and Google Flights for more helpful responses. As of December
2023, third-party extensions are not yet available in Bard. Google is currently exploring features
that will enable users to connect with third-party services.
Data
Data Sources used to train Gemini Pro:
Gemini Pro is trained on datasets that are both multimodal and multilingual. Our pre-training
datasets use data from publicly available web documents, books, and code, and include image,
audio, and video data.
Safeguards:
We have implemented the following measures to improve the safety and quality of the LLMs
for use in products like Bard.
• Harms mitigation: Prior to training, various steps were taken to mitigate potential
downstream harms at the data curation and data collection stage for Gemini Pro.
Training data was filtered for high-risk content and to ensure all training data is
sufficiently high quality. Beyond filtering, steps were taken to ensure all data collected
meets Google DeepMind’s best practices on data enrichment.
• Mitigations for quality and safety, specific to Gemini Pro: Quality filters were
applied to all datasets used to train the pre-trained Gemini Pro model. Safety filtering
was applied to remove harmful content. Evaluation sets were filtered from the training
corpus. The final data mixtures and weights were determined through ablations on
smaller models. Training was staged to alter the mixture composition during training —
increasing the weight of domain-relevant data towards the end of training.
Additional mitigation measures are applied in Google’s products, including Bard, over time (as
described in Google’s AI Principles Progress Updates).
43Personal data collected and processed in providing the Bard service:
When people interact with Bard, Google collects:
• Conversations
• Location
• Feedback
• Usage information
This data helps provide, improve, and develop Google products, services, and machine-
learning technologies, like those that power Bard. Bard shows user-interface elements at the
bottom of the Menu that offer continuous transparency about location data processed by
Bard. Users can review their prompts, delete Bard activity, and turn off Bard activity at any
time. For more details, visit the Bard Privacy Help Hub, and read the Google Privacy Policy and
the Bard Privacy Notice.
Model Training Process
Pre-training
LLMs are pre-trained on the pre-processed data. Pre-training helps LLMs learn the patterns
and relationships in data (which ultimately will be used to generate responses for Bard in
Gemini Pro’s case).
LLMs built on Transformer1 architecture fundamentally map the statistical relationships
between words, phrases, and sentences, to predict what next words, images, video, or other
content will be most likely to follow when prompted with a new set of words. These models
first build these relationships in an “unsupervised” way — that is, without the data being
categorized or labeled. To do that, they need large quantities of language and language-like
data (such as code, math proofs, etc.) that cover the spectrum of use of language to model all
the possible relationships that exist across the vast breadth and complexity of words in a single
language and among the many languages used around the world.
1 A Transformer is a deep learning neural network architecture that Google introduced in 2017 (see https://ai.googleblog.
com/2017/08/transformer-novel-neural-network.html).
44Gemini models are trained to accommodate textual input interleaved with a wide variety of
audio and visual inputs, such as natural images, charts, screenshots, PDFs, and videos, and
they can produce text, image, video, and audio outputs.
Pre-training large AI systems requires substantial computational, human, and energy
resources, and currently can take substantial time to complete. While Google continues to
work to reduce such costs,2 these practical and technical challenges significantly limit how
often new foundational LLMs can be pre-trained. For details on the training infrastructure of
Gemini models, please see the Gemini technical paper.
Fine-tuning
Once a pre-trained model is created, it is then adjusted and adapted for use in a specific
application. This fine-tuning process takes additional data, some of which the model may have
already seen, and formats it in a way to match the expectations of the application the model is
being used in.
This process is undertaken with human supervision/feedback and by using reinforcement
learning. Fine-tuning can relatively quickly adapt LLMs to new policies and allow for more
experimentation to optimize outputs.
Instruction tuning encompasses supervised fine tuning (SFT) and reinforcement learning
through human feedback (RLHF) using a reward model. Instruction tuning was applied in both
text and multimodal settings. Instruction tuning for Gemini Pro in Bard was carefully designed
to balance the increase in helpfulness with decrease in model harms related to safety and
hallucinations. Curation of “quality” data included a mix of data to balance the metrics on
helpfulness (such as instruction following, creativity) and reduction of model harms.
Risks of harmful text generation for Gemini Pro are mitigated with technical approaches. For
example, a dataset of potential harm-inducing queries was generated to reflect risks and
societal harms, guided by the AI Principles. For Gemini Pro, this overall approach was able to
mitigate a majority of identified text harm cases without any perceptible decrease in response
helpfulness. These mitigations were made before integration into the Bard service.
2 See, e.g. how Google is minimizing our AI carbon footprint:
https://blog.google/technology/ai/minimizing-carbon-footprint/, https://ai.googleblog.com/2022/02/good-news-about-carbon-
footprint-of.html, and https://cloud.google.com/blog/topics/systems/tpu-v4-enables-performance-energy-and-co2e-efficien-
cy-gains.
45Model evaluations
Google engages in extensive evaluation and testing to ensure classifiers and other safeguards
are operating effectively in all of our novel generative AI models that power services like Bard,
including the following:
1. Pre-Launch testing: Prior to launch, our AI Principles, Trust & Safety and Responsible AI
teams engage in rigorous testing of safety guardrails, including classifiers, to evaluate
Bard’s performance after these guardrails are put in place. The teams generate large
sets of adversarial queries, as well as queries in sensitive verticals, to evaluate how
often an unsafe response is generated. Our Trust & Safety team consistently evaluates
such safety metrics to ensure the rates meet launch goals.
2. Targeted adversarial testing and red teaming: The Trust & Safety, AI Principles and
Responsible AI teams also conduct targeted adversarial testing and red teaming to
better understand how Bard, and the classifiers it leverages, are performing against
certain areas and identify failure patterns that need to be addressed.
3. Feedback: Google analyzes and measures the feedback our users provide. Users
have the ability to provide feedback if a response by Bard is low quality in the form
of a “thumbs-down” vote. A user submitting such feedback then indicates whether a
response is unsafe/offensive, not factually correct, or they can specify another reason.
Deployment and continued iteration
Specialized models built off the base LLM can be further fine-tuned for the specific needs of
the product or service where they are deployed. For example, Bard uses a specialized model
fine-tuned from the Gemini Pro base model. We continue to iterate on the Bard fine-tuning
recipe (data mixtures, fine-tuning parameters).
46Risk Assessment and AI Principles Review outcomes
Google’s AI Principles team conducted a risk assessment and review of Bard. Recommendations
resulted in additional extensive deep-dive dogfooding and adversarial testing in the areas of
safety, accountability, and inclusion to prepare for the initial experimental rollout of Bard and
subsequent updates. Further cross-functional work helped to ensure appropriate mitigations
were adopted before Bard and its updates, such as Bard with Gemini Pro, launched. These
product mitigations included the following:
• Clear and relevant explanations to set appropriate expectations that describe Bard as a
helpful service that offers collaboration with AI for specific types of tasks. Explanations
make clear that this AI-powered system is useful for brainstorming ideas, developing
plans, creating first drafts of written outlines, emails, blog posts, or for quick summaries
of complex topics.
• Disclosures in the Bard Privacy Notice stating that people should not rely on Bard’s
responses as medical, legal, financial or other professional advice.
• Disclosure in product stating that Bard responses should be double-checked for
information accuracy.
• Added ability to use Google Search to find content that helps users assess and further
research the information they get from Bard.
• Feedback channels and operational support were defined and built to help ensure rapid
response to user feedback to improve the model and address issues.
47Testing
For Bard and its updates, the testing approach is:
• Red teaming for security using the Secure AI Framework (SAIF)
• Adversarial testing for unfair bias
• Applying mitigations, for example, for content policy violations or abuse
• Conducting ongoing trusted testing of Bard with external users
in experimental releases
Safeguards
A set of “model policies” guided the model development of Gemini Pro and evaluations.
Model policy definitions act as a standardized criteria and prioritization schema for responsible
development and as an indication of launch-readiness. Gemini model policies cover a number
of domains including child safety, hate speech, factual accuracy, fairness and inclusion,
and harassment.
Other outcomes of ongoing AI Principles reviews of Bard and its updates include the following:
• Dedicated Bard safety teams and policies
• Review of Bard user safety feedback
• Technical safeguards such as classifiers and filters were used to enforce policies
• A restricted use or gating policy is used when appropriate
• A Generative AI Additional Terms of Service exists
• Ongoing technical testing to inform decisions and improvements
48In addition, ongoing AI Principles guidance for transparency and user control practices have
been implemented, including:
• Disclosure in product stating that Bard should be double-checked for
information accuracy
• Bard Privacy Help Hub
• FAQs
• Help article(s)
• In-product safety guardrails to add contextual help, like Bard’s “Google it” button to
more easily double-check answers
• Feedback opportunities for users
• Clear user controls
Note: information on Google’s overall model safety strategy and classifiers is highly confidential, commercially sensitive, and
proprietary information of Google. Any public availability of this information could expose people who use Google’s products and
the greater public to security and safety risks. As clear, consistent directives emerge, we aim to share additional transparency
artifacts in the context of scientific excellence (as stated in AI Principle # 6), with appropriate third parties on how best to offer
additional details while both remaining competitive and prioritizing people’s safety. At this time, for the responsible reasons stated
above, this document doesn’t offer specific details on any model size, training methods or compute, or other similarly proprietary
or sensitive information. The design and details of our transparency artifacts evolve over time to reflect the evolution of
technologies, product specifications, and user interface design. The content may be adjusted for the needs of various audiences.
49

****************************************************************************************************An AI Opportunity
Agenda for India04 Executive Summary
06 Introduction
07 India's AI Opportunity
Improving access to quality healthcare
Boosting agricultural productivity and sustainability
Improving educational and employment opportunities
Harnessing India’s linguistic and socio-cultural diversity
Enhancing citizen engagement with public services
Improving financial inclusion
12 An Affirmative AI Policy Vision for India
13 Investing in Innovation Infrastructure
Strengthening and democratising compute capacity
Adopting a cloud-first approach
Enhancing the accessibility and quality of open government datasets
Investing in R&D
Pro-Innovation Legal Frameworks
20 Building an AI-ready workforce
Modernising Skilling Programs for the AI Era
Supporting workers in transition
223
Promoting Inclusive Adoption and Accessibility
Promoting inclusive access to foundational infrastructure for AI, with
particular attention to marginalised communities
Government Adoption of AI
India's AI Opportunity Small Business and Traditional Industries
Enabling Regulatory Framework
27 Conclusion
An Affirmative AI Policy Vision for India
3Executive Summary
Executive Summary
We stand at a pivotal moment in the development of Infrastructure (DPI), including in healthcare, education,
artificial intelligence (AI) in India. India’s outstanding tech e-governance and finance, will help unleash AI’s potential
talent, growing economy and vibrant start-up ecosystem and accelerate the benefits of DPI.
mean that the country is well placed to fulfil its potential
for AI global leadership. By maximising the potential of AI, At Google, we are deeply committed to working with
India can drive productivity growth, raise living standards public and private sector stakeholders to realise India’s
and tackle some of the country’s most pressing social AI ambitions. Our work with our partners in India has three
problems. To do so, stakeholders must come together to core components:
build a comprehensive AI opportunity agenda to help
• Investing in foundational infrastructure for AI
deliver the IndiaAI Mission and harness the power of AI for
development: Google's ongoing investments in
the benefit of India’s citizens.
