Sources and References

• Turing, Alan. “Computing Machinery and Intelligence.” Mind 59, no. 236 (1950): 433–460.
• McCarthy, John, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. 1955.
• Dreyfus, Hubert L. What Computers Can’t Do: A Critique of Artificial Reason. Harper & Row, 1972.
• Minsky, Marvin. The Society of Mind. Simon & Schuster, 1986.
• McCorduck, Pamela. Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence. W. H. Freeman, 1979.
• Nilsson, Nils J. The Quest for Artificial Intelligence: A History of Ideas and Achievements. Cambridge University Press, 2010.
• LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep Learning.” Nature 521 (2015): 436–444.
• Kang, Minsoo. Sublime Dreams of Living Machines: The Automaton in the European Imagination. Harvard University Press, 2011.
• Suchman, Lucy. Human-Machine Reconfigurations: Plans and Situated Actions, 2nd ed. Cambridge University Press, 2007.
• Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.
• Couldry, Nick, and Ulises A. Mejias. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press, 2019.
• Cave, Stephen, and Kanta Dihal, eds. Imagining AI: How the World Sees Intelligent Machines. Oxford University Press, 2023.
• Cardon, Dominique, Jean-Philippe Cointet, and Antoine Mazières. “Neurons Spike Back: The Invention of Inductive Machines and the Artificial Intelligence Controversy.” Réseaux 5–6, no. 211–212 (2018): 173–220.
• Gray, Mary L., and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019.
• Lanier, Jaron. You Are Not a Gadget: A Manifesto. Alfred A. Knopf, 2010.
• Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT 2021.
• Online resources: Stanford HIST 244S public materials; MIT OpenCourseWare STS.050; Harvard HISTSCI 119 syllabus; Cambridge HPS Part II Histories of Computing; AI Now Institute annual reports (ainowinstitute.org).

Chapter 1: Why AI Has a History

Artificial intelligence presents itself, with remarkable consistency, as a technology without a past. Press releases from major AI laboratories speak of systems that “have learned” to do something, as if the process were purely natural, unfolding outside the normal circuits of institutional pressure, funding politics, and cultural expectation. Academic papers cite benchmarks rather than genealogies. Startup founders announce that their products will change everything — always forward-looking, never backward-glancing. The futurism problem (未来主义问题) is perhaps the defining rhetorical characteristic of AI as a field: it presents itself as prophecy while suppressing its own history. The task of this course is to correct that suppression, to bring the history of artificial intelligence into view as a proper subject of historical inquiry, complete with contingency, failure, ideology, and the full messiness of human enterprise.

What does it mean, concretely, to write a history of a technology that claims to transcend its own historical moment? It means, first, attending to ideas — the intellectual lineages through which concepts of mind, logic, symbol, and learning were forged, debated, and institutionalised. It means, second, attending to materials and infrastructures — the laboratories, the funding agencies, the server farms, the lithium deposits, and the underpaid human annotators whose invisible labour makes AI systems possible. And it means, third, attending to imaginative culture — the novels, films, and popular narratives through which societies have dreamed, feared, and shaped their expectations of intelligent machines, often long before such machines existed. These three historiographical lenses — intellectual history (思想史), material and infrastructure history (物质与基础设施史), and cultural and imaginative history (文化与想象史) — organise the arc of this course.

The intellectual history of AI reaches back well before the word “artificial intelligence” was coined. It runs through the logicist tradition in philosophy and mathematics, through wartime cybernetics, through postwar cognitive science, and into the deep learning era. McCorduck, in Machines Who Think, traces this lineage with unusual candour about the personalities and social dynamics that drove it, arguing that AI was never merely a technical project but always also a project about the nature of mind and the place of human beings in a universe of symbol-processing matter. Nilsson’s Quest for Artificial Intelligence provides a denser, more technically oriented genealogy, but both authors share an insistence that ideas have histories — that the concepts researchers used were borrowed, contested, and revised in ways that left lasting marks on the systems they built.

The material history of AI is newer as a scholarly project, and considerably more unsettling. Crawford’s Atlas of AI makes the case that every act of machine learning draws on a planetary infrastructure of extraction: minerals mined under brutal conditions, data centres consuming colossal quantities of water and electricity, and a global workforce of human raters and annotators whose labour is deliberately obscured in the presentation of AI as autonomous. This material history is not incidental to AI’s intellectual development — it shapes which systems get built, which problems get prioritised, and which communities bear the costs of progress. Couldry and Mejias extend this analysis into the conceptual register, arguing that the mass collection of human behavioural data constitutes a form of colonialism structurally homologous to earlier resource extraction.

