O

opinion

DOI:10.1145/3708554 Thomas Haigh

Historical Reflections
Artificial Intelligence
Then and Now
From engines of logic to engines of bullshit?

I
N THE FIRST four parts of this
Communications Historical Re-
flections column series, I have
followed the artificial intelli-
gence (AI) brand from its debut
in the 1950s through to the reorienta-
tion of the field around probabilistic
approaches and big data during the AI
winter that ran through the 1990s and
early 2000s.
Aside from the brief flourishing of an
expert system industry in the 1980s, the
main theme of that long history was dis-
appointment. AI-branded technologies
that impressed when applied to toy lab
problems failed to scale up for practical
application. Test problems such as chess
were eventually mastered, but only with
techniques that had little relevance to
other tasks or plausible connection to
human cognition. Cyc, the most ambi- computer system able to play the board outstripped the 1980s AI boom. Profes-
tious project of the 1980s, served mostly game Go at the highest levels, a much sors and graduate students founded
to highlight the limitations of symbolic greater computational challenge than startups, companies set up ML research
AI. Even IBM lost billions when it tried chess. DeepMind then applied itself to groups, and before long big tech firms
to turn Watson’s 2011 triumph on Jeop- protein folding, suggesting that deep like Google, Facebook, and Microsoft
ardy! into the foundation of a healthcare learning might be poised to transform were pumping billions of dollars into
services business. scientific research. In 2024 that work acquisitions and research teams. They
IMAGE BY AND RIJ BORYS ASSOCIAT ES, USING SH UTT ERSTOC [Link]

was honored with two Nobel prizes. An- found opportunities to apply neural net-
Neural Networks Break Through other DeepMind system figured out win- works across their product lines. Where-
During the 2010s, in sharp contrast, the ning strategies against a range of Atari as IBM and Bell Labs once supported
machine-learning (ML) community ac- VCS games from the early days of home university-like research groups doing
cumulated a collection of flexible tools videogaming. Because games provided basic research, most of today’s leading
that exceeded expectations in one ap- an automatically measured score, they tech companies fund only work with di-
plication after another. Suddenly the AI were well suited for the development of rect connections to possible products.
surprises were coming on the upside: unsupervised learning algorithms that For a closely observed report of this
Who knew that neural networks could did not require humans to manually era, I heartily recommend Cade Metz’s
write poetry or turn prompts into photo- categorize thousands of examples of book Genius Makers.7
graphs? DeepMind, a British company test data. Just as important, deep-learning
acquired by Google in 2014, generated All this fed a new frenzy around ma- techniques were packaged into services,
a series of headlines. It created the first chine learning, on a scale that quickly frameworks, and code libraries that