India's compute infrastructure, including the launch
With the right policy framework in place, AI can help of cloud regions in Mumbai and Delhi, demonstrate
accelerate the transformation of India’s economy, its commitment to the country's digital growth.
society, and public services. These efforts are crucial for scaling operations and
democratising AI technology across India, fostering
First, if fully harnessed, AI has the potential to innovation and development in various sectors.
substantially boost the Indian economy. AI will play
a crucial role in delivering the target of a $1 trillion digital • Improving AI skills through training: Google is
economy by 2028, which is expected to account for 20 continually developing its AI courses, such as the
percent of India’s GDP. A Google-commissioned report AI Essentials online course, and partnerships with
estimated that at least INR 33.8 lakh crore of economic the Indian public sector to help workers enhance
value from AI adoption can be achieved in India in 2030. their AI skills. Google in India has a clear goal to
empower 10 million people with AI digital literacy
Secondly, AI can accelerate positive social including students, job seekers, educators, startups
transformation and social inclusion in India. Applied to and developers and civil officials, and offer technical
agriculture, AI can improve agricultural sustainability and knowledge to help contribute towards building an
help prevent diseases. Agrostar, for example, launched a AI-ready workforce.
multilingual mobile app using Google Cloud that is helping
to boost crop yields and encourage sustainable practices • Developing innovative AI services and use cases:
for small farmers in India. This is contributing to raising Google is partnering with public and private sector
rural living standards in poorer parts of the country. stakeholders to foster AI innovations in essential
sectors and build a robust AI ecosystem across the
Thirdly, India has a unique opportunity to become country.
a global leader in leveraging AI for public services.
Further integrating AI across India’s Digital Public
4Executive Summary
• Through a collaboration with the Government Building on these efforts, we offer three key
of Maharashtra, Google will leverage its AI recommendations in this report on how India can
capabilities to provide AI-enabled solutions harness AI responsibly and to its fullest potential:
across sectors like agriculture, healthcare,
sustainability, education, and startups in the
State. To maximise the AI Opportunity for India,
collaborative efforts across government, industry,
• Google and the Government of Tamil Nadu and civil society must prioritise three key areas:
are collaborating to drive technological
advancement and foster AI innovation in the • Invest in infrastructure and innovation -
State, focusing on initiatives in key areas optimising the opportunities presented by this
such as AI start-up enablement, skilling, and technology by investing in AI research and
industrial ecosystem enablement including development, access to and quality of digital
MSMEs, to create impactful, scalable AI infrastructure and compute capacity, and
solutions. providing a balanced regulatory environment
to convert ideas and data into new discoveries,
• Google is contributing to the agriculture
products and services.
sector in partnership with local players
through AI-powered solutions that address
• Build human capital and an AI-empowered
challenges across the entire farming lifecycle.
workforce - investing in people to ensure they
By leveraging AI, stakeholders gain landscape
can use and benefit from AI, from students
insights, digital management tools, farmer-
to workers, and from small businesses to
focused resources, real-time crop prices, and
traditional industries.
personalised information platforms. All these
AI-driven collaborations foster innovation,
• Promote widespread adoption and
inclusivity, and efficiency, ultimately benefiting
universal accessibility - harnessing AI
farmers, consumers, and the environment.
across governments and all sectors of
the society to address major societal and
economic challenges and ensure the benefits
of AI are widely shared.
5Introduction
Introduction
The choices made by governments, industry, and civil Building on Google’s three-pillar agenda for responsible AI
society at early stages of technological development will progress – unlocking opportunity, promoting responsibility,
determine the speed and scale of adoption and the extent and enhancing security — this paper proposes three key
to which all parts of Indian society can benefit. AI has recommendations for Indian policymakers, companies,
the potential to fundamentally change the ways we live, and civil society to deliver AI's benefits to as broad a range
work and learn through its ability to assist, complement, of people as possible. To achieve this, we must work in
empower and inspire people in almost every field of human partnership to:
endeavour. It is already opening up new possibilities in India,
such as helping with flood forecasting, making it easier
1. Invest in innovation infrastructure;
to identify diabetic retinopathy, supporting agricultural
sustainability and maternal healthcare, improving
reading proficiency and gaining a richer understanding
2. Build an AI-ready workforce; and
of language diversity across India.
3. Promote inclusive adoption and accessibility.
Google believes that AI can drive innovative solutions
tailored to India's defining challenges. The possibilities
are immense: from addressing major public health
challenges, to boosting productivity and living standards,
and providing high-quality fulfilling jobs for many more
Indians.
Together, we must ensure that AI is introduced safely in a
way that improves wellbeing, helps solve India’s complex
challenges and enables India to lead globally in unique
areas of research and innovation. It is important that the
private sector works in partnership with the Government,
civic society and other bodies to help India meet the
ambitious goals outlined in the IndiaAI Mission.
6India's AI Opportunity
India’s AI Opportunity
With its world-leading AI skills and start-up ecosystem,
technologically optimistic population and strong public
and private sector investment in digital transformation,
India can leverage its unique strengths to harness
the AI opportunity. AI has the potential to bring real
societal benefits to India, boost economic growth, lift
living standards and provide India’s economy with the
competitive advantages it needs to be a global leader.
Improving access to quality
healthcare
AI has the potential to transform healthcare, potentially
saving millions of Indian lives through earlier diagnosis,
broader access and better understanding of diseases.
The use of AI in imaging and diagnostics is already speeding
up the diagnostic process. Google Cloud is working with
Karkinos Healthcare, an AI-powered oncology platform,
to provide hundreds of thousands of underinsured Indians
with easier access to cancer risk assessments, screening
and treatments.
AI is also being used to tackle health problems that have
a particularly devastating impact on parts of the Indian
population. The direct causes of maternal mortality are
largely avoidable and treatable with sufficiently early
intervention. Google has partnered with ARMMAN to
help them build a solution that uses AI to identify Indian
women at risk of dropping out from their health information
programme, which is designed to target preventative care
information to expectant and new mothers. The AI-powered
early targeting system helps ARMANN personalise
interventions and retain these individuals.
New AI applications are also improving accessibility to
healthcare for people living in remote or rural areas. Practo,
India’s largest digital healthcare start-up, seeks to promote
access to healthcare and increase the ease of navigating
the system by bridging the gap between patients and
healthcare providers. It uses AI and data analytics to deliver
affordable and personalised healthcare insights for people
in underserved communities.
7India's AI Opportunity
Boosting agricultural productivity
and sustainability
Agriculture is an essential part of India’s economy, with
70% of rural households depending on agriculture for their
livelihood and 82% of farmers being small-scale. Extreme
climate events such as floods, droughts, and infestations
are devastating for the livelihoods of India’s agricultural
workers. AI-driven data analytics can help India’s farmers
minimise these climate risks and optimise agricultural
productivity and sustainability.
AnthroKrishi and Google Partner Innovation, two teams
at Google, are leveraging AI to tackle this challenge. To
enhance agricultural sustainability, the teams are currently
exploring the use of AI-powered technologies to organise
and utilise India’s agricultural data. By combining satellite
imagery and machine learning to draw boundaries
between fields, their work has the potential to enable
sustainable farming practices and improve crop yields.
AI is also empowering India’s farmers with access to
real-time crop pricing information, levelling the playing
field with buyers and safeguarding farm livelihoods. Jiva,
a mobile app aimed at improving smallholder farmers’
livelihoods, provides services for farmers built with
Google’s Vertex AI. This includes providing farmers access
to adequate financing, high-quality agricultural inputs,
agronomic advisory services, and a fair market to sell
crops at harvest.
AI is further assisting farmers by providing early warning of
potential crop threats, enabling them to implement timely
mitigation strategies. A recipient of Google.org’s grant
funding, Wadhwani AI supports Indian farmers with pest
control through CottonAce, its AI-powered early warning
system. Cropin, an Indian start-up, provides farmers with
valuable insights into potential infestations and diseases by
combining AI, remote sensing, and geographic information
systems.
Farmer Empowerment: This initiative, in partnership with
Protean, ONEST, and ONDC, empowers farmers with
end-to-end solutions. By leveraging Google's location-
based insights, AI recommendations, and generative AI
8India's AI Opportunity
capabilities, farmers gain access to valuable information,
wider markets, and simplified processes for listing and
selling their produce.
Farmer Bot: This AI-powered voicebot provides farmers
with real-time crop price information in multiple Indian
languages. By integrating with the e-NAM portal API and
utilising Google Cloud's Chat Bison model, Farmerbot
ensures accessibility and delivers accurate market data
to farmers.
Improving educational and
employment opportunities
High-quality education is integral to maintaining
India’s rapid economic growth. Despite considerable
improvement in recent decades, India still has important
educational gaps, as highlighted by the 2023 Annual Status
of Education Report (ASER). AI can help bridge these
gaps, support underserved communities to gain access
to high quality education and improve employment
opportunities.
To this end, Google is partnering with the Rocket Learning
Foundation to use AI to improve early education for
children in India. Rocket Learning provides an AI coach
that creates localised content, automated grading and
personalised learning paths, making education more
accessible and effective.
AI is not only changing what the skills are that need to
be adopted, but also how these skills can be taught
and disseminated. It presents an opportunity to
personalise learning, reach more people, and maximise
impact. Initiatives like the Read Along Initiative (Bolo),
which uses voice-recognition technology to help
improve learning skills, demonstrate AI’s potential.
Additionally, AI is helping job seekers identify suitable
opportunities. APNA, a professional networking
platform using Google Vertex AI, is connecting millions
of Indian workers with opportunities and delivering
personalised job-matching based on their skills and
experience.
9India's AI Opportunity
Harnessing India’s linguistic and
socio-cultural diversity
India’s vast linguistic diversity and the presence of multiple
dialects – even within a single district like Muzaffarpur
- underscore the need for innovations that bridge the
linguistic divides. Project Navrassa, based on Google’s
Gemma family of open models, is already training large
language models (LLMs) on multiple Indic languages to
make AI accessible across a wide Indian audience.
As part of its commitment to ensuring an equitable AI
landscape that reflects India’s linguistic and socio-
cultural diversity, Google has also launched Project Bindi
(Bias Interventions for Natural Language Processing and
Data in the Indian context). The project aims to address
biases related to technology access and socio-cultural
factors including gender and caste, helping to ensure
the responsible development and implementation of AI
across India.
