BCS, The Chartered Institute for IT—written evidence (LLM0094)
House of Lords Communications and Digital Select Committee inquiry:
Large language models
We acknowledge the incremental evolution of LLMs, characterised by a trend
towards smaller, more accessible models that offer personalisation and
integration into various applications. However, we emphasise the
importance of addressing the environmental and economic challenges posed
by the intensive training requirements of LLMs, which may limit their
widespread adoption. BCS also underscores the need for responsible
funding, monitoring, and ethical considerations to navigate the risks
associated with LLMs, including the potential for misinformation and biased
responses. Ultimately, with careful management, LLMs can reshape various
domains, offering opportunities and challenges that require ongoing research
and ethical usage.
Question 1: How will large language models develop over the next three years?
In the upcoming three-year period, the evolution of large language models
(LLMs) is expected to follow a path marked by incremental changes,
although the precise course remains uncertain. LLMs have been undergoing
a reduction in size, while concurrently expanding in capabilities. The size of
an LLM is significant as it affects both energy efficiency and processing
speed. The trend towards smaller models is gaining traction due to their
heightened accessibility, decreasing reliance on specific providers such as
Microsoft or OpenAI.1
These compact LLMs offer personalisation,2 allowing responses to be tailored
to individual user requirements, and they can be seamlessly integrated into
operating systems. Given that such pre-trained models are based on
Machine Learning (ML), they possess the capacity to continue to learn from
new data gathered about a user’s preferences and interests. This potential
for personalisation means smaller LLM models can be bundled in with an
operating system (OS), increasing the capabilities of personal devices. It is
worth noting, however, that while LLMs like those underpinning ChatGPT
have been trained on a wide breadth of general- purpose knowledge
sources, their capability is limited when delving into specific domains. For
example, while they may be able to answer questions about general human
biology, they would struggle to accurately understand, interpret and
diagnose specific medical conditions or do non-trivial arithmetic.3 This
limitation gives rise to the possibility of specialised models focused on
distinct topics, heralding an era where numerous models excel in specific
areas, rather than relying on a single generalist model.
Despite their advantages, LLMs present challenges due to their demanding
training requirements, which can be costly and create barriers to entry. LLMs
1 [Link]
2 [Link]
3 [Link]
are trained by feeding them vast amounts of data, teaching them to identify
patterns and trends and use that information to reach a desired outcome.
This process requires processing vast amounts of data in data centres, which
comes at a great economic and environmental cost.4
Only the most well-resourced actors can explore the boundaries of LLMs at
any significant scale. One reason is that many organisations lack enough
Graphical Processing Units (GPUs) – specialised computer chips which,
among other things, help computers handle video games, cryptocurrency,
mathematical and ML-based applications (like LLMs) more efficiently. There
are ‘GPU-Rich and GPU Poor’5 organisations, which is a barrier to
innovation.
Training algorithms, retaining data and the uncurbed growth of ML models
like LLMs is already proven to be having detrimental impacts on the
environment, and adverse impacts on the global effort to reduce the
emission of greenhouse gases.6
The development patterns with LLMs often see gradual progress, with an
occasional leap that elevates their capabilities. LLMs build upon existing
neural network architectures from 1958,7 with a big leap forward in 2017
with the invention of transformer models.8 These are ML techniques that
learn statistical patterns from data rather than explicit logic. There is also
another branch of AI research, symbolic AI, which seeks to categorise and
build interpretable rules and models based on knowledge and logic.
ChatGPT uses LLMs, lots of conventional software code, and apparently also
some symbolic AI. The LLM recognises patterns from extensive examples
but lacks any understanding of concepts founded in a knowledge base. The
use of other technologies in combination with LLMs are the breakthrough
represented by ChatGPT and similar systems.
Elements of symbolic AI are likely to have been integrated to improve
ethical responses, including reducing bias. This means that it has been
given rules that, for example, make it refrain from giving answers that
perpetuate racism, sexism and other forms of discrimination, despite the
fact that the data used to train it will have undoubtedly contained elements
of these forms of bias. Reinforcement learning is also incorporated to
improve an LLM’s capabilities based on feedback in the form of human
responses.