The cultural and imaginative history of AI is the dimension most familiar from popular discourse, but it is also the most frequently misread. Fiction about intelligent machines is not merely entertainment that reflects technical reality; it actively shapes research agendas, policy debates, and public expectations. Cave and Dihal’s Imagining AI demonstrates that AI imaginations vary dramatically by national and cultural context — that the robots of Japanese manga carry entirely different cultural freight than the rogue AIs of Hollywood science fiction — and that these imaginative frameworks condition what researchers think they are building, what policymakers think they are regulating, and what publics think they are welcoming or resisting. This course treats the relationship between AI fiction and AI fact as one of mutual constitution, not mere reflection.

A note on method: the course sits in conscious relationship to two companion offerings in the UWaterloo AI track. PHIL 459b (Philosophy of AI) takes up the conceptual questions — what would it mean for a machine to think? — that this course deliberately sets aside in favour of the historical question of how those conceptual questions were framed and institutionalised. PHIL 451 (AI Ethics, Law, and Governance) addresses the normative question of what ought to be done about AI’s present trajectory. History, this course argues, is prior to both: it tells us how we arrived at the questions that philosophy and ethics then attempt to answer.

Chapter 2: Precursors and Prehistory — Automata, Engines, and Dreaming Machines

Long before anyone spoke of artificial intelligence, European courts entertained themselves with mechanical animals and simulated human beings. The automata of the eighteenth century — Vaucanson’s celebrated duck, which appeared to eat, digest, and excrete; the Jaquet-Droz draughtsman and musician, who could write sentences and play real melodies on a real organ — were not proto-computers in any meaningful engineering sense. They were, as Kang argues in Sublime Dreams of Living Machines, cultural objects that expressed and contested ideas about the relationship between mechanism and life, matter and spirit, the artisan’s craft and the divine act of creation. The duck did not think; it did not even truly digest, as later examination revealed the digestive mechanism to be separate from the food-intake mechanism. But the appearance of digestion was enough to provoke serious philosophical dispute about whether the boundary between living and non-living matter was principled or merely conventional.

This prehistory matters for the history of AI not because automata were technically ancestral to computers, but because they established a cultural vocabulary — and a set of recurring anxieties — that the later field would inherit. The Romantic period’s anxiety about the created being turning against its creator, crystallised in Mary Shelley’s Frankenstein (1818), drew directly on the automaton tradition. The idea that a sufficiently lifelike mechanism might harbour genuine interiority, or might exceed the intentions of its maker, or might constitute a threat to human uniqueness — all of these themes were thoroughly worked over by the automaton tradition before a single transistor was switched. As Kang shows, this is not mere intellectual background colour; it shapes the metaphors that AI researchers and their critics reach for instinctively, even when they believe they are speaking purely technically.

The nineteenth century produced two developments of more direct technical relevance. Charles Babbage’s design for the Analytical Engine, developed through the 1830s and 1840s, was the first conception of a general-purpose, programmable calculating machine — a machine that could, in principle, perform any sequence of arithmetic and logical operations specified in advance on punched cards. Ada Lovelace’s notes on the Analytical Engine, published in 1843, went further than Babbage’s own prose in articulating what such generality might mean: she observed that the Engine could manipulate any symbols whatsoever, not merely numbers, and that its operations were in a genuine sense those of a formal rule-following process. She also, crucially, insisted that the Engine could originate nothing — it could only do what it was programmed to do — a limitation that Turing would famously revisit a century later under the rubric of “Lady Lovelace’s Objection.” Whether or not Lovelace anticipated computation in the modern sense, her notes mark a moment when the idea of a general symbol-manipulating machine was given articulate, public expression.

The logicist tradition in mathematics and philosophy provided the conceptual raw material that would eventually be assembled into the intellectual core of symbolic AI. George Boole’s Laws of Thought (1854) demonstrated that the operations of logical reasoning could be expressed as algebraic manipulations over a two-valued system — a result that connected formal logic to the kind of mechanical operation that machines could, in principle, perform. Gottlob Frege’s Begriffsschrift (1879) pushed further, developing a formal notation capable of expressing the full structure of mathematical proof, free from the ambiguities of natural language. The logicist programme of Frege and Russell — the attempt to derive all of mathematics from logical axioms — was eventually undermined by Gödel’s incompleteness theorems (1931) and Russell’s own paradox, but it left behind a powerful idea: that rigorous reasoning is, at bottom, a formal process that might be mechanised.