Enhancing citizen engagement with
public services
AI holds significant potential to build on India’s successful
Digital Public Infrastructure (DPI) model, facilitating
innovation, reducing bureaucracy, and boosting citizen
engagement through personalised services.
India is leading the way in using AI to improve public
services. According to a report by EY, generative AI is now
extending these benefits to more citizens across India,
particularly in underserved communities.
In collaboration with Axis My India, Google Cloud
introduced an AI-powered multilingual super-app
designed to build awareness of government social welfare
schemes, healthcare benefits and day-to-day amenities.
AI has allowed the app to provide personalised, real-time
answers to citizen queries about public services, in their
preferred language.
10India's AI Opportunity
Improving financial inclusion
Improving financial inclusion remains a crucial goal in India.
According to the 2021 World Bank Findex, India has a large
unbanked population, accounting for nearly 17 percent of
global unbanked adults. This issue often intersects with other
social challenges, disproportionately affecting poorer and
rural Indians. To tackle this, organisations are increasingly
turning to AI to enhance financial inclusion.
India’s Unified Payments Interface (UPI), to which Google Pay
has played a pivotal role, has made significant strides in this
area. AI is now further offering new ways for underbanked
communities to access financial services. Platforms like
OnFinanceAI, for example, are using AI to analyse data on
mobile phone usage patterns and e-commerce payments, to
help identify and onboard unbanked individuals.
11An Affirmative AI Policy Vision for India
An Affirmative AI Policy Vision for India
The examples above only scratch the surface of what The IndiaAI Mission has clearly displayed India’s ambition to
is possible. There is potential for AI to do so much more develop a robust AI ecosystem that catalyses large-scale
for India, significantly improving the lives of everyone in socio-economic transformation. If India wants to fully
the country. However, as we have seen from prior waves harness AI’s transformative potential, it must focus
of technology, these benefits are not automatic. Unless its attention on what it wants to achieve, not just what
people trust and see the benefits in using the technology, it wants to avoid.
it will not be adopted at scale.
In this light, we offer three key recommendations on how India can harness AI responsibly and to its fullest
potential:
Invest in infrastructure and innovation - meeting the moment of this technology by investing in
compute capacity, cloud infrastructure, cybersecurity measures, open government datasets, AI research
and development; and establishing policy frameworks / principles that enable responsible innovation.
Build an AI-ready workforce - investing in skilling to make sure people can use and benefit from AI,
from students to workers, and from small businesses to traditional industries.
Promote inclusive adoption and accessibility - harnessing AI across governments and all sectors of
the society to address major societal and economic challenges and ensure the benefits of AI are widely
shared, while adopting a regulatory framework that supports a healthy AI ecosystem.
12Investing in Innovation Infrastructure
Investing in Innovation Infrastructure
Countries have historically excelled when they support technological
change and harness it to improve living standards. India’s openness to
technological development positions it well in this regard. For India to unlock
the immense AI opportunity, it will require sound and deliberate investment
in innovation infrastructure, which includes not just technical infrastructure
but also legal and policy frameworks that enable responsible AI innovation.
There is no one AI investment strategy that will work for all governments,
but one basic formula for success is to invest in basic and applied
research and technologies (such as graphics processing units
and supercomputers), cloud infrastructure, and open government
datasets – and then to put in place policies encouraging private sector
innovation and product development that are built on top of these
foundational initiatives. Such a model can drive innovation leadership
by creating a sense of shared responsibility between the public and
private sectors for developing AI and other emerging technologies.
Strengthening and democratising compute capacity
Substantial investment in compute capacity – essential for building
India’s own foundation models and delivering AI solutions at scale – will
propel India on its path to global leadership in AI. The IndiaAI Mission
has outlined the clear direction of procuring 10,000 or more Graphics
Processing Units (GPUs) through partnering with the private sector. This
provides important momentum. Looking ahead, India will benefit from a fair
and transparent procurement process for these GPUs and other AI compute
infrastructure.
Meanwhile, India will benefit from taking a holistic approach to strengthening
compute capacity, further articulating government objectives in promoting
investment in cloud adoption, data centres and server optimisation
algorithms, and development of software to build AI systems. This is
important for ensuring that all three layers of the compute stack – hardware,
infrastructure, and software – meet the growing demands of Indian
start-ups and research communities to promote AI innovation.
13Investing in Innovation Infrastructure
Democratising access to compute is key to metrics, and transparent vendor management.
strengthening India’s AI ecosystem. The success of The roadmaps should also prioritise promoting
India’s Digital Public Infrastructure (DPI) in fostering competition to create value for governments and
digital inclusion may be instructive in how to inclusively avoid restrictive practices that hinder long-term
scale up compute capacity. Increasing investment in flexibility.
subsea cables can further deliver better connectivity by
expanding the supply of international bandwidth, which • Conduct targeted cloud and AI opportunity
is a crucial enabler of high-performance computing. assessments, focusing on services with the
greatest potential for citizen impact. Sectors like
Adopting a cloud-first approach healthcare, education, and transportation should
be prioritised. This can be done in partnership with
industry, which can help governments deploy AI
Cloud computing is the gateway through which
solutions on cloud.
businesses and governments can fully harness the
power of AI. Its vast computational resources, scalable
data storage, management, and analysis capabilities • Implement policies that address cybersecurity
are crucial for developing and deploying AI applications. concerns. The PwC’s 2024 Global Digital Trust
The Indian Government has led important cloud-first Insights survey finds that 52% of global business
initiatives and focused on improving the public sector’s leaders have heightened concerns about cyber
readiness for mass cloud adoption. As part of Meghraj attacks facilitated by generative AI usage,
– the national cloud computing initiative that prioritises underscoring the importance of addressing the
the adoption of cloud solutions over traditional security challenges to the cloud. Meanwhile,
on-premise IT systems – the Ministry of Electronics AI can be harnessed to detect and analyse
and Information Technology (MeitY) has implemented potential cybersecurity threats, strengthening
a cloud-first policy and developed guidelines that help cybersecurity defences in India.
Indian government departments with cloud adoption.
Google Cloud: Bringing AI-driven security
In order to maximise the benefits of AI, we recommend
operations to India
that governments:
Google is committed to developing quality cloud
• Continue to articulate and adopt a Cloud First
services and using AI to strengthen cloud security
policy. By showing the benefits of cloud-based AI
in India. Google opened its first two Cloud regions
solutions in delivering citizen services, the Indian
in India in Mumbai (2017) and Delhi (2021) and
Government can catalyse private sector use of AI
expects further expansion over the next few years.
tools on the cloud. Emerging homegrown cloud
Additionally, Google Cloud announced the plan to
solutions providers, such as NeevCloud which
use Gemini in Google’s new security operations
launched the country’s first AI SuperCloud, have
(SecOps) region in India. Google Cloud’s AI-driven
underscored the potential of cloud deployment
security operations aim to supercharge our public
in helping democratise access to AI and
and private sector customers’ security operations,
supercomputing for enterprises and startups.
to reap the benefits of secure cloud infrastructure.
• Establish clear frameworks to ensure successful
government-industry collaboration, such as
data governance and responsible AI guidelines,
procurement guidelines, well-defined performance
14Investing in Innovation Infrastructure
Enhancing the accessibility and
quality of open government datasets
Data held by central and state government agencies To tap on the potential of its data to empower AI
can be great enablers of growth and innovation. Enabling innovation, India will benefit from deepening efforts to
access to these diverse, high-quality government make open government datasets easily accessible
datasets is important to the commercialisation and across sectors and regions. India’s INDIAai
scaling up of AI solutions in India. While India holds an programme, launched in 2021, was an important early
advantage in the vast quantum of raw data that can sign of the Government’s commitment to opening up
be used for training AI models, the fragmentation and the Government’s anonymised data to researchers and
accessibility of government datasets can hinder AI start-ups.
development and adoption in the country.
Project Vaani
A Google-funded initiative led by the Indian Institute of Science, Project Vaani aims to collect speech data of 1
million Indians and open-source it for use in automatic speech recognition and speech-to-speech translation.
Now one of the largest datasets of Indian dialects ever to exist, Vaani will contain more than 150,000 hours
of audio across all districts in India upon the project’s completion. The vast data will boost the creation of an
AI-based LLM that captures the diverse Indian languages and dialects. This will lead to the development of
AI use cases that are relevant and tailored to Indians across the country.
15Investing in Innovation Infrastructure
Investing in R&D
Sustained investment in long-term AI R&D, while India can also leverage its strong relationships with
nurturing a culture of industry-academia collaboration, Global South countries to create a conducive global
will be crucial to ensure the global competitiveness of R&D environment for AI. A novel approach which the
India’s AI ecosystem and talent development. Indian Government could consider supporting globally
to augment local AI research would be the establishment
Public-private partnerships will be pivotal to of a Global Resource for AI Research (GRAIR) that
accelerating research across the AI ecosystem in India. would pool financial, technical and data resources
Both governments and industry can support academic across borders to help countries overcome resource
and civil society researchers through programmes such constraints. If successful, the initiative could make AI
as tech transfer frameworks, fellowships, and direct accessible to many more of the world’s entrepreneurs
support for research. Google Research, for example, and scientists.
has partnered with Indian academic institutions such as
IIT Madras. Such public-private initiatives must include
a broad range of participants to reflect the geographic,
linguistic and cultural diversity of Indian society.
Support the establishment of a Global Resource for AI Research
Inspired by successful models such as European Center for Nuclear Research (CERN) and the International
Space Station (ISS), the GRAIR would be a collaboratively governed, multinational AI research infrastructure
and research consortium working to ensure ethical development, equitable access, and the pursuit of AI
applications that foster local innovation. A collective computing resource such as the GRAIR would also help to
address concerns about AI’s carbon footprint, as it would reduce duplicative efforts and environmental impact.
The proposed GRAIR would comprise three key elements. A cloud-hosted Global Dataset Library would
feature diverse, curated, high-quality datasets, with continuous programmes addressing representational
gaps. A Distributed Compute Network would span data centres across multiple countries, particularly those
currently lacking dedicated AI infrastructure, providing researchers worldwide with essential computational
resources. An Operations Team would manage infrastructure, outreach programmes, and user initiatives to
ensure smooth functioning of the resource.
A GRAIR could undertake a range of activities, depending on the priorities of its members, including:
• Issue periodic requests for proposals (RFPs) that would allow researchers and organisations to apply
for compute time.
• Solicit proposals focused on creating high-quality datasets where gaps exist e.g. data related to
low-resource languages and cultural knowledge.
• Support in-person or remote safety testing, evaluations, and red-teaming on AI models for locally relevant
characteristics and development of associated benchmarks and testing suites.