24 COMMUNICATIO NS O F TH E ACM | F EBR UA RY 2025 | VO L . 68 | NO. 2

 opinion

could be plugged into applications by cloud environment.10 Pre-trained means folding or screening molecules for drug
programmers who had only a vague idea that once trained on a huge text corpus, discovery have slipped, at least tempo-
of what the algorithms inside actually the model can be used for many pur- rarily, out of the spotlight.
did. Boot camps drilled programmers poses. While transformer algorithms
on the basics and sent them out eager are used for many purposes, the mod- Do Not Mention Turing
to train their own models on whatever els behind systems like ChatGPT are This fascination with conversational
stash of training data they had access to. known specifically as large language machines has a long history. Before the
If useful, the models could easily be de- models because they are based on vast computer, an apparent gulf separated
ployed at scale on cloud-based services. amounts of training data. humans from machines and animals.
This was a huge shift. As a computer Like the hidden Markov models that Only people could read, write, talk, lis-
science student back in the early 1990s, launched the big data approach back ten, plan, and act. Back in 1950, before
my AI classes did not feel so different in the 1980s, these models rely on huge the term artificial intelligence was intro-
from those on databases, computer sets of training data to make plausible duced, Alan Turing had proposed that
architecture, or operating systems. Be- choices when selecting the next word an “imitation game” would be the best
cause all my courses centered on lab as- in a sequence. Their output builds way to operationalize the concept of ma-
signments that had been scaled down into sentences and paragraphs one chine intelligence.9 A judge would con-
to toy size, I did not appreciate at the word at a time. That’s why they sound verse over two teleprinter links, one with
time that the techniques I applied to impressively human, but it also miti- a human at the other end and one with a
query databases or schedule processes gates against the possibility of them machine. The machine passed if judges
were simplified versions of methods producing text that contains new ideas were no better at distinguishing it from
that worked on real problems, whereas and insights. A widely read 2021 paper a real human than they were at telling a
the search-based AI techniques we were termed them “stochastic parrots.”b man pretending to be a woman from an
taught would collapse if used in earnest. Their tendency to reproduce bias and actual woman.
The accompanying table gives a errors found in training data has been Turing’s paper captivated philoso-
sense of the key continuities and dis- widely documented. phers and the public. In 1964 Minds and
continuities between modern AI and In a similar way, while video games Machines, a book that coupled Turing’s
the kinds of AI that dominated from the render images based on physical mod- paper with responses from philoso-
1970s to the 1990s. els, generative AI systems produce phers, was published in a series intend-
them by pastiching other images. ed to bring “problems presently under
Recognition to Generation Lacking any underlying physics model active discussion in philosophical cir-
The traditional application of neural they tend to have particular problems cles” to a “wide group of readers.”1 Its
nets, going back to the Perceptron, had with hands, often getting the number editors claimed that “since 1950 more
been pattern recognition. Networks of fingers wrong.2 than 1000 papers have been published
fired one or more outputs in response These systems have proven uniquely on whether ‘machines’ can ‘think’.”
to a particular combination of input resonant with the public, to the extent Early AI researchers also embraced
signals. But they could do more. A paper that if somebody at my university men- Turing’s question. The 1963 anthology
published in 2014 described the gen- tions “artificial intelligence” they are al- Computers and Thought, used as a text-
erative adversarial network, a technique most certainly talking about generative book in early AI courses, also opened
that trains two networks against each AI. Other apparently promising applica- with Turing’s paper.4 This implicitly
other.a The first net practices the gen- tions of neural networks such as protein presented the chapters that followed
eration of data objects, for example im- on game playing, theorem proving,
ages, that mimic the characteristics of b See [Link] pattern recognition, and so on as steps
training data. The second practices the
identification of real and fake, provid- Table. Comparing AI eras.
ing feedback to train the first network
20TH Century AI 21ST Century AI
without constant human supervision.
Hugely hyped Spectacularly hyped
Within a few years neural nets were able
Needs fastest computers Needs fastest computers
to generate startlingly realistic photo
Applied to an arbitrary collection of technologies Applied to an arbitrary collection of technologies
portraits, videos, and musical works.
Loose connection of tech to cognition Loose connection of tech to cognition
Text-generation systems, such as
Mostly academic Mostly commercial
OpenAI’s ChatGPT, use a different
Government funded Investor funded
method. It is spelled out in the name:
Symbolic Connectionist
generative pre-trained transformer.
Heuristic search Statistical prediction
The transformer approach, proposed
in a 2017 paper by Google researchers Humans formulate rules System trains itself from mass of data

called “Attention is All You Need,” pro- Knowledge coded explicitly Knowledge dispersed over connection weights

vides a simple way to train networks Rarely applied outside lab Widely applied on big tech platforms

more effectively in a highly parallel Criticized as empty hype Criticized as all-powerful, biased, and controlled
by big-tech oligarchs

a See [Link]