Alan Turing’s 1936 paper “On Computable Numbers, with an Application to the Entscheidungsproblem” synthesised these strands into something new and transformative. By imagining a simple hypothetical machine — a reading head moving along an infinite tape, changing its internal state and writing symbols according to a finite set of rules — Turing demonstrated that any well-defined computation could be performed by such a device, and that a single “universal” such machine could simulate any other. The paper’s primary aim was mathematical: to prove that certain problems were not mechanically decidable. But its conceptual by-product — the notion of a universal symbol-manipulating machine that could, in principle, do anything that any other such machine could do — became the theoretical foundation for the stored-program computer and, eventually, for the idea that intelligence itself might be a form of computation. The temptation to read all of this prehistory as a straight line leading inevitably to modern AI must, however, be resisted. Teleological readings — what historians call Whig history (辉格史学) — flatten the contingency of the actual path and make alternative possibilities invisible. Babbage’s engine was never built in his lifetime; Lovelace’s notes were largely forgotten until Turing cited them; the logicist programme failed on its own terms. The connections are real but retrospective, constructed by later actors who selectively assembled a usable past.

Chapter 3: The Turing Moment — Computation, Imitation, and the Question of Mind

In October 1950, the journal Mind published a paper by Alan Turing that opened with a sentence of deliberate provocation: “I propose to consider the question, ‘Can machines think?’” What followed was not an answer to that question but a replacement of it — an argument that the question, as ordinarily posed, was too philosophically loaded to be answerable, and that a more tractable substitute could be found in what Turing called the Imitation Game (模仿游戏). The game’s setup is well-known: a human interrogator communicates via typewritten messages with two respondents, one human and one machine, and tries to determine which is which. Turing proposed that if a machine could play this game well enough to fool a significant proportion of interrogators, the question of whether it “really” thought would be of no practical importance. The rhetorical structure of this move — reframing an apparently metaphysical question as an empirical and operational one — was as significant as any specific claim Turing made about machine capabilities.

What Turing was and was not claiming in 1950 has been debated intensively by philosophers, computer scientists, and historians ever since. He was not claiming that a machine passing the imitation game would thereby possess genuine consciousness or inner experience — the paper was agnostic about such matters. He was claiming, more modestly, that the question of machine intelligence was tractable, empirically approachable, and likely to be answered affirmatively within about fifty years (a prediction that turned out to be considerably optimistic in some dimensions and surprisingly apt in others). He was also deploying operationalism (操作主义) — the philosophical stance that a concept’s meaning is exhausted by the procedures for measuring it — as a rhetorical weapon against those who wished to dismiss machine intelligence on purely conceptual grounds. The paper anticipated nine objections to machine intelligence and responded to each; it was, in effect, a comprehensive defence of the research programme that would come to be called artificial intelligence, written six years before that name was coined.

The materiality of early computation is easy to lose sight of behind the clean theoretical abstractions of Turing’s paper. The Manchester Small-Scale Experimental Machine (the “Baby”), which ran its first programme in June 1948, was a room-filling assembly of cathode ray tubes and vacuum tubes that required constant maintenance and consumed prodigious quantities of electricity; it stored a grand total of 128 bits of data. ENIAC, completed at the University of Pennsylvania in 1945, contained 18,000 vacuum tubes and required a team of operators — the famous “computers,” mostly women with mathematics degrees — to set up each new calculation by physically reconfiguring patch cables. These machines were not implementations of Turing’s universal machine in any direct sense; they were engineering achievements with their own design trajectories, built for specific purposes (ballistics calculations, code-breaking) before being adapted to more general research. The gap between the theoretical universality of the Turing machine and the practical constraints of actual hardware is itself a historical phenomenon with lasting consequences.

The intellectual context of the late 1940s was shaped decisively by wartime experience. Norbert Wiener’s cybernetics (控制论), developed during the war in the context of anti-aircraft fire control and formalised in his 1948 book of the same name, proposed that feedback-governed purposive behaviour was the key property shared by living organisms and certain machines — a proposal that dissolved the boundary between biology and engineering in ways that had profound effects on how researchers thought about intelligence. Claude Shannon’s information theory, also formalised in 1948, provided a mathematical framework for quantifying the transmission of arbitrary symbols over noisy channels — a framework that would be applied (sometimes too casually) to questions about thought, language, and mind. RAND Corporation, the DARPA-funded think tank established in 1948, became a crucible for interdisciplinary thinking about rational decision-making, game theory, and machine cognition, drawing together mathematicians, psychologists, economists, and engineers under the broad umbrella of what we might now call cognitive science.