• Support countries at different levels of development in building up domestic AI workforce capabilities,
including application developers, tech entrepreneurs and researchers, through training and accreditation
programmes.
16Investing in Innovation Infrastructure
Building a responsible AI research ecosystem in India
While AI has the potential to make our lives easier and address some of society’s most complex challenges —
such as preventing disease, making cities work better and predicting natural disasters – there are important
questions about fairness, bias, misinformation, security and the future of work. Answering these questions
requires deep collaboration among industry, academia, governments and civil society. This is why Google is
committed to building and supporting a responsible AI research ecosystem in India.
In 2023, with the support of a USD 1 million grant from Google, the Indian Institute of Technology Madras
(IIT Madras) opened its Centre for Responsible AI (CeRAI), a first-of-its kind multidisciplinary centre aimed
at ensuring ethical and responsible development of real-world AI applications in India. CeRAI has convened
responsible AI workshops to discuss key responsible AI challenges such as fairness, explainability and
accountability, privacy, and security. CeRAI also continues to focus on producing high-quality research
outputs, curating technical resources, and offering specialised training programmes on responsible AI. This
commitment to supporting a robust responsible AI research ecosystem in India is aligned with Google’s efforts
to share responsible AI best practices with the broader AI community.
Last September, Google launched the Digital Futures Project, establishing a 20 million dollars fund, an initiative
that aims to bring together a range of voices to promote efforts to understand and address the opportunities
and challenges of artificial intelligence (AI). Through this project, we are supporting researchers, organising
convenings and fostering debate on public policy solutions to encourage the responsible development of AI.
In India, the Digital Futures Project has supported the Aapti Institute to conduct exploration and research
on bias and digital integrity in the context of AI development and deployment. This research recognises the
need to unpack the risks associated with large-scale AI deployment, while attempting to delineate strategies
that leverage its transformative potential proactively.
17Investing in Innovation Infrastructure
Pro-Innovation Legal Frameworks
AI is too important not to regulate – and too important India is actively formulating and implementing policy
not to regulate well. At this moment, the challenge frameworks governing key aspects of AI regulation.
faced by all policymakers is how to govern AI in a way Google believes that building and optimising holistic
that mitigates risks and potential harms without policy frameworks can unlock public trust in AI and
impeding beneficial innovation. There is a risk that AI-derived opportunities in India.
conflicting and fragmented regulatory approaches
will block innovators and governments around the
world from harnessing trustworthy and beneficial AI
applications to achieve strengthened economies, find
cures for cancer, and ensure longer, better lives for
billions of people.
We believe there are seven major policies that policymakers in India should consider to ensure AI
researchers and innovators can convert ideas and data into new discoveries, products, and services.
1. Adopt a risk-based and proportionate approach to AI regulation focused on use cases: This is
crucial to provide clarity to developers, deployers, and regulatory agencies about which risks to mitigate
in specific contexts and which uses are completely disallowed. A risk-based approach will encourage
alignment around addressing the most severe concerns related to particular AI applications. This
approach recognises that AI is a general-purpose technology that will be applied in different ways in
different contexts.
2. A copyright framework that supports innovation and R&D: A strong predictor of a country’s leadership
in AI is its copyright framework, particularly when it is able to support broad usage of data inputs and
datasets by AI systems for learning and interaction with diverse information sources. To achieve this,
copyright frameworks must incorporate input from users, scientists, innovators, researchers, and creators
in the policymaking process. Creating exceptions to India’s Copyright Act (1957) to allow for innovation
in AI development and model training could help ensure copyright owners are appropriately protected
while not stifling innovation. A general AI framework must enable the limited use of copyrighted data
that allows startups and innovation companies to keep the large language models updated and relevant.
3. Facilitate better cross-border data flows: Deepening India’s links to the global data ecosystem will
be essential for catalysing India’s indigenous AI development and deployment. Providing increased
support to trusted cross-border data flows can enhance the capability of partners to work together
to ensure AI systems are trained on demographically and geographically diverse datasets, which helps
mitigate potential bias.
Data flows also empower Indian firms by helping them significantly reduce IT costs, leverage AI
technologies, serve cross-border customers and access global markets. A 2021 study published by
the Information Technology and Innovation Foundation estimated that increasing data restrictions is
associated with a 7% decrease in gross output traded and a 2.9% decrease in productivity over 5 years.
Creating flexible mechanisms that enable cross-border data flows can help India maximise economic
opportunity while safeguarding privacy and sovereign objectives.
18Investing in Innovation Infrastructure
4. Cohesive government AI policy: A cohesive approach to AI regulation is one that acknowledges
the cross-cutting nature of the technology, focusing on end applications rather than the underlying
technology. In India, AI governance typically requires participation from multiple agencies. It is therefore
vital that India develops an inter-agency approach that effectively coordinates and balances the emerging
AI-related work streams, avoiding a siloed approach to national AI regulations.
5. Conduct a regulatory gaps analysis to assess relevant legislations’ application to AI: The Indian
Government will benefit from undertaking holistic audits of regulations relevant to AI across the ecosystem,
to have a clear view of the existing regulatory landscape. The focus should be on examining whether
existing regulations sufficiently address the emerging risks from AI. Any new regulations on AI should
only be considered and introduced where legislative gaps are identified.
6. Promote international alignment and interoperability on AI regulations: India can be a pioneering
force in promoting regulatory interoperability that helps advance AI innovation and adoption globally.
As a founding member and the lead chair in 2024 of the Global Partnership on Artificial Intelligence
(GPAI), India can spearhead important discussions around AI, including: promoting the use of common
AI technical standards; amplifying the interests of the Global South in AI governance; and encouraging
AI adoption among SMEs.
7. Improve access to open datasets: Facilitating access to data is crucial for empowering Indian
researchers and entrepreneurs to develop reliable, high-impact AI models. Building an open data
ecosystem through initiatives that promote data commons and open government datasets, will play a
key role in this. To support such efforts, it's essential to establish a robust data infrastructure, especially
in critical areas like demographics, transportation, and environmental conditions.
19Building an AI-ready workforce
Building an AI-ready workforce
AI will transform the world of work by changing the nature of existing roles and
creating millions of new jobs and even new sectors. The required skills will also
evolve. Students need to leave education AI-ready and workers require the
ability to continually update their skills to remain relevant.
Realising India’s global leadership on AI will require an AI-ready workforce – an
area which India already has unique strengths. According to a recent Nasscom
report, India has the third largest AI talent base globally. The 2024 Stanford
University AI Index finds that India leads the world in AI skills penetration,
with a remarkable 263% growth in AI talent concentration from 2016 to 2023.
India produces a significant number of STEM graduates and its Institutes of
Technology are known for producing high quality tech talent.
Supported by its education system, India’s talent pool has fostered a thriving
ecosystem for start-ups, with over 1,900 startups focused on AI-driven
solutions in areas such as conversational AI, video analytics, deep fakes
detection, and disease detection. Indian workers are also already more likely
to be using AI in their day-to-day work, as highlighted in a report by Deloitte
which notes that India has the highest proportion of generative AI users in
Asia Pacific.
To sustain this momentum, India should leverage these strong foundations to
prepare workers for the potential disruption AI can bring. Addressing the AI
skills gaps – estimated to be up to 300,000 in 2024 alone – and bridging
the gender disparity in AI skills are crucial. India should also leverage its
talent and skills to support the Government’s mission of building local AI
models and engage the Indian diaspora in building AI infrastructure within India.
Compared to prior waves of technology, AI will present unique challenges
that will also require new solutions. The question then is: how can Indian
policymakers equip the workforce to harness AI, so that it empowers
workers, helps them become more productive, bumps up their expertise
level, and makes their skills valuable? Additionally, how can potential risks to
the workforce be mitigated through partnerships between the government,
industry and civil society?
20Building an AI-ready workforce
Modernising Skilling Programs for
Building an AI-empowered workforce will require
the AI Era
a shared vision – and a shared responsibility –
across a number of sets of stakeholders:
To tailor policy interventions, it will be important to
understand how AI is both similar to and different
• Industry has a crucial role to play in developing
from prior waves of technology. Early research
new skilling programmes that focus on AI
indicates that generative AI may help up level certain
preparedness. Given the transformative
skills, enhance labour productivity, create new
impact of AI across all sectors of the
occupations, and democratise access to higher paid
economy, individual company efforts would
occupations. However, as generative AI can automate
be insufficient on their own - companies
non-routine cognitive tasks, it may impact a wider range
will need to stand up new cross-sectoral
of tasks and occupations than earlier technologies.
AI training partnerships to ensure workers
in all industries are ready to harness AI.
We are only now building our understanding of what
kinds of new skills AI-enabled work will require. There
• Civil society, foundations, and academia
are some things we know already. This includes the
should drive new research to understand
importance of basic AI literacy and how critical
what has and hasn’t worked in the past
thinking, cross-disciplinary problem-solving, effective
in terms of worker preparedness for
collaboration, and empathy are likely to increase in
new technologies, and then apply those
value. Nevertheless, there are other open questions
insights to ensure that the benefits
about AI’s impact on work that will need further study.
of AI are able to be felt across society.
These include how best to use AI to support re-skilling
and how to minimise the risk of “skill atrophy”, as routine
• Crucially, policymakers have an essential role
tasks that previously provided training opportunities
to play in making sure that India’s population
for novice employees are increasingly automated.
have the skills able to make the most of
Companies, civil society, and policymakers will
the AI opportunity. Policymakers must help
need to constantly evolve skilling programmes to
scale up AI training programmes so that they
address these questions.
reach all communities, while building more
effective “trampolines” to catch workers that
We need an education and training system that prepares
are impacted by AI and reskill them so they
workers to thrive in a dynamic environment and to
can quickly bounce back into new and better
augment their existing skills and talents with AI. This
jobs. The expansion of the IndiaAI Future
must extend beyond the secondary education system
Skills programme will be an important part
to equip all students and workers with foundational
of ensuring that foundational AI skills reach
AI skills. FutureSkills Prime, a joint initiative by the
well beyond major Indian cities and into more
MeitY and the trade association, Nasscom, is offering
rural and underserved areas.
affordable digital upskilling courses, including in AI.
In a country with a burgeoning young population
like India, treating AI as a core component of the
education and professional development systems
can deliver important benefits. According to a report
by Educational Perspectives on Digital Technologies in
Modeling and Management, integrating AI in vocational
education and learning will significantly enhance skills
acquisition in India by 2030. AI-based tutorial and
simulation systems can meaningfully personalise
learning and provide interactive educational content.