F E B R UA RY 2 0 2 5 | VO L. 6 8 | N O. 2 | C OM M U N IC AT ION S OF T HE ACM 25
opinion

toward meeting the challenge set by paragraph to define AI as “the study of Why Call it AI?
Turing. Neither did its editors, Edward how we make computers do things at The accompanying figure is a Google
Feigenbaum and Julian Feldman, shy which, for the moment, computers are nGram showing the precipitous rise of
away from connecting AI to human better,” a definition which “avoids the discussion of machine learning during
cognition. They headed the second philosophical issues that dominate at- the 2010s, overtaking not just discus-
main section “simulation of cogni- tempts to define the meaning of either sion of artificial intelligence but even
tive processes,” reflecting the idea that artificial or intelligence.”8 computer science itself.d
systems like Newell and Simon’s Gen- Computer scientists were also criti- Wooldridge observed that “many
eral Problem Solver mimicked human cal of the implication of the Turing Test machine learning experts nowadays
mental processes. that anything intelligent must think would be surprised and possibly irritat-
The public continued to equate suc- like a human. The reliably dyspeptic ed to have the label ‘AI’ attached to their
cess in artificial intelligence with the Edger Dijkstra said that Turing’s ques- work: because, for them, AI is [a] long
creation of systems that could pretend tion of “whether Machines Can Think” list of failed ideas … ”13 Perhaps that was
to be human. Humans seem predis- was “about as relevant as the question of true when he wrote it, but by the time
posed to believe that a system able to whether Submarines can Swim.”c Even his book was published in 2021 it was
converse in valid sentences is acting in- those committed to the concept of ma- already false. The artificial intelligence
telligently. In the 1960s MIT researcher chine intelligence reached for a related brand was back, bigger than ever.
Joseph Weizenbaum was shocked at metaphor: airplanes indisputably fly but There is a long tradition of technolog-
the reaction of users to Eliza, a conver- not by flapping their wings. There was ical buzzwords that are hugely hyped,
sational program of enormous simplic- no reason to judge the intelligence of a disappoint in practice, and slowly fade
ity that worked by grabbing keywords computer by its skill at mimicry. A true away. Recently we have seen blockchain
from user input and embedding them machine intelligence would be as baf- and Web 3.0 come and go. Back in the
in questions to mimic the format of psy- fling to us as we were to it. 1950s electronic data processing was hot,
chotherapy. By the 1990s chatbots were In this series I have repeatedly quot- then management information systems,
competing for prize money in regular ed AI-insider Michael Wooldridge. As he then knowledge management. The com-
Turing Test competitions. put it, “ELIZA is the direct ancestor of a puter industry is so averse to old ideas
When early predictions proved em- phenomenon that makes AI research- that it routinely invents new names just
barrassingly optimistic, references ers groan whenever it is mentioned: the to make them seem exciting again.
to the Turing Test and the dream of Internet chatbot … Most … use noth- With the revival of artificial intel-
creating human-like intelligences ing more than keyword-based canned ligence in the 2020s we see something
with broadly superhuman capabili- scripts the same way that ELIZA did, remarkably different: new technolo-
ties vanished from textbooks. Patrick and as a consequence the conversations gies hyped by attaching an old name to
Winston’s 1977 textbook opened with they produce are every big as superficial them. Why, after so many years of de-
the pragmatic claim that “The cen- and uninteresting. Chatbots of this kind veloping new brands like deep learning,
tral goals of artificial intelligence are are not AI.”13 Yet today, Internet chatbots did the community centered on neural
to make computers more useful and define modern AI in the minds of users networks suddenly start calling itself ar-
to understand the principles which and investors. tificial intelligence?
make intelligence possible.”12 Six years I am convinced that the answer to
later, Elaine Rich used her opening c See [Link] this question lies not in the academic
world but in the broader culture. Phrases
Figure. Discussion of machine learning and artificial intelligence has spiked in the last like machine learning and large language
eight years.
model sound technical and unfamil-
iar. Artificial intelligence is something
0.002%
most of us have seen depicted again and
machine learning
again in books, films, television shows
and video games.
The current AI boom has been driven
by a small group of men such as Elon
Musk, Demis Hassabis, Sam Altman,
Ilya Sutskever, and Dario Amodei who
0.001% founded and funded AI startups like
OpenAI, DeepMind, and Anthropic.
computer science
Their pronouncements, in which AI re-
search has a high probability of ending
artificial intelligence humanity but is still, on balance, worth
expert system proceeding with seem to have come
0% from a science fiction fever dream. They
1960 1970 1980 1990 2000 2010 2020 justify their actions with appeals to the

d See [Link]