Chapter 4: The Founding Moment — Dartmouth 1956 and the AI Programme

In the summer of 1955, four researchers — John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon — submitted a proposal to the Rockefeller Foundation requesting funds for a two-month summer research project to be held at Dartmouth College. The proposal’s opening sentence announced its subject with crisp confidence: “We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956.” The phrase “artificial intelligence” had been used occasionally before, but this document gave it institutional form, naming a new research discipline and implicitly claiming for it a set of problems, methods, and ambitions. The Dartmouth Conference (达特茅斯会议) of 1956 is conventionally dated as the founding moment of AI as a distinct academic field — a status it retains despite the conference’s actual proceedings being rather less dramatic than its retrospective reputation suggests.

The name “artificial intelligence” was itself a deliberate and consequential rhetorical choice. McCarthy later recalled choosing it in part to distinguish the new project from Wiener’s cybernetics, which he found too focused on analogue feedback systems and too committed to biological analogies. The word “artificial” carried the connotation of deliberate construction rather than natural growth; “intelligence” claimed, at once, that intelligence was the subject of study and that machines could in principle possess it. The proposal’s intellectual bet was stated with remarkable clarity: “every aspect of learning or every other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” This is the hypothesis of symbolic AI (符号人工智能) in its purest form — that cognition is computation over discrete symbolic representations, and that a machine capable of the right symbolic computations is thereby capable of intelligence.

The first demonstrations that this hypothesis might be more than wishful thinking came quickly. Allen Newell and Herbert Simon, working at RAND and Carnegie Mellon, presented the Logic Theorist at the Dartmouth Conference — a programme that could prove theorems in propositional logic by heuristic search through a space of possible proof steps, mimicking (they argued) the problem-solving strategies of human mathematicians. Within two years they had developed the General Problem Solver (通用问题求解器), a programme designed not to solve any particular class of problems but to embody a general strategy — means-ends analysis — applicable in principle to any well-defined problem. Newell and Simon were explicit about the theoretical significance of this achievement: they believed they had identified the fundamental mechanism of human intelligent behaviour, and they said so with a confidence that later events would render premature. Simon famously predicted in 1957 that within ten years a computer would be chess champion of the world and would discover and prove an important mathematical theorem — predictions that turned out to be correct, but on a timescale of forty years, not ten.

The institutional conditions that made such optimism structurally sustainable deserve attention. DARPA’s funding model in the late 1950s and 1960s was extraordinarily permissive by later standards: programme managers could and did write large cheques for research groups with interesting ideas, with minimal bureaucratic oversight and long time horizons. MIT’s AI Lab (founded by Minsky and McCarthy in 1959), Stanford’s AI Laboratory (founded by McCarthy in 1963), and Carnegie Mellon’s AI programme (under Newell and Simon) became magnets for talented graduate students and beneficiaries of a resource flow that was ultimately tied to Cold War anxieties about Soviet technological leadership. Within this funding environment, confident claims about imminent breakthroughs were not merely expressions of genuine scientific optimism — they were also, structurally, the appropriate genre for grant-seeking. The hype cycle (炒作周期) was built into AI’s institutional architecture from the beginning.

The Dartmouth conference also marked AI’s differentiation from cybernetics as a research tradition. Wiener’s cybernetics had proposed a unified science of control and communication in animal and machine, and had attracted extraordinary interdisciplinary interest through the Macy Conferences of the late 1940s and early 1950s. The symbolic AI programme that emerged from Dartmouth was narrower in scope — focused on discrete, rule-governed symbol manipulation rather than analogue feedback — and more confident in its claims about the nature of intelligence. This narrowing would prove both productive and limiting: productive because it generated a clear research programme with tractable sub-problems, limiting because it ruled out from the start the embodied, distributed, and statistical approaches to intelligence that would eventually prove more powerful.