21Building an AI-ready workforce
Encouragingly, the Central Board of Secondary
Education (CBSE) has introduced AI as a subject in Key steps that Indian policymakers can take to build
its affiliated schools to familiarise students with AI an AI-empowered workforce and support workers
concepts. It is important to build on and expand existing in transition include:
government initiatives to leverage AI in the classroom
and transform how students learn. • Encouraging companies that have developed
career certificate and apprenticeship
Looking ahead, India will benefit from further supporting programmes to work across sectors to develop
educators to update curriculum frameworks, doubling more comprehensive cross-sectoral skilling
down on STEM education with an emphasis on AI literacy and certificate programmes that reflect the
(while avoiding narrow recommendations like ‘learn to full spectrum of skills needed for an AI-driven
code’ that may be less relevant if generative AI can future.
cover basic coding skills), and emphasising skills-based
learning models, including apprenticeship programmes. • Building reskilling into the skills element of the
IndiaAI Mission and ensuring that workers have
the capability to update their skills at a variety
Supporting workers in transition
of points in their career. The $15 million AI
Opportunity Fund: Asia-Pacific - a collaborative
Every major technological revolution disrupts existing effort with Google.org and Asian Development
roles while creating new ones, and AI will be no exception. Bank - is supporting underserved workers
It is important for Indian workers across sectors, in both and job-seekers to build AI capabilities and
urban and rural areas, to gain the skills that help them confidence across the region, including India.
stay ahead of industry changes. This means that skills
development must not stop when people leave school or • Developing an AI adjustment assistance
college, and that one should instead embrace a “whole programme to provide support for workers
career” approach to learning and skills. impacted by AI, with a tailored set of skilling
options that can adapt to different worker
AI is already helping to democratise access to skills needs in different geographies, and a focus on
platforms, particularly benefiting underserved groups lower-wage workers and rural or underserved
in rural areas historically excluded from technological communities.
advancements. While the transition to an AI era may
entail some job losses in the short term, there is long-
term potential for AI to generate more jobs in India.
India must continue developing programmes and
strategies focused on skills for workers in transition
and new entrants to the workforce. Google is
continuing to update the Google Career Certificates
(GCC) programmes to help learners in India attain
critical digital skills. Over 80% of GCC graduates are
already reporting a positive career outcome within six
months of completion (e.g. a new job, promotion or
raise). The continued evolution of the IndiaAI Future
Skills programme should consider, in partnership with
industry and academia, how the skills environment will
continue to change and what that might mean for skills
courses that are provided.
22Promoting Inclusive Adoption and Accessibility
Promoting Inclusive Adoption and
Accessibility
In addition to building AI and preparing students and workers, India will benefit
from ensuring that AI is applied in a universally accessible and useful way to
help solve real world problems. To do this, we have identified four key goals:
1. Promote inclusive access to necessary infrastructure to access
and benefit from AI, with particular attention to marginalised
communities;
2. Increase governmental adoption of AI to make people’s lives
easier and better and address major public priorities;
3. Ensure that small businesses and traditional industries are able
to adopt AI applications; and
4. Regulate AI applications in a way that facilitates their adoption
across different industries.
Promoting inclusive access to foundational
infrastructure for AI, with particular attention to
marginalised communities
As earlier noted, AI has the potential to positively transform India’s society
and economy significantly. Indeed, many of the most substantial benefits
will be realised in rural areas and agriculture, where AI can play a crucial role
in tackling poverty and bridging the urban-rural economic divide. However,
fulfilling and maximising these benefits requires robust infrastructure to be in
place throughout India.
23Promoting Inclusive Adoption and Accessibility
Bridging the urban-rural digital divide, especially in access help mature the quality and safety of commercial and
to the internet, is crucial. There are a number of ways that enterprise AI products.
India can ensure underserved and diverse communities
are engaged in each element of the AI design and An integral element of governmental use of AI is ensuring
implementation process: that AI is built into public services as part of India’s
Digital Public Infrastructure (DPI). AI can shift DPI from
• Building safeguards into model design to ensure being an important element of public service delivery to
that different communities within India are taken using DPI to genuinely transform public services around
into account. Google is continuing to undertake the citizen. AI can be used to analyse complex data sets,
research into how fairness can be ensured with an anticipate potential issues such as financial distress
often complex and sometimes overlapping range and craft proactive interventions. As early examples
of social groups. of what is possible, Indian Government’s projects
such as Bhashini have shown the benefits of ensuring
• Inter-disciplinary and participatory design that digital services reach all citizens, regardless of
processes, with AI developers working directly language.
with different communities that might be impacted
by their systems. Additionally, AI can revolutionise the delivery of
welfare services. A Bengaluru-based NGO, Societal
• The Government should incentivise fundamental Thinking, has highlighted that AI could empower India’s
research on the development of responsible and underserved communities to better engage with the
inclusive AI and ensure that stakeholders from state. For instance, AI-enabled personalisation can
different groups are included in the research tailor services to individual needs of underserved
process. This should also involve convening citizens. Furthermore, AI streamlines Indian citizens’
research conferences and responsible AI interactions with government agencies. India’s diverse
workshops to help socialise India’s own models of population also enriches datasets for AI development.
fairness into the development of global responsible Finally, AI solutions can evolve over time in order to work
AI frameworks. at scale, based on the data that becomes available on
the effectiveness of their results.
• Ensuring that a rich variety of datasets are utilised,
including those of diverse and underserved groups. India will benefit from continuing to assess the most
Such Indian specific datasets are crucial in making effective uses of AI in local and regional public service
sure that AI solutions are developed for the local delivery. Based on such comprehensive national
context. AI opportunity assessments, the Government can
then consider whether such initiatives can be scaled
nationally, potentially through partnering with the
Government Adoption of AI industry. This process may include identifying and
addressing any procurement barriers to AI adoption and
moving towards more transparent rules that facilitate
The Indian Government can lead by example in AI
fair competition and accelerate AI adoption.
adoption and, in turn, promote economy-wide adoption
of AI in two ways. First, it can leverage AI to improve the
It is also essential to build awareness and understanding
delivery of services to citizens, which has the additional
of AI within all government departments. Working with
benefit of familiarising people with the underlying
industry through training courses, such as Google’s
technologies and building trust that AI can be used
AI Essentials course, presents an effective way of
in helpful ways. Second, by adopting AI, it can further
equipping the public sector with foundational AI
strengthen its domestic technology sector. The scale of
knowledge.
government deployment and investment can ultimately
help further catalyse a domestic AI ecosystem and, by
requiring standards for AI system performance, can also
24Promoting Inclusive Adoption and Accessibility
Finally, the Indian Government will need to build period. Flagship manufacturers are quickly introducing
adequate AI expertise to effectively harness AI. AI into their factories, design processes and customer
Policymakers should build and scale up “in-house” AI engagement. Tata Steel, for example, is using AI to
skills for the government IT workforce. Google took help predict when machinery might fail and Mahindra
a similar step a few years ago requiring all software Group is using AI for quality control and customer
engineers to enrol in an internal machine-learning personalisation.
curriculum.
To fully harness AI’s potential for traditional industries
like manufacturing and agriculture and fulfil the Indian
Small Business and Traditional
Government’s “Make In India” ambitions, it is important
Industries that the Government identifies barriers to increasing
the use of AI within these key sectors. As an example,
India’s business services, finance, transportation, digital infrastructure and skills training will be essential
education, retail, and healthcare are expected to for AI adoption in agriculture, particularly among
benefit significantly from AI, driven by their focus on smallholders.
digitalisation, productivity, efficiency, and personalised
experiences. For start-ups and SMEs across sectors,
early use of AI could prove transformative, boosting
their productivity and ability to reach new customers. To further address this AI implementation gap,
the Indian government should build on existing
MSMEs form a key pillar of India’s economy, contributing work and:
29% to GDP in 2021-2022. As MSMEs stand to benefit
from greater AI adoption, government support on • Identify priority national sectors that have the
MSMEs’ AI deployment becomes crucial. For example, highest need and/or the lowest uptake of AI
the Ministry of Micro, Small & Medium Enterprises (M/o tools, such as the agriculture, manufacturing,
MSME) can play a helpful role in engaging with MSMEs healthcare, and energy sectors, and work
to evaluate their state of AI readiness. with these sectors on “proof of concept”
initiatives to model effective AI deployment.
The private sector also has a critical role to play in
enabling MSMEs’ AI adoption. A useful step is to • Give small businesses a “digital jumpstart”
give MSMEs access to the essential AI technical through new models of technical assistance
infrastructure. A number of technology companies, and engagement, including digital coaches
including Google, are working to improve access who can help businesses understand and
to compute capacity for small businesses. leverage AI to capitalise on new opportunities.
Private-sector solutions can also be invaluable in
helping small business owners develop AI skills - the • Improve access to capital, including through
Grow With Google Educators programme, for example, low-interest loan and grant programs
provides digital upskilling for educators and teachers, designed to support AI-driven transformation.
including those providing skills training within The Indian Government can build on its grant
companies. Public-private initiatives to support the initiatives like those included in the IndiaAI
creation of impactful AI use cases for small businesses Mission. Singapore’s Infocomm Media
should also be encouraged, with Google’s Accelerator Development Authority (IMDA) recently
Programme for Indian startups as a positive model. announced a plan to provide funding support
to up to 300 SMEs to trial generative AI
India’s manufacturing sector is rapidly adopting AI, enterprise solutions developed by IMDA.
with 54% of Indian manufacturers already using AI and
advanced analytics, a 20% increase over a two year
25Promoting Inclusive Adoption and Accessibility
• Target AI training resources towards small
businesses and traditional industries
in underserved communities to build
confidence and competency.
Enabling Regulatory Framework
The Indian Government has proactively recognised
the need to create frameworks that empower Indian
enterprises, including SMEs and traditional industries,
to embrace AI, rather than hinder their adoption. Indian
regulators should further consider what approaches
will facilitate the adoption of AI, including adoption by
SMEs with fewer resources. Any AI regulation should
therefore be proportionate, light touch, risk-based
and focused on applications, recognising that AI is a
general purpose technology. Regulatory requirements
should be calibrated to the particular risk and use case
to avoid inhibiting low risk AI innovation and adoption of
AI technologies, at large.
The use of common international and technical
standards can also facilitate AI adoption and growth for
SMEs. Common standards mean that where a SMEs is
required to show its compliance with a regulation, it can
do so by showing adherence to the common standard,
rather than having to meet a bespoke requirement.
It will be beneficial for the Government to maintain
its engagement with international standards bodies
focused on the responsible development of AI systems,
in particular the ISO.
26Conclusion
As the Indian Government looks to realise the potential of AI technologies
to serve Indian society, and increase the public’s trust in AI, it has a critical
global role to play in developing AI policy frameworks whereby safety, security,
innovation, and opportunity are addressed cohesively. We look forward to
partnering with the Indian Government, industry, and civil society to build an
AI-driven digital future that works for everyone.