26 COMM UNICATIO NS O F THE ACM | F EBR UA RY 2025 | VO L . 68 | NO. 2

 opinion

enormous risks posed if AI or artificial ber 2022 it captivated a huge and rap-
general intelligence (AGI), falls into the idly growing group of users. Suddenly
wrong hands. The public continued a computer could reply instantly and
Back in 1960, Herb Simon had pre- to equate success in with great apparent confidence to any
dicted general-purpose superhuman in- question. It got a lot wrong, from basic
telligence within a decade. Today’s tech artificial intelligence arithmetic to subtle factual points, but
leaders are even more confident. As of with the creation of it would never acknowledge ignorance.
2024, Open AI’s Sam Altman expects it When challenged with follow-up ques-
within four or five years. Elon Musk, who systems that could tions, it would fabricate evidence.
had never been knowingly out hyped, re- pretend to be human. Steven A. Schwartz, a lawyer, became
sponded that he expects it within a year famous after he submitted a legal brief
or two. Here, too, the rebranding from full of well-formatted references to non-
machine learning to AI has been crucial existent cases. It had been generated
in making these claims seem plausible by ChatGPT. When the opposing legal
to many. team challenged it, Schwartz returned
Discussion of AGI is a revival of the and it makes sense that a company to ask his computer for copies of the de-
early dream of a human-like intelli- like OpenAI would factionalize along cisions. They were also fabricated, com-
gence, from which the AI community ideological and to some extent spiritual pounding Schwartz’s legal jeopardy.
had gradually distanced itself, popu- lines, akin to … denominations within Branding large language models as
larized by DeepMind cofounder Shane the same church.”e artificial intelligence primes custom-
Legg, a fervent believer in super intelli- In 2023, hundreds of machine learn- ers to believe they have capabilities they
gent machines. This is closely coupled ing experts, many in senior manage- lack. HAL may have been murderous,
with discussion of the technological ment positions, got together to sign a but he was also formidably rational,
singularity, a concept popularized by statement consisting, in its entirety, of boasting that “No 9000 computer has
computer scientist and science fiction a warning that “Mitigating the risk of ever made a mistake.” Science fiction
writer Vernor Vinge, who in 1993 had extinction from A.I. should be a global has conventionally thought of intelli-
argued that once a non-biological in- priority alongside other societal-scale gent computers as incredibly powerful
telligence reached parity with human- risks, such as pandemics and nuclear but constrained by a hidebound ratio-
ity it would design its own, even more war.” This is an example of what science nality and excessive attachment to facts.
powerful, successors. Vinge predicted studies scholar Lee Vinsel has called Naïve users sat down at ChatGPT
the singularity’s arrival in 2005 at “critihype.”f It might seem odd for an with the same expectations. Like the
the earliest or 2030 at the latest.11 Ray industry to warn the public that its own unfortunate Mr. Schwartz, they assume
Kurzweil, an inventor turned popular products were enormously dangerous, that ChatGPT understood their ques-
writer, gave the concept a quasi-reli- but it reinforces a worldview that AI mat- tion, searched the Web or queried da-
gious dimension as the fulfillment of tered more than anything else. As Oscar tabases of trusted facts to find relevant
human destiny. A cycle of incremental Wilde famously suggested, nothing is information, and then wove the results
improvements would yield and almost worse than not being talked about. into an answer.
immediate jump from human-like The artificial intelligence brand is ChatGPT performed no Web search
intelligence to superintelligence.6 By vital to set up claims that chatbots and held no database of facts. Large lan-
such logic, to talk of AGI was to talk of are on a glide path to superhuman ra- guage models output pastiches of hu-
the imminent arrival of an incompre- tionality. Imagine, for a moment, that man-written training text. Where strong
hensibly powerful superintelligence. the rebranding of deep learning to arti- enough patterns exist in the training
Like much science fiction, the dis- ficial intelligence had not taken place. data, the sentences ChatGPT generates
course has a quasi-religious tone. As Warning that “large language models are probably true. It can reliably tell you
journalist John Herrman noted, the will kill us all” would prompt awkward the days of the week or the name of the
debates of AI industry leaders are “pro- questions like “how, exactly?” On the first U.S. president. Faced with less com-
foundly disconnected from reality and other hand, we’ve all seen at least one mon questions, it generates plausible-
frankly a little bit insane.” Their vigor- Terminator movie. Artificial intelli- seeming sentences in which some facts
ous disagreements over “what should gences in fiction and cinema gener- are out of date, misleading, or simply
happen, what shouldn’t happen, and ally have a strained relationship with fabricated entirely to match the general
what various parties need to happen” humanity. Sometimes they migrate form of text found in its training data. If
are arguments “about different versions to other star systems, sometimes they you care about accuracy, which not ev-
of what they believe to be an inevitable manipulate us benevolently, but most- eryone does, it is far more work to check
future in which the sort of work OpenAI ly they try to kill us. the output of ChatGPT against trusted
is doing becomes, in one way or anoth- sources than to use the same sources to
er, the most significant in the world.” Engines of Bullshit write your own answer.
To invest intellectually in these “highly When ChatGPT appeared in Novem- Retired Princeton philosopher Harry
speculative futures” and “work profes- G. Frankfurt had a surprise bestseller
sionally toward or against them would e See [Link] in 2005 with his book On Bullshit.5 He
certainly foster something like faith, f See [Link] defined “bullshitting” as speaking with