The future development trajectory remains uncertain, and refining the
techniques employed by LLMs could lead to more accurate responses and
enhanced technology.
4 [Link]
5 [Link]
6 [Link]
7 [Link]
8 [Link]
2
Improving Understanding of Future Trajectories
Question 1a: Given the inherent uncertainty of forecasts in this area, what
can be done to improve understanding of and confidence in future
trajectories?
Given the inherent uncertainty associated with predicting developments,
enhancing our comprehension of future trajectories calls for focused actions.
Streamlining funding processes for AI research and more technical support
like increasing access to GPUs could empower researchers to operate more
efficiently. Additionally, investing in local technology talent and providing
tech-specific funding for AI and LLMs can nurture innovative businesses and
reduce dependency on foreign sources like the US and China.
Efficient monitoring of LLM and AI development, along with responsible
funding practices, is essential. To foster meaningful discussions,
establishing a platform for scientifically informed discourse rather than
anecdotal exchanges is vital.
Acknowledging the emergence of abilities in LLMs and redefining success
metrics can provide deeper insights into their progress. While leaps in
development are unpredictable, supporting continuous incremental
improvements can contribute to a better understanding of future
trajectories.
Greatest Opportunities and Risks
Question 2: What are the greatest opportunities and risks over the next three
years?
Anticipated over the next three years are substantial opportunities and risks
within the realm of LLMs. One significant risk is the potential for
"hallucination", which is when LLMs generate false information, incorrectly
combining data sources, or generating randomised tokens that lead to
factual inaccuracies. This phenomenon, observed during domain-specific
tasks like generating legal cases9 as well as general enquiries about
Australian mayors10 or American law professors,11 underscores the
challenges posed by the sequential nature of automated generation.
Dependence on major tech companies and presumed reliability of said
companies can also pose risks, as demonstrated by the lawyer case cited
above. Caution is advised in anthropomorphising scenarios, as LLMs are
trained to defer to human expertise, potentially reinforcing user biases.
Concerns about data breaches and inadequate regulation, exemplified by
the Italian ban on data storage, further highlight the need for careful
consideration of data handling.
9 [Link]
10 [Link]
first-defamation-lawsuit-over-chatgpt-content
11 [Link]
[Link]
3
The rise of mis- and disinformation and content generated by LLMs poses
significant societal risks, potentially exploited by malicious actors to
manipulate public opinion. The potential harms arising from misuse remain
largely uncharted territory.
Furthermore, the integration of LLMs across all educational settings raises
questions about assessment methods and authentic evaluation.
A lack of understanding of AI risks scaremongering, ‘fake news’, and poor
decision- making. On the other hand, there is an opportunity to roll out a
programme of education in AI literacy for everyone, rather than just
computer scientists as has been the norm until now. Such a programme
could cover how to use AI, understanding how AI works, and the ethical
and societal implications of AI.
Balancing Risk Considerations
Question 2a: How should we think about risk in this context?
Considering risks within their context offers insights into potential
outcomes. LLMs hold promise in enhancing efficiency and automating
routine tasks. However, the impact on employment remains uncertain and
dependent on existing expert knowledge to effectively utilise these
technologies.
The emergence of new businesses and job opportunities is possible as LLMs
streamline processes. Rather than erasing jobs, LLMs could reshape roles by
automating routine tasks and enabling humans to focus on more complex
aspects of their work. Nurturing innovation and staying ahead of LLM
advancements is advised to avoid playing catch-up.
“Automation Bias”, or the tendency of humans to believe the machine, is a
major risk of LLMs, and we have to insist that information we rely on must
be backed by verified citations (as the legal case showed).
Question 3: How adequately does the AI White Paper (alongside other
Government policy) deal with large language models? Is a tailored
regulatory approach needed?
BCS, along with The Royal Statistics Society, National Physics Laboratory,
the Alan Turing Institute, The Operational Research Society and the Institute
for Mathematics and its Applications is part of the Alliance for Data Science
Professionals. This means that registration, professional standards and
responsible computing form the basis of our approach to engage with AI
safely in general.12
The government’s UK AI Regulation White Paper highlights five principles
for UK regulators to achieve responsible AI in their specific sectors. Section
4 recognises a central role for tools in trustworthy AI, including technical
12 Helping AI grow up without pressing pause: [Link]
research/helping-ai- grow-up-without-pressing-pause/
4
standards and assurance techniques to help implement the principles.
The technical and operational standards provide management systems,
processes and measurement methods to support implementation of the five
principles.
Organisations should align themselves with these best practices and ensure
that their AI technology providers also adopt them as and when they are
published.
While significant advances have clearly been made in LLMs, it is not entirely
clear that this rapid pace of development will continue, although we expect
that LLMs will become more integrated with other tools and become
increasingly multi-modal (for instance, consuming and generating images
and sound in addition to text).
The concerns from most experts are not that AI is too powerful, but that
basic guard rails are not in place to ensure AI is deployed responsibly. It
should be noted that the guard rails that need to be put in place are
governance-based rather than being technical innovations.
The Institute broadly welcomes the government’s regulatory proposals but
with caveats.13 While the light touch approach is positive in that it enables
innovation, there are areas that need more consideration to ensure the
proposals maximise the public benefit of AI. The proposed cross sectorial
principles are appropriate and useful, but should be extended further.
For example, all AI systems must have appropriate safeguards to ensure
they remain technically sound and are used ethically under reasonably
foreseeable exceptional circumstances, as well as under normal
circumstances. Organisations must show they have properly explored and
mitigated against reasonably foreseeable unintended consequences of AI
systems.
In a sense, we view ChatGPT (and other LLMs) as a distraction in relation
to increasing the reasoning power of AI systems. While LLMs are AI systems
that can generally respond to human queries well, they do not represent a
significant improvement in reasoning ability. We acknowledge that LLMs do
appear to be able to reason better on first glance, but that is because they
have illusory anthropomorphic qualities.
At this point it is important to keep monitoring the consequences and
impact of LLMs, however as this is an emerging technology it is too early to
say if we need tailored regulatory approach for this specific use of AI.
13 Light touch approach to AI regulation welcomed by IT industry body:
[Link] and-research/light-touch-approach-to-ai-regulation-
welcomed-by-it-industry-body/
5
Question 4. Do the UK’s regulator have sufficient expertise and resources to
respond to late language models? [5] If not what should be done to address
this?
As with most emerging technologies, there is a steep learning curve. We
would suggest that all stakeholders have a basic understanding of the
ethical considerations around AI and LLMs, and the challenges and
opportunities this technology brings.
As the professional body for information technology BCS’ remit is the
development of digital skills and literacy, promoting the ethical use of
technology, and the value that competent professionals from diverse
backgrounds bring when delivering change.
We would welcome the opportunity to discuss further with the government
how we can help.
Long-Term Implications and Responsible Usage
The development of large language models in the next three years remains
uncertain yet promising. Their evolution holds the potential to reshape
various domains, offering both opportunities and challenges that require
careful consideration, ethical usage, and ongoing research.
Thanks to the following for their contributions to this consultation:
Adam Leon Smith FBCS - Chair of BCS, The Chartered Institute for IT
Fellows Advisory Group (F-TAG) and CTO Dragonfly; Professor James
Davenport FBCS FIMA CITP CMath; Professor Adrian Hopgood FBCS (BCS AI
Specialist Group)
About BCS
The purpose of BCS, as defined by its Royal Charter, is to promote and
advance the education and practice of computing for the benefit of the
public. We bring together industry, academics, practitioners, and
governments to share knowledge, promote new thinking, inform the design
of new curricula, shape public policy and inform the public.
As the professional membership and accreditation body for Information
Technology, we serve 70,000 members, including practitioners, businesses,
academics, and students, in the UK and internationally. We accredit the
computing degree courses in ninety-eight universities around the UK and as
a leading IT qualification body, we offer a range of widely recognised
professional and end-user qualifications.
September 2023