27

****************************************************************************************************2021 AI Principles
Progress Update2021 AI Principles Progress Update
Table of contents
Overview ...........................................................................................................................................2
Internal governance and operations ........................................................................5
Resources, research, tools, and responsible practices ......................... 8
Product impact ..........................................................................................................................12
Supporting global dialogue, standards, and policy ..................................14
Conclusion .....................................................................................................................................16
End Notes ........................................................................................................................................17
12021 AI Principles Progress Update
Overview
AI continues to help people around the world with everyday tasks, like real-time
translation, driving directions, and optimized video lighting so everyone is visible in virtual
meetings. AI can also be used to address some of the world’s most challenging problems.
It is already helping pathologists grade prostate cancer more quickly and consistently,
optimizing the efficiency of traffic lights, predicting floods and other impacts of climate
change, and informing COVID-19 policy decisions and drug discovery.
Google’s approach to AI is a result of more than two decades of responsible innovation
at the company. We began machine learning (ML) work in 2001, deep learning research
in 2011, user research on algorithmic fairness in 2014, and a cross-functional ML Fairness
initiative in 2016, supporting a number of projects in ML fairness, explainability, privacy,
and safety. In 2017 we began developing an ethical charter to guide the development and
use of AI, resulting in publication of the Google AI Principles in 2018. Since then we have
been working to improve implementation, operationalization, and governance of the AI
Principles at Google.
The fact that AI carries risks, alongside enormous benefits, is increasingly recognized
by the private sector as well as policymakers and civil society. AI principles, governance
proposals, and regulations have proliferated around the world, with a growing focus on
issues of fairness, safety, and privacy of AI systems. The OECD, G20, and G7 have
adopted AI principles, and AI governance standards are being developed by IEEE, ISO,
and national standards bodies such as NIST in the US. In April 2021, the EU released the
first horizontal AI regulatory proposal, the AI Act, which outlines detailed requirements
and a regulatory structure to manage high-risk AI systems. Korea’s Basic Act on
Intelligent Information Society, which includes protections against potential risks of AI
systems, went into effect in December 2020, and in September 2021, Brazil’s House of
Representatives introduced draft legislation on AI governance. Other AI governance
proposals are being developed in many countries, including India, Israel, Colombia, China,
and the UK.
Within Google, we’ve built a three-tiered governance program underpinned by industry-
leading research and a growing library of resources, tools, and recommended practices.
We’ve launched new features and options to empower our users with more control of
their data, and explanations of how our products work. We’ve developed policies and
frameworks to guide internal teams, and made tough decisions, for example in 2018
affirming not to release general-purpose facial recognition APIs despite potential
business opportunities. And we continue to work with governments and international
organizations on AI standards and rules, recognizing the enormous importance of AI
innovation and experimentation together with the need for regulation.
While we have made a lot of progress, we are committed to continuing to learn and
improve. For example, after a prominent researcher left Google in the past year, we
updated our research review process to increase consistency and transparency,
appointed HR specialists to review certain sensitive employee exits, and as part of our
22021 AI Principles Progress Update
ongoing racial equity commitments1 more than doubled the retention team to ensure we
are meaningfully addressing experiences in the workplace, including relationships with
managers and peers. In order to better coordinate and execute on advancing progress
in this space, we also consolidated dozens of research and technical teams working in
responsible and ethical AI into the same organization.
There are a number of open technical questions. For example, we are still learning how to
improve the development of machine vision systems that work well for a variety of skin
tones. Earlier this year we launched our DermAssist tool to help users better understand
skin conditions. While initial experiments demonstrate the tool performs well across
self-reported ethnicities, more work needs to be done to understand its performance
on the darkest and lightest skin tones. This year we also launched Real Tone, which helps
our imaging products like the Pixel phone camera and Google Photos more accurately
and beautifully represent a diverse range of skin tones. We developed Real Tone in
collaboration with photographers, cinematographers, and directors known for their
understanding of the challenges in lighting and capturing a spectrum of skin tones, and
who also represented communities of people with darker skin tones. We will continue
work on alternative and more inclusive measures that could be useful in the development
of our products, in collaboration with scientific and medical experts as well as groups
working with communities of color.
While we still have a lot to learn—and will continue learning given the dynamic and evolving
nature of technology and society—we remain committed to sharing our progress and
findings. In this year’s report, we outline our progress in AI Principles implementation to date,
highlighting advances in internal governance and operations; resources, research, tools, and
responsible practices; product impact; and supporting global dialogue, standards, and policy.
32021 AI Principles Progress Update
Google AI Principles
We will assess AI in view of the following objectives. We believe AI should:
1. Be socially beneficial: with the likely benefit to people and society substantially exceeding the foreseeable
risks and downsides.
2. Avoid creating or reinforcing unfair bias: avoiding unjust impacts on people, particularly those related to
sensitive characteristics such as race, ethnicity, gender, nationality, income, sexual orientation, ability and
political or religious belief.
3. Be built and tested for safety: designed to be appropriately cautious and in accordance with best practices
in AI safety research, including testing in constrained environments and monitoring as appropriate.
4. Be accountable to people: providing appropriate opportunities for feedback, relevant explanations and
appeal, and subject to appropriate human direction and control.
5. Incorporate privacy design principles: encouraging architectures with privacy safeguards, and providing
appropriate transparency and control over the use of data.
6. Uphold high standards of scientific excellence: Technology innovation is rooted in the scientific method
and a commitment to open inquiry, intellectual rigor, integrity and collaboration.
7. Be made available for uses that accord with these principles: We will work to limit potentially harmful or
abusive applications.
In addition to the above objectives, we will not design or deploy AI in the following application areas:
1. Technologies that cause or are likely to cause overall harm. Where there is a material risk of harm, we will
proceed only where we believe that the benefits substantially outweigh the risks, and will incorporate
appropriate safety constraints.
2. Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate
injury to people.
3. Technologies that gather or use information for surveillance violating internationally accepted norms.
4. Technologies whose purpose contravenes widely accepted principles of international law and human rights.
42021 AI Principles Progress Update
Internal governance and operations
At the same time as we announced the Google AI Principles in 2018, we founded a central
Responsible Innovation team, drawing from existing specialists across the company
to operationalize implementation. Starting with half a dozen full time employees,
engagement in AI Principles governance has since grown in scale; today, hundreds of
Google employees across dozens of teams with a wide range of expertise—human rights,
user experience research, ethics, trust and safety, privacy, public policy, machine learning,
and more—make up an internal AI Principles ecosystem that supports employees in
incorporating responsible practices into their work.
Privacy & Data
Escalation Advanced Technology Review Council Protection Office
(If needed) ATRC Steering
Committee
Review Health Ethics Central Responsible Product Area Privacy
Innovation Review AI Principles Review Advisory
processes Committee Committee Committees Council
Dedicated Privacy
User Product
functions in Trust & Safety DEI Councils Working
Experience Inclusion
product teams* Groups
*This is not an exhaustive list, and does not include product-specific teams (e.g., Search Quality)
Google operationalizes responsible innovation practices via
a three-tiered internal AI Principles Ecosystem.
At the core of this internal ecosystem is a three-tiered governance structure. It starts with
our product teams themselves, which include dedicated user experience (UX), privacy,
and trust and safety (T&S) experts who provide deep functional expertise consistent with
the AI Principles.
The second tier is a set of dedicated review bodies and expert teams. The central
Responsible Innovation team is available to support implementation across the company,
and all Google employees are encouraged to engage with the AI Principles review process
throughout the project development lifecycle. Some product areas have set up review
bodies to address specific audiences and needs, such as enterprise offerings in Google
Cloud, hardware in Devices and Services, and medical expertise in Health. In addition,
the Privacy Advisory Council (PAC) reviews all projects for potential privacy concerns,
including (but not exclusively) issues related to AI.
52021 AI Principles Progress Update
One example of a body in this second tier is the Health Ethics Committee. This is a forum
for guidance and decision-making regarding ethical issues arising within the context
of health products, health research, or health-related organizational decisions. It was
created in 2020 to provide a forum for moral deliberation and decision-making on health-
related initiatives to keep Google’s users and products safe with expertise across multiple
domains. The Health Ethics Committee is a multidisciplinary forum that includes subject
matter experts in bioethics, clinical medicine, policy, legal, privacy, compliance, research,
and business. In 2021 the Google Bioethics Program created the Health Ethics Cafe, an
informal forum to discuss bioethical questions with anyone in the company and at any
stage of project development. The most vexing cases from the Health Ethics Cafe are
escalated to the Health Ethics Committee for review.
The second tier also includes review committees tailored for specific product areas. This
includes Google Cloud’s Responsible AI Product and Deal Review Committees, which are
designed to ensure Cloud’s AI products and projects align with the Google AI Principles in
a systematic, repeatable way, and are built with ethics and responsibility by design, across
products and geographies. The first committee focuses on the products built by Cloud
AI & Industry Solutions. These reviews undertake a comprehensive analysis that includes
an evaluation of the sociotechnical landscape, opportunity and harm assessments
across each of the AI Principles, and live discussion with a cross-functional and diverse
committee, resulting in an actionable alignment plan. The second review committee
covers early stage customer engagements leveraging custom AI solutions beyond
our generally available products. It is a committee composed of four cross-functional,
senior executive members. All decisions require full agreement from all four committee
members, and are escalated as needed. Relevant stakeholders across the AI Principles
ecosystem at Google help inform these discussions; this input and involvement is critical
to ensure decisions are not made in a vacuum.
As one example of how these review processes combine to make and build on decisions,
credit risk and worthiness have been noted as an area of concern for algorithmic
unfairness at Google since 2017. In 2019, Cloud’s Responsible AI Product Review
Committee evaluated product opportunities in this space. While our hope is that one
day AI can provide access to credit and play a role in increasing financial inclusion and
health, the Committee ultimately determined that a creditworthiness product—built with
today’s technologies and data—could create disparate impact related to gender, race,
and other marginalized groups, and conflict with Google’s AI Principle to “avoid creating
or reinforcing unfair bias.” In mid-2020 the Product Review Committee re-evaluated
and reaffirmed this decision. Over the course of last year, Cloud’s Responsible AI Deal
Review Committee evaluated multiple proposed custom AI engagements related to the
assessment of creditworthiness. Each engagement is evaluated for its particular use case,
and the Deal Review Committee determined not to pursue many of these. These learnings
over multiple years built on each other, leading to a decision to pause development of
custom AI solutions related to creditworthiness until risks can be appropriately mitigated.
This Cloud-wide policy went into effect last year, and remains in place today.