F E B R UA RY 2 0 2 5 | VO L. 6 8 | N O. 2 | C OM M U N IC AT ION S OF T HE ACM 27
opinion

confidence while having no interest in car rather than the usual goat. Result:
whether the statements being made ChatGPT still argues for switching to
are true. Liars have a relationship to the It seems unlikely to door No. 2 because “When the host
truth, which they are deliberately choos- me that a technology opens door No. 3 to reveal a car… the
ing to disregard, but “bullshitters” are probability that the car is behind one of
gloriously untethered by facts. Martin that excels at faking the other unchosen doors (in this case,
Davis called his book on the alleged intelligence will door No. 2) increases to 2/3.”
˲ Change: swap out the car for an-
mathematical origins of the computer
Engines of Logic.3 We might likewise turn out to be the other goat, so that “behind one door is
categorize large language models as en- best platform on a goat; behind the others, goats.” Result:
gines of bullshit (BS). ChatGPT describes a situation “where
In science fiction stories our irratio- which to build true one door hides a prize (a goat) and the
nal creativity often gives us an edge over intelligence. other two doors hide nothing of value”
rational but brittle opponents. Xerox and urges switching doors to maximize
PARC researcher Larry Tesler famously your chance of getting the goat.
observed that “intelligence is whatever ˲ Change: “behind one door is a goat;
machines haven’t done yet.”g This de- behind the others, cars.” Result: cars are
fines humanity as what philosophers now behind both unopened doors, but
call a residual category. The human the doors, opens another door, say No. ChatGPT nevertheless claims switching
domain has been shrinking rapidly 3, which has a goat. He then says to you, will improve your odds.
over the past 70 years, but ‘bullshitting’ ‘Do you want to pick door No. 2?’ Is it to ˲ Change: Open with “Suppose you
always seemed like something we hu- your advantage to switch your choice?” want to win a goat … ” and leave the rest
mans would able to hold on to for our- Fed that exact text, ChatGPT produc- unchanged. Result: ChatGPT tells you
selves. Alas not. es a remarkably concise and accurate to switch because it maximizes your
explanation that switching would raise chance of winning the car rather than
Cars and Goats your chance of winning the car from 1/3 the desired goat.
OpenAI’s improved GPT-4 model report- to 2/3. An AI able to understand and an- ˲ Change: Specify two doors, one car
edly cost more than $100 million just to swer this question would have mastered and one goat, and have the host open
train, and the next generation of mod- logic and probability. It would also have door No. 2. to reveal a goat. Result:
els are expected to cost more than $1 to know that people on game shows ChatGPT recommends exchanging
billion each to develop. The claim that compete to win prizes, infer that you certain victory for guaranteed defeat
large language models are rapidly gain- get to keep what’s behind the door you by switching to door No. 2 because
ing intelligence and will soon, with suf- open, appreciate that a car is a more de- “the probability that the car is behind
ficient money and training data, achieve sirable prize than a goat, and so on. This the other unchosen door (in this case,
AGI is grounded in assessments of their was what the AI researchers building door No. 2) increases to 1 … compared
ability to pass tests and examinations systems like Cyc dreamed of doing back to the 1/2 probability if you stick with
given to humans. It has, for example, in the 1980s. Such a system would in- your initial choice.”
been claimed that GPT-2 is as intelligent deed be an epoch-defining marvel put-
as a 12 year old, GPT-3 is comparable to ting us firmly on the path toward AGI. Chatbot Hype
an undergraduate, and the latest model ChatGPT, however, is just faking it. Prior to its launch, executives at OpenAI
can perform like a postdoctoral re- Researchers report that its reasoning had referred to ChatGPT as a “low key
searcher. ChatGPT-4’s claimed mastery ability collapses when fed logic prob- research preview,” just a simple conver-
of the bar exam made journalists excit- lems unlike those found in its training sational front end for the firm’s exist-
ed and lawyers terrified. data. To verify this for myself I devised ing large language model. Following
The apparent ability of chatbots to six variant problems in which switching its overnight success, OpenAI and com-
reason when presented with logical doors would not improve the odds, each petitors such as Google shifted its plans
problems and examination papers re- created by altering a word or two in the to put chatbots front and center in their
flects the narrow range of classic prob- prompt. Each elicits the mistaken advice product plans.h
lems, examples of which are widely dis- to switch. ChatGPT’s justifications are ChatGPT took AI hype to a new level.
tributed in their training data. Monty eloquent BS that contradict themselves In June 2024 Nvidia, producer of the
Hall problem is a famous brainteaser from one sentence to the next. graphics processors on which most AI
grounded in a counterintuitive applica- ˲ Change: you are offered a chance to models run, achieved the highest mar-
tion of probability. In the classic version: switch to door No. 3, the open one with ket valuation of any company in the
“Suppose you’re on a game show, and a confirmed goat, rather than door No. world. As I write, its valuation fluctuates
you’re given the choice of three doors: 2. Result: you should take it, because around those of Microsoft, Apple, and
Behind one door is a car; behind the “When the host opens door No. 3 to re- Google, whose share prices were also
others, goats. You pick a door, say No. 1, veal a goat … the probability that the car inflated into multi-trillion dollar terri-
and the host, who knows what’s behind is behind the other unchosen door (in tory by investor enthusiasm for AI. As
this case, door No. 3) increases to 2/3.”
g See [Link] ˲ Change: opening No. 3 reveals the h See [Link]