Chapter 5: First Promises and First Failures — The 1960s–70s and the First AI Winter

The decade following Dartmouth saw the establishment of the major tools and languages of symbolic AI. John McCarthy’s LISP, developed at MIT in 1958, became the standard programming language of AI research for three decades. LISP treated programs and data as instances of the same object — lists of symbols — and made it natural to write programs that manipulated other programs, an important property for AI systems that needed to reason about their own knowledge. LISP’s homoiconicity (the property that code and data share the same representation) was not merely a technical convenience; it expressed a philosophical commitment to the view that intelligence was the manipulation of symbolic structures, and that there was no principled distinction between programs (rules for manipulation) and data (the objects being manipulated). The AI lab culture that grew up around LISP — hacker culture in its original sense, a culture of late-night programming sessions and intrinsic motivation — was itself a social formation with lasting effects on how the field understood its own work.

The first serious intellectual challenge to the symbolic AI programme came not from within the field but from philosophy. Hubert Dreyfus, a philosopher of phenomenology at MIT and later Berkeley, published a RAND Corporation working paper in 1965 — later expanded into the book What Computers Can’t Do (1972) — that argued symbolic AI faced structural, not merely technical, obstacles. Drawing on Heidegger’s analysis of being-in-the-world and Merleau-Ponty’s account of embodied cognition, Dreyfus contended that human intelligence was not, at its core, a matter of applying explicit rules to discrete symbolic representations. Rather, it depended on a background of tacit knowledge (默会知识) — practical, embodied, contextual understanding that could not be made fully explicit without generating an infinite regress. The frame problem (框架问题) — how a reasoning system can determine which facts in its knowledge base are relevant to a given situation, without checking all of them — was Dreyfus’s sharpest technical illustration of this philosophical difficulty. A symbolic system that represents the world as a database of explicit propositions faces, in principle, an intractable problem of relevance that human beings solve effortlessly through embodied situatedness.

The AI community’s response to Dreyfus was not, on the whole, gracious. Minsky famously arranged for Dreyfus to play chess against an AI programme — which beat him — as if this demonstrated the erroneousness of philosophical critique. The episode is revealing: it showed the symbolic AI community’s tendency to interpret any demonstration of machine competence in a bounded, well-defined domain as evidence against the general critique, when Dreyfus’s argument was precisely that such domains were untypical of the full scope of human intelligence. Over the subsequent decades, many of Dreyfus’s specific predictions about the limits of symbolic systems — in natural language understanding, in visual scene interpretation, in robotics — proved accurate, even if his philosophical framework remained contested.

The formal reckoning came in 1969 and 1973. Minsky and Papert’s Perceptrons (1969) proved mathematical theorems about the limitations of the single-layer neural networks that had been proposed in the late 1950s as an alternative to symbolic AI. The book showed that perceptrons — simple threshold-logic units — could not compute certain important functions, notably the XOR function. The argument was technically correct but rhetorically overreaching: the theorems applied to single-layer networks, not to the multi-layer networks that would eventually prove powerful. Nevertheless, Perceptrons effectively froze neural network research for roughly fifteen years, partly because Minsky and Papert were the most prestigious figures in AI and their negative assessment carried institutional weight, and partly because the funding agencies that sustained AI research could not easily distinguish valid mathematical critique from general dismissal. In the United Kingdom, the mathematician Sir James Lighthill produced a report for the Science Research Council in 1973 concluding that AI had failed to deliver on its promises across all its major areas — robotics, language understanding, theorem proving — and recommending substantial cuts to funding. The cuts followed.

The concept of an AI winter (人工智能寒冬) — a period of reduced funding and diminished public enthusiasm following a cycle of unfulfilled promises — is central to the field’s self-understanding, though the term was not coined until the late 1980s. The winter of the 1970s was not a simple failure of intelligence; it was a sociotechnical phenomenon produced by the interaction of unrealistic capability claims, a funding model that rewarded optimism, and the inevitable collision between those claims and actual performance. The Lighthill Report’s critique was partly motivated by inter-disciplinary politics — the report was relatively sympathetic to robotics and relatively hostile to the grander theoretical claims of symbolic AI — but its underlying point, that AI had consistently overpromised and underdelivered, was factually well-grounded. The lesson that AI winters are structural, not accidental — that they are produced by the same funding dynamics that generate the preceding booms — is one that the field has repeatedly failed to internalise.