62021 AI Principles Progress Update
The third tier of our AI governance structure is the Advanced Technology Review Council
(ATRC), a rotating committee of senior product, research, and business executives
representing a broad cross section of Google. The ATRC addresses escalations and the
most complex and precedent-setting cases, and establishes policies impacting multiple
product areas. It has made a number of tough decisions, weighing potential business
opportunities against the ethical risks of certain applications. In 2018 this executive council
affirmed Google Cloud’s Product Review proposal to not offer general-purpose facial
recognition APIs before working through important technology and policy questions, and
advised the team to focus on narrowly focused solutions. Drawing on this decision, and
following a multi-year effort with significant input from internal and external stakeholders,
Google Cloud developed, sought approval from the ATRC, and then released a highly
constrained and specifically designed Celebrity Recognition API2. One topic considered
by the ATRC this year was the development of large language models. Following review,
the ATRC determined that research involving large language models could continue
cautiously, but that no such model should be launched without a full AI Principles review.
72021 AI Principles Progress Update
Resources, research, tools, and
responsible practices
Implementation of the AI Principles is underpinned by a growing body of company-wide
input, training and education resources, technical tools, and recommended practices that
equip Googlers across various functions and roles to develop and deploy AI responsibly.
We’ve created more formalized channels for taking into account a wider variety of
perspectives throughout the product development lifecycle and in our review processes.
Our Responsible AI and Human Centered Technology team, which is growing to 200
experts, is a consolidation of dozens of research and technical teams working in
responsible and ethical AI. The team works with hundreds of partners in product, privacy,
security, and other teams in the company, providing insights through research, tools,
and other technical solutions. Our Product Inclusion team helps product teams tap into
Google’s Employee Resource Groups to test for equitable product experiences with a
range of different users and audiences. We’ve also created an AI Principles Ethics Fellows
program, in which fellows share their perspectives on the responsible development
of future technologies and develop hypothetical case studies to inform how Google
prioritizes socially beneficial applications. In 2021 the fellows created timely case studies
on topics including the responsible development of genome datasets and a COVID-19
content moderation workflow. In late 2021 we complemented this fellowship effort with
another internal program focused on scaling product fairness (ProFair) testing across
Google. This ProFair program consists of more than 70 trusted and trained AI Principles
advisors in 40 offices around the world. Each advisor has received dedicated training on
tech ethics and algorithmic fairness, and engaged in projects such as identifying fairness
concerns in an image dataset for a Next Billion Users product. We also continue
collaborating with human rights experts to conduct Human Rights’ Impact Assessments
for understanding new and emerging technologies.
Our educational resources include the foundational Technology Ethics Training, which
guides Googlers through the philosophy of technology ethics and how to assess potential
benefits and harms, and a suite of courses that explain the Google AI Principles and
internal governance practices. This year, we launched an AI Principles and responsible
innovation training course for new employees in a variety of engineering and UX roles to
understand Google’s ethical charter and available resources. In the two years since we
began internal AI Principles training, these courses have been taken by more than 19,000
Googlers in a variety of roles, including engineers, researchers, product managers, and
account representatives. We continue to expand education to more employees.
In 2021 we also launched interactive online puzzles3 designed to help employees build
awareness of, and test their recall of, the AI Principles. More than 6,000 employees have
engaged with the puzzles in the six months since launch. We continue to experiment
with gamification of our education efforts to improve engagement and retention of key
82021 AI Principles Progress Update
takeaways. We have also piloted a two-day, live-video immersive “Moral Imagination”
workshop for 15 product teams of 6-15 people who are working on a project that may
lead to a potential AI product or feature. Through semi-structured conversation and
various practical exercises, facilitators make new concepts concrete and actionable for
the participating teams by walking them through the ethical implications of their project
and coaching them to see their work from various points of view. The goal is to provide
a pragmatic and constructive reflection of these often unexamined parts of the role of
an engineer, researcher, or product manager. Teams are asked to articulate and interpret
key ethics concepts and processes that relate to their project throughout the workshop,
helping to instill these learnings. Feedback from the pilots has been positive and we will
scale the program next year.
Technical research informs our practices and tooling, and is another avenue for progress.
Since 2018 Google has published more than 3300 research papers on a variety of topics,
including more than 500 on responsible innovation. This year, publications included
the impact of data cascades4 (downstream effects from poorly managed data used
to train high-stakes AI applications), the environmental impact of large models5, and
accountability frameworks6 for the data used in AI systems. In parallel we have developed
a growing library of responsible AI tools, informed by how we put our AI Principles into
practice. We also continue to expand our Responsible AI Toolkit7, launched in 2020; this
year we added hands-on technical tutorials on topics such as privacy in machine
learning8, Know Your Data (KYD)9 (a tool to better understand datasets and improve data
quality), and an updated People + AI Research (PAIR) Guidebook10. We’ve also continued
to improve the Language Interpretability Tool (LIT), designed to help identify and mitigate
bias in language models. LIT Version 0.411, released in November 2021, adds a number of
new features, including support for Google Cloud tools and the ability for developers to
explore tabular and image data.
Researchers are continuing to develop a technical infrastructure approach to supporting
how fairness research is applied in Google products directly. For example, in the last
year, we have been able to apply new, state-of-the-art tools developed and announced
over the last couple of years, such as Fairness Indicators12 (launched in 201913) for model
evaluation and our Min-Diff14 technique (launched in 202015) for remediation to a growing
number of product use cases, enabling the scale of our learned best practices to
proactively address fairness considerations. We are just at the beginning of the journey to
scale these approaches, and are committed to doing so and sharing results responsibly.
92021 AI Principles Progress Update
This screenshot from Know Your Data shows a relationship between words associated with age and
movement. KYD analysis can be used to identify and address potential unfair biases in data labels.
To do so, we continue to develop practices for creating transparency documents around
datasets used for fairness evaluation and training; this year we released new datasets and
data cards—including for avoiding unfair bias in Translate16 and more inclusive annotations
of people17—together with guidance on how to document these datasets responsibly via
the Data Cards Playbook18, launched this year with an interactive external workshop19 at
the 2021 FAccT (Fairness, Accountability and Transparency) conference.
Important progress has also been made on privacy-preserving technologies. For example,
federated learning20, used in products like Gboard, allows models to be centrally trained
and updated based on real user interactions without collecting centralized data from
individual users. Since releasing federated learning in 2017, Google researchers have
developed federated analytics21, which uses similar techniques to get insight into how
product features and models perform for different users without collecting centralized
data. Among other things, federated analytics allows model developers to conduct certain
types of fairness testing on federated systems despite not having access to raw user data,
102021 AI Principles Progress Update
overcoming a significant challenge with federated systems in the past. This year, the team
released federated reconstruction22, a model-agnostic approach to faster, large-scale
federated learning and personalization under privacy and communication constraints.
Researchers have also continued to build on the 2020 GShard results23 for more compute-
and energy-efficient models. This year the team developed GSPMD24, an automatic,
compiler-based, and scalable parallelization system: compilation time stays constant with
an increasing number of devices. The team has also used the combination of pre-training,
self-training, and scaling up model size to enable significant efficiency improvements
in other models25, matching state-of-the-art performance in an automatic speech
recognition model with only 3% of the training data.
Researchers and practitioners at Google continue to use and improve existing tools for
examining fairness. For example, the research team developing novel techniques for an
AI-based model for COVID-19 epidemiology26 published new findings27 this year on using
the What-If Tool (WIT)28 and fairness analysis techniques to identify and address potential
challenges that could impact public health. Another team used KYD to explore gender
bias29 in the COCO Captions dataset, which includes over 300K human-captioned images
used to train models in image labeling and classification tasks.
Finally, we note the importance and responsible practice of continuing to build a
strong and diverse team. Google is committed30 to building a workforce that is more
representative of our users and a workplace that creates a sense of belonging for
everyone. One goal is to improve leadership representation of Black+, Latinx+, and
Native American+ Googlers in the US by 30% by 2025. We’ve already reached this goal
and are on track to double the number of Black+ Googlers at all other levels by 2025. In
addition, all VPs and above at Google are now evaluated on progress in diversity, equity,
and inclusion. These efforts are crucial if we’re going to have a more representative set of
researchers and engineers building future technologies. With respect to representation
in AI, while there are dozens of Black+ and Hispanic/Latinx+ Googlers and hundreds of
female Googlers on our AI teams, we continue to work to expand this representation. This
includes attracting talent and building partnerships in regions with diverse talent pools,
and investing in the external research community through grants, workshops, and other
initiatives to support new voices entering into the field of computer science and more
equitable outcomes.
112021 AI Principles Progress Update
Product impact
Our progress in responsible innovation is reflected in improvements in our products. We
are focused on user-centric design and our mission of universal accessibility and utility.
This goes back to our earliest efforts to use AI to improve our flagship Search product.
For example, in 2011, we announced Panda31, a machine learning algorithm that allowed
us to take the overall content quality of a website into account and adjust its Search rank
accordingly. Ten years later, our Search Quality team continues to leverage AI tools to
improve results for users, address misleading information, and minimize user exposure to
offensive query predictions.
We continue to develop and improve Search responsibly, carefully testing and evaluating
new features to ensure that they are beneficial to our users. A big advance this year was
the announcement of our Multitask Unified Model (MUM)32, which will allow Search to
understand information across a wide range of formats, like text, images, and video, and
draw implicit connections between concepts, topics, and ideas about the world around
us. Applying MUM will not only improve how people around the world find the information
they need, but will also play a role in increasing economic opportunity for creators,
publishers, startups, and small businesses. Ahead of MUM deployment in Search, the
team is carefully analyzing quality gains and losses and examining the impact on a broad
set of queries and specialized slices to identify any unexpected or concerning
performance.
Another area of progress in 2021 was improving the performance of our products
for darker skin tones. The Pixel and Photos teams partnered with a diverse range of
renowned image makers who are celebrated for their beautiful and accurate depictions
of communities of color to help our teams understand where we needed to do better.
With their help, we significantly increased the number of portraits of people of color in
the image datasets that train our camera models. This culminated in our announcement
of Real Tone33, a feature in the Pixel 6 and Pixel 6 Pro, enabling better performance of
features like face detection, auto-exposure, and auto-enhance for users with darker skin
tones.
Accessibility in our products has also improved with the application of AI. In 2021, the
Android team launched an updated version of Lookout34, an Android app developed for
blind and low vision individuals which uses computer vision to provide information about
a person’s surroundings. In June, Lookout v2.3 launched with a much-improved Explore
mode: object identification is now faster and more accurate. At the same time, the
Android team also released the latest version of Voice Access35, including improvements
for entering passwords and the gaze detection beta. This year we also announced Project
Relate36, which uses machine learning to help people with speech impairments
communicate and use technology more easily.
We have continued to refine Privacy Sandbox37, a collaboration with the digital ads
industry to enhance privacy through the addition of AI-related techniques. For example,
one Privacy Sandbox proposal is to use a technology called Federated Learning of
Cohorts (FLoC) to enable the digital ads ecosystem to serve relevant ads to users without
122021 AI Principles Progress Update
tracking user identities across the web. While still a work in progress, Sandbox aspires to
take a significant step forward in delivering a more private user experience while
supporting publishers, advertisers and content creators and preserving the vitality of the
open internet.