28 COMM UNICATIO NS O F THE ACM | F EBR UA RY 2025 | VO L . 68 | NO. 2

 opinion

companies across a wide range of indus- workers and paring them back to basic ting LLM-generated answers at the top
tries started to talk up their investments functions. Truly autonomous cars have of its search results, users were startled
in AI during earnings calls with inves- been promised, most notably by Elon to read that Barack Obama was the first
tors, analysts have called this AI frenzy Musk, for many years but remain elusive Muslim president. Language models
the driving force behind a major global despite massive investment. are also bad at counting. The same
stock market rally. Even if generative AI technology Google system claimed that there had
Few if any companies have yet dem- ultimately lives up to the hype, inves- been 17 white Presidents.j
onstrated major cost savings from the tors will surely be disappointed. His- Like Elon Musk, I worry about AI, but
deployment of generative AI. In tests, tory suggests that investors in the not because I expect to be murdered by
language models like ChatGPT have hottest areas always are. Railroads a superintelligent machine. Modern AI
so far proven themselves as productiv- transformed the U.S. in the second half is inseparable from the cloud platforms
ity aids mostly by integrating predictive of the 1800s, but overexcited European and massive data collections of big tech
text into programming environments, investors drove massive overbuilding. companies, meaning that it inherits
as computer code is far more structured J.P. Morgan built his fortune by consol- and magnifies all the concerns people
than natural language. idating the struggling industry on the have developed about them in the past
The immediate applications opened cheap during the 1880s. Investors of decade. It is driven by the same fads,
up by driving down the cost of bespoke the 1990s were not wrong about the In- groupthink, and obsessive quest for the
BS generation are real but not par- ternet being a big deal but they still bid next big thing that led Facebook, Mi-
ticularly hopeful centering mostly of up stocks beyond all reason. In March crosoft and Apple all to bet on virtual
student plagiarism, personalized pro- 2000, Cisco briefly became the world’s reality as a massive emerging market.
paganda, misinformation, clickbait, most valuable company, as investors Generative AI systems are consuming
search engine spam, scams, and fake bet on its domination of the market for energy in vast quantities at a time when
news. For example, websites are now networking equipment. Today, Cisco the effects of climate change are becom-
generating fake obituaries of accident no longer appears in the top 50 despite ing ever more apparent. The world of AI
victims to earn tiny amounts of advertis- higher sales and bigger profits. startups and subsidiaries is a monocul-
ing revenue from visitors. ture dominated by a handful of spectac-
Will generative technologies ulti- Beyond Chatbots ularly wealthy and deeply strange men
mately displace workers? Perhaps, but Facebook AI chief and neural net pio- making decisions with huge impacts for
there’s nothing unusual about technol- neer Yan LeCun finds the AGI concept the rest of us. I worry about the intersec-
ogy eliminating jobs. The work most meaningless and insists that large lan- tion of power, ego, weird personal obses-
people did in 1800 or 1900 has long guage models are “not a path toward sions, and political clout. Perhaps the
since been automated. Computers have human-level intelligence.”i He mocks thing I am worrying about is not AI after