Chapter 6: The Knowledge Revolution — Expert Systems and the 1980s Boom

The second wave of AI enthusiasm, which crested in the early 1980s, was built on a different intellectual foundation from the grand theoretical ambitions of the Dartmouth era. Rather than seeking programs that embodied general intelligence, researchers in the 1970s turned to expert systems (专家系统): computer programs that encoded the specialised knowledge of human domain experts in the form of rules — typically thousands of “if-then” conditional statements — and applied those rules via a logical inference engine to answer questions or make recommendations within a well-defined domain. DENDRAL, developed at Stanford in the late 1960s, could infer the molecular structure of organic compounds from mass spectrometry data. MYCIN, also from Stanford, could diagnose bacterial infections and recommend antibiotic treatments, in clinical trials performing comparably to specialist physicians. XCON (later R1), deployed by Digital Equipment Corporation in the early 1980s to configure customer orders for VAX minicomputers, reportedly saved the company $40 million per year.

The commercial success of these systems drove a massive expansion of the AI industry through the early 1980s. Japan’s Ministry of International Trade and Industry announced the Fifth Generation Computer Systems project in 1981, a ten-year programme to develop massively parallel machines capable of running logic programming languages at speeds sufficient for real-time AI applications. The announcement provoked alarm in the United States and Europe and triggered a flood of new government and corporate AI investment. LISP machine companies — Symbolics, Lisp Machines Inc., Xerox — sold specialised hardware optimised for AI workloads at premium prices. AI consultancies proliferated. The business press, with characteristic optimism, declared that expert systems would shortly replace large swathes of professional labour: physicians, lawyers, financial analysts. The institutional structures of AI academia expanded in parallel: new departments, new journals, new conferences.

The European AI scene developed a distinctive identity around logic programming (逻辑编程) and the language Prolog, developed by Alain Colmerauer and colleagues at the University of Marseille in the early 1970s. Prolog allowed programmes to be written as collections of logical facts and rules, with computation proceeding by automatic proof search — a style that felt closer to formal reasoning than to the procedural programmes typical of American AI. The Fifth Generation project adopted Prolog as its primary language, contributing to Prolog’s brief moment of transatlantic prominence. The European AI tradition generally placed greater emphasis on formal foundations — type theory, categorical logic, constructive mathematics — and was somewhat more sceptical of the engineering-over-theory pragmatism that characterised the American AI mainstream. This divergence in style had lasting effects on how questions about machine reasoning, formal verification, and the relation between logic and learning were framed.

The second AI winter arrived in the late 1980s with characteristic abruptness. Expert systems, it turned out, were extraordinarily brittle: they worked well within the exact domain for which they had been engineered, but failed in unpredictable ways when queries fell outside that domain, because they lacked any mechanism for recognising the boundaries of their own competence. Building and maintaining a large expert system required years of painstaking “knowledge engineering” — the extraction and formalisation of expert knowledge — a process that was both expensive and often resented by the experts whose knowledge was being extracted. The knowledge engineers’ fundamental problem — how to elicit and represent tacit, contextual, practical expertise in a form amenable to explicit rule encoding — was, in retrospect, precisely the problem that Dreyfus had identified in 1965. The LISP machine market collapsed as general-purpose workstations from Sun and Apollo caught up in performance while undercutting dramatically on price. Japan’s Fifth Generation project quietly underdelivered on its targets. The AI bubble deflated.

It is at this historical juncture that the analysis of Cardon, Cointet, and Mazières becomes indispensable. Their “Neurons Spike Back” article reconstructs, with careful attention to archival evidence, how the connectionist approach — the tradition of modelling intelligence through networks of simple, neuron-like units — staged a gradual comeback during the very period of symbolic AI’s apparent dominance. The connctionist tradition had never entirely disappeared; researchers like James Anderson, Teuvo Kohonen, and Stephen Grossberg continued to publish throughout the 1970s, though largely outside the mainstream AI conferences. What Cardon, Cointet, and Mazières show is that the return of neural networks was not a sudden revolutionary discovery but a slow accumulation of results — new training algorithms, new architectures, new mathematical analyses — that gathered momentum beneath the surface of the symbolic AI consensus. The narrative of a sudden paradigm shift, they argue, is itself a retrospective construction that obscures the longer continuities and the extent to which the “victory” of connectionism depended on institutional conditions — hardware, data, funding — as much as on purely intellectual breakthroughs.

Chapter 7: The Connectionist Comeback — Neural Networks, Big Data, and Deep Learning

The intellectual prehistory of modern deep learning begins not in 2012 but in 1986, with the publication of David Rumelhart, Geoffrey Hinton, and Ronald Williams’s paper “Learning Representations by Back-propagating Errors” in Nature. Backpropagation (反向传播), the algorithm described in that paper, provided an efficient method for training multi-layer neural networks — for adjusting the weights connecting artificial neurons in multiple layers so as to reduce the network’s error on a training set. The idea was not entirely new: versions of backpropagation had been described earlier by Seppo Linnainmaa, Paul Werbos, and others. But the Rumelhart–Hinton–Williams formulation, combined with the parallel distributed processing (PDP) framework developed in the accompanying two-volume collection, gave the approach a compelling theoretical framing and attracted serious attention from a research community that had largely dismissed neural networks since Minsky and Papert.

The 1990s and early 2000s were, for neural networks, a period of accumulation under adverse conditions. The dominant applied machine learning methods were support vector machines (支持向量机), developed by Vladimir Vapnik and Corinna Cortes at AT&T Bell Labs, which had strong theoretical foundations in statistical learning theory and consistently outperformed neural networks on standard benchmarks with smaller datasets. Bayesian probabilistic methods were ascendant in fields like speech recognition and natural language processing. Neural network research continued — Yann LeCun’s work on convolutional networks for digit recognition, Sepp Hochreiter and Jürgen Schmidhuber’s long short-term memory networks — but without the institutional resources or the public profile that would come later. Hinton, LeCun, and Yoshua Bengio continued to push the neural approach through a long period of relative obscurity, sustained by the Canadian Institute for Advanced Research’s CIFAR programme, which created a protected space for unfashionable but potentially transformative research.

The transformation came from outside the algorithm: it came from data and hardware. The ImageNet dataset, assembled by Fei-Fei Li and colleagues at Princeton and later Stanford beginning in 2007, contained more than fourteen million hand-labelled images in over twenty thousand categories — a scale of labelled training data previously unavailable for visual recognition research. The ImageNet Large Scale Visual Recognition Challenge, launched in 2010, became the benchmark competition through which deep learning demonstrated its superiority over all prior approaches. In 2012, a deep convolutional network called AlexNet, designed by Alex Krizhevsky, Ilya Sutskever, and Hinton at the University of Toronto, won the challenge with an error rate of 15.3%, more than ten percentage points better than the second-place entry. The gap was large enough to be unambiguous: something qualitatively new was happening.

LeCun, Bengio, and Hinton have been retrospectively designated the “Three Godfathers of Deep Learning,” a framing that the 2018 ACM Turing Award reinforced. The narrative serves important functions — it provides the field with founding heroes and a clear story of visionary persistence rewarded — but Cardon, Cointet, and Mazières’s historiographical caution applies here too. The AlexNet breakthrough was as much a materials and infrastructure story as an algorithmic one. The Graphics Processing Unit (GPU), originally developed for video game rendering, turned out to be extraordinarily well-suited to the parallel matrix computations required by neural network training; NVIDIA’s CUDA programming platform, released in 2007, made it possible for researchers to exploit GPU acceleration without specialised hardware engineering skills. Amazon Web Services, launched in 2006, made large-scale cloud computing accessible without capital investment in physical infrastructure. And ImageNet itself represented the labour of many thousands of human annotators — a fact to which we shall return in the next chapter — whose work was organised through Amazon Mechanical Turk.

The decade following AlexNet saw an acceleration of capability that consistently outpaced both pessimistic predictions and optimistic ones. The transformer architecture (变换器架构), introduced by Vaswani and colleagues at Google in 2017, replaced recurrent connections with self-attention mechanisms that could be parallelised across the entire input sequence, enabling the training of far larger models than had previously been feasible. The GPT series (Radford et al., 2018, 2019; Brown et al., 2020), BERT (Devlin et al., 2018), and the subsequent proliferation of foundation models (基础模型) — large-scale models pre-trained on internet-scale text and image data and then fine-tuned for specific tasks — constituted a genuine paradigm shift, in the sense that a single trained model could, with appropriate prompting or fine-tuning, perform competently across a wide range of tasks without domain-specific re-engineering. As LeCun, Bengio, and Hinton noted in their 2015 Nature review of deep learning, the combination of large datasets, increased compute, and improved training algorithms had produced systems whose capabilities would have seemed implausible a decade earlier — while also producing systems whose failure modes, biases, and internal representations remained poorly understood.

Chapter 8: The Infrastructure of Artificial Intelligence — Labour, Extraction, and Data Colonialism

Kate Crawford’s Atlas of AI, published in 2021, opens with a visit to the Thacker Pass lithium mine in Nevada, and the choice is deliberate and precise. Before a neural network can be trained, before a server can be powered on, before a data centre can cool its racks, the mineral substrate of computation must be extracted from the earth — lithium for batteries, cobalt for cathodes, rare earth elements for magnets and displays, coltan for capacitors. This extraction is not incidental to AI; it is constitutive of it. Crawford’s argument is that the discourse of AI as pure intelligence — disembodied, weightless, running on “the cloud” — is a systematic misrepresentation that serves to conceal the planetary-scale material costs of the technology. Every act of machine learning draws on a cascade of prior extractions: from the earth, from human labour, from human behavioural data. The Atlas (地图集) metaphor is deliberate: Crawford is providing a geographic and material map of a technology that presents itself as having no geography and no materiality.

The energy and water consumption of AI infrastructure merits detailed attention. Training a single large language model — GPT-3, for instance — has been estimated to require the energy equivalent of hundreds of transatlantic flights; running inference across millions of queries daily requires sustained power at scales that are already reshaping energy markets in regions where large data centres cluster. Water cooling, required to maintain the thermal performance of dense server racks, places AI infrastructure in direct competition with agricultural and domestic water use in areas that are simultaneously experiencing climate-driven water stress. Microsoft’s announcement of expanded data centre capacity in the American Southwest — a region facing severe water shortage — crystallises the tension between the discourse of AI as a solution to environmental problems and its actual material footprint. Crawford documents these figures not to argue against AI as such, but to insist that the question “what does AI cost?” must be answered materially, not merely economically.

Couldry and Mejias develop a parallel and complementary argument in The Costs of Connection, introducing the concept of data colonialism (数据殖民主义) to characterise the extraction of human behavioural data as a new form of appropriation structurally homologous to the historical colonial extraction of natural resources. The analogy is not merely rhetorical. Historical colonialism, on their analysis, involved the appropriation of territories, peoples, and resources from outside the market economy and their incorporation into capitalist circuits of accumulation — a process that was presented ideologically as civilisational progress and the spread of rational organisation. Data colonialism, they argue, involves the appropriation of human social life — attention, behaviour, relationship, identity — from outside the domain of economic exchange and its incorporation into the data accumulation strategies of platform corporations. The ideological presentation is structurally identical: this too is framed as progress, connection, empowerment. What is suppressed in both cases is the one-sidedness of the appropriation and the structural inequality it produces.

Mary Gray and Siddharth Suri’s Ghost Work directs attention to a more immediately visible form of exploitation: the micro-task labour (微任务劳动) that sustains AI systems at every stage of their development and operation. Training a supervised machine learning system requires labelled data — images with correct category labels, text with correct sentiment annotations, audio with correct transcriptions. The labelling is done by human beings, typically hired through online platforms like Amazon Mechanical Turk, Appen, or Scale AI, paid at rates of pennies per task with no employment protections, no benefits, and no job security. Content moderation — the removal of violent, abusive, or illegal material from social media platforms — is performed by human workers, often in the Philippines, Kenya, or other countries with lower labour costs, who are exposed to psychologically damaging content with minimal support. Gray and Suri document the structural dynamics that keep this workforce invisible: the platforms’ deliberate use of “artificial” branding for systems that are heavily dependent on human labour, the piece-rate payment structures that prevent workers from organising, and the geographic dispersion that makes collective action difficult.

The geography of AI labour is not accidental but reflects structural inequalities in the global economy. The concentration of annotation work in Venezuela, India, the Philippines, and Kenya reflects wage differentials that make the economics of micro-tasking viable — workers in these locations can earn above local minimum wages for tasks that would be economically infeasible to perform at developed-country wages. This geographic distribution embeds structural inequality into AI systems at the most fundamental level: the training data that shapes a system’s capabilities and biases is produced by workers whose social position, cultural context, and economic interests are systematically different from those of the end users for whom the system is ostensibly designed. Debates about digital sovereignty (数字主权) — the EU AI Act’s attempt to regulate AI systems deployed in European territory, India’s data localisation requirements, China’s cybersovereignty doctrine — can be understood in part as responses to this structural asymmetry, attempts by states to assert control over infrastructure and data that currently flows predominantly through American and Chinese corporate channels.

Lucy Suchman’s Human-Machine