Expert teams provide dedicated resources and consultations to product teams on AI
Principles-related issues, so that product teams can learn about—and proactively avoid
or mitigate—common failure modes. The product fairness (ProFair) team, for example,
provides socio-technical advice and conducts proactive algorithmic fairness testing to
ensure new AI technologies do not reflect or perpetuate sociological or socioeconomic
inequalities. The team has also taken on responsibilities for investigating inclusion and
equity of language in data labeling and diversity and representation in training data for
AI. For instance, before launching a new feature called Style AI in Google Shopping38
in countries like India and Brazil, ProFair held local focus groups in collaboration with
Google’s Product Inclusion team. The process helped the Style AI team find diverse
images and clothing for these specific demographics. In the past year, a variety of teams,
including YouTube, Meet, Translate, and Lookout improved projects related to potential
new product features consulted with ProFair testing.
We also continue to evolve internal policies and frameworks to support product teams,
gathering lessons learned from reviews and product engagements to guide teams on how
to think about complex AI Principles challenges. For example, over the past year we have
seen increases in reviews and consultations on surveillance, synthetic media, and affective
technologies. To ensure a consistent and higher-throughput review process we are
piloting a few frameworks on these topics with assessment guidelines, including objective
criteria drawn from precedent centered around the concepts of dignity, autonomy and
consent, guidance from human rights experts, and sample case studies developed by
Google’s AI Principles Ethics Fellows.
132021 AI Principles Progress Update
Supporting global dialogue, standards,
and policy
Google remains committed to engaging with external experts and civil society
stakeholders around the world. Establishing norms and best practices for responsible AI
development will succeed only as a community endeavor, and it is vital to share learnings
and receive feedback on our efforts from the wider community.
This year, we have continued to support programs on inclusion, diversity, and equity in AI
research and product development. We sponsored 114 conferences in computer science
and related fields, including NeurIPS and ICML. And we continue to partner with the
Algorithmic Fairness and Opacity Group (AFOG) at UC Berkeley, an interdisciplinary
research group that bridges disciplinary boundaries to support the equitable
development of AI systems. We continue to host quarterly Equitable AI Research
Roundtables (EARR), launched in 2020 to focus on the potential downstream harms of AI
with experts from the Othering and Belonging Institute at UC Berkeley, PolicyLink, and
Emory University School of Law. In 2021, we co-created with our EARR partners a set of
internal exercises for product teams to apply equitable research practices.
We’re developing new engagements with students who may represent communities
not currently represented in the technology industry. In fall 2021, for example, in the
US, we’ve added sessions focused on responsible innovation topics including ethics in
ML, applying AI principles, algorithmic unfairness, and bias in technology into two of
Google’s classroom programs for Historically Black Colleges and Universities (HBCU),
Google In Residence (GIR) and TechExchange. We continue to expand our engagements
with a growing variety of educational institutions, through ongoing partnerships with
schools such as Berea College, where we host hands-on app-development workshops
focused on building critical thinking skills toward responsible development of advanced
technology. Our Research Scholar39 Program, started this year, gave grants to more than
50 universities in 15+ countries -- and 43% of the principal investigators identify as part
of a group that’s been historically marginalized in tech. Similarly, our exploreCSR40 and CS
Research Mentorship41 programs support thousands of undergrads from marginalized
groups.
We’ve also made a concerted effort to continue hands-on learning opportunities globally.
Our global community educational outreach efforts have included partnerships42 with
dozens of universities around the world to support development of new ML courses,
diversity, and inclusion. We continue to host two-hour interactive digital ML for Policy
Leaders workshops and other related trainings. Since their launch in May 2020, we have
provided ML education to more than 450 policymakers and 150 organizations across
North America, Latin America, Europe, Africa, and the Asia Pacific. And we continue to
share recommended responsible AI practices via TensorFlow43, Cloud44, and Google AI45 to
a range of audiences.
142021 AI Principles Progress Update
Governments have an important role to play in the responsible development of AI, and
we remain committed to sharing our learnings and responsible practices and partnering
with governments to support responsibility across the world’s AI ecosystem. For example,
in addition to providing input on the OECD AI Principles, which were adopted by member
countries in 2019, Google participates in the OECD Network of Experts on AI (ONE AI)
trustworthy AI working group, developing practical guidance around policies that lead to
trustworthy AI. Google also participates in the Global Partnership on AI (GPAI), a multi-
stakeholder initiative established by the G7 which aims to bridge the gap between theory
and practice on AI by supporting cutting-edge research and applied activities on AI-
related priorities.
In addition to international organizations, Google works with national governments around
the world on AI policy. Google partners with governments to support small and medium
enterprises and academics developing responsible AI techniques and services, including
$5M support46 for the US National Science Foundation’s Human-AI Interaction and
Collaboration research center and $2M in Google Cloud Platform credits47 to Portugal’s
Foundation for Science and Technology (FCT) to support research in natural language
understanding and responsible AI.
AI is too important not to regulate, and we are working closely with governments to
establish policies that support innovation while managing risk. This year we provided
submissions48 to consultations from the EU on the draft AI Act and proposed reforms
to the Product Liability Directive for AI, provided feedback on proposed AI legislation in
Brazil, responded to a request for information on AI regulation in Israel, worked with NITI
Aayog to develop AI policy proposals for India, submitted comments on the National AI
Research Resource Task Force, and provided input on NIST’s growing body of responsible
AI frameworks, including the AI Risk Management framework and Explainable AI
framework, among many others. Standards is another key area of partnership and Google
is also an active participant in a variety of efforts, including serving as a founding member
of the International Organization for Standardization (ISO)’s SC42 working group49, and
contributing to the Institute of Electrical and Electronics Engineers (IEEE) AI standards50.
152021 AI Principles Progress Update
Conclusion
Responsible AI at Google has come a long way since we launched our AI Principles in 2018.
We remain committed to these principles, and in the last three years we have expanded
our AI Principles governance and support functions from a small internal “start-up” to a
cross-functional ecosystem.
There is still a lot we don’t know. Foundational questions like how to define fairness for AI
systems are not yet fully answered, and techniques to assess and manage AI risks are still
being refined. AI technology also continues to evolve, introducing new potential benefits
and challenges. Synthetic data, for instance, has the potential to replace the use of
sensitive data in model training, but we are still learning how to generate and use synthetic
data responsibly and safely. Similarly, emergent areas like large multipod models and
affective computing promise exciting new capabilities, but also introduce new risks that
we must learn to understand and manage.
We are improving our ability to implement responsible AI through research, tools,
frameworks, recommended practices, and engagement with a range of experts and
institutions, and will continue to share progress reports (as we have in 2019,51 202052) and
product case studies at ai.google/responsibilities53.
162021 AI Principles Progress Update
End Notes
1. https://blog.google/outreach-initiatives/diversity/racial-equity-update-nov-2021/
2. https://cloud.google.com/blog/products/ai-machine-learning/celebrity-recognition-now-
available-to-approved-media-entertainment-customers
3. https://blog.google/technology/ai/crossword-puzzle-big-purpose/
4. https://dl.acm.org/doi/10.1145/3411764.3445518
5. https://arxiv.org/abs/2104.10350
6. https://arxiv.org/pdf/2010.13561.pdf
7. https://www.tensorflow.org/responsible_ai?hl=ro
8. https://www.tensorflow.org/responsible_ai/privacy/guide?hl=ro
9. https://knowyourdata.withgoogle.com/
10. https://pair.withgoogle.com/guidebook/
11. https://github.com/PAIR-code/lit/blob/main/RELEASE.md
12. https://www.tensorflow.org/responsible_ai/fairness_indicators/guide
13. https://ai.googleblog.com/2019/12/fairness-indicators-scalable.html
14. https://www.tensorflow.org/responsible_ai/model_remediation
15. https://ai.googleblog.com/2020/11/mitigating-unfair-bias-in-ml-models.html
16. https://storage.googleapis.com/gresearch/translate-gender-challenge-sets/Data%20Card.pdf
17. https://storage.googleapis.com/openimages/open_images_extended_miap/Open%20Images%20
Extended%20-%20MIAP%20-%20Data%20Card.pdf
18. https://pair-code.github.io/datacardsplaybook/
19. https://facctconference.org/2021/acceptedcraftsessions.html#data_cards
20. https://federated.withgoogle.com/
21. https://ai.googleblog.com/2020/05/federated-analytics-collaborative-data.html
22. https://arxiv.org/abs/2102.03448
23. https://arxiv.org/abs/2006.16668
24. https://arxiv.org/abs/2105.04663
25. https://arxiv.org/abs/2109.13226
26. https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-is-releasing-the-
covid-19-public-forecasts
27. https://ai.googleblog.com/2021/10/an-ml-based-framework-for-covid-19.html
28. https://pair-code.github.io/what-if-tool/
29. https://ai.googleblog.com/2021/08/a-dataset-exploration-case-study-with.html
30. https://blog.google/outreach-initiatives/diversity/racial-equity-update-nov-2021/
31. https://googleblog.blogspot.com/2011/02/finding-more-high-quality-sites-in.html
32. https://www.blog.google/products/search/introducing-MUM/
33. https://store.google.com/intl/en/discover/realtone/
34. https://support.google.com/accessibility/android/answer/9031274?hl=en
35. https://support.google.com/accessibility/android/answer/6151848?hl=en&ref_topic=6151842
36. https://blog.google/outreach-initiatives/accessibility/project-relate/
172021 AI Principles Progress Update
37. https://privacysandbox.com/
38. https://blog.google/technology/ai/this-googlers-team-is-making-shopping-more-inclusive/
39. https://ai.googleblog.com/2021/04/announcing-2021-research-scholar.html
40. https://research.google/outreach/explore-csr/
41. https://research.google/outreach/csrmp/
42. https://blog.tensorflow.org/2021/06/2021-request-for-proposals-ml-faculty-awards.html
43. https://www.tensorflow.org/responsible_ai
44. https://cloud.google.com/responsible-ai
45. https://ai.google/responsibilities/
46. https://blog.google/technology/ai/partnering-nsf-human-ai-collaboration/
47. https://portugal.googleblog.com/2021/07/2-milhoes-de-dolares-para-apoiar.html
48. https://ai.google/responsibilities/public-policy-perspectives/
49. https://www.iso.org/committee/6794475.html
50. https://standards.ieee.org/initiatives/artificial-intelligence-systems/index.html
51. https://ai.google/static/documents/ai-principles-2019-progress-update.pdf
52. https://ai.google/static/documents/ai-principles-2020-progress-update.pdf
53. https://ai.google/responsibilities
18

****************************************************************************************************