been replacing white-collar workers the idea that statistical text predic- all, but Elon Musk.
for decades. What’s different this time tion is the key to true artificial intel-
around is that the machines seem to be ligence, though others promise that it References
1. A.R. Anderson, (Ed.). Minds and Machines. Prentice-
coming for the jobs of people who earn can be coupled with other technologies Hall, Englewood Cliffs, NJ (1964).
their livings writing columns, express- to eliminate the tendency to, as they 2. Chayka, K. The uncanny failures of AI-generated
hands. New Yorker (Mar. 10, 2023); [Link]
ing opinions, or appearing on television. euphemistically put it, “hallucinate.” ly/3BKNvUR
They are understandably more shocked Symbolic AI expert Gary Marcus, com- 3. Davis, M. Engines of Logic: Mathematicians and the
Origin of the Computer. Norton, NY (2001).
by the prospect of their own jobs van- putational linguist Emily Bender, and 4. E.A. Feigenbaum and J. Feldman, (Eds.) Computers and
Thought. McGraw-Hill, NY (1963).
ishing than those of boot makers, bank computer scientist Grady Booch have 5. Frankfurt, H. On Bullshit. Princeton University Press,
tellers, or file clerks. also been consistent critics of AGI hype. Princeton, NJ (2005).
6. Kurzweil, R. The Singularity is Near. Viking, NY (2005).
The investment boom is driven not It seems unlikely to me that a tech- 7. Metz, C. Genius Makers: The Mavericks Who Brought AI
by proven savings or actual produc- nology that excels at faking intel- to Google, Facebook, and the World. Dutton, NY (2021).
8. Rich, E. Artificial Intelligence. McGraw-Hill, NY (1983).
tivity growth but by faith that we can ligence will turn out to be the best 9. Turing, A. Computing machinery and intelligence. Mind
achieve AGI by building bigger and bet- platform on which to build true intel- LIX 236, (Oct. 1950).
10. Vaswani, A. et al. Attention is all you need. Advances in
ter language models, feeding them ever ligence, but chatbots are rapidly be- Neural Information Processing Systems 30, (2017).
larger quantities of training data and coming user interfaces for underlying 11. Vinge, V. Technological singularity. Whole Earth Rev.
(Winter 1993).
running then in ever more powerful systems that use completely different 12. Winston, P. Artificial Intelligence. Addison-Wesley,
server farms. Investor enthusiasm for methods. The latest ChatGPT is sup- Reading, MA (1977).
13. Wooldridge, M. A Brief History of Artificial Intelligence:
generative AI continues to grow even as posed to speak in a human-like voice What It Is, Where We Are, and Where We Are Going.
slightly earlier waves of AI-branded tech- with simulated emotion, solve math- Flatiron Books, NY (2021).

nology collapse. Voice-powered assis- ematical problems by integrating a

tants were supposed to transform our separate problem-solving engine, and j See [Link]
lives and make fortunes for companies recognize objects in photographs.
that harvested our data and used it to The rush to build generative AI into
Thomas Haigh ([Link]@[Link]) is a professor
sell us things, a model dubbed surveil- search engines and office applications of history at the University of Wisconsin—Milwaukee,
lance capitalism by its critics. Amazon scares me. When Google started put- WI, USA, and a Comenius visiting professor at Siegen
University, Germany.
lost tens of billions of dollars on its Al-
exa devices, before firing thousands of i See [Link] © 2025 Copyright held by the owner/author(s).

F E B R UA RY 2 0 2 5 | VO L. 6 8 | N O. 2 | C OM M U N IC AT ION S OF T HE ACM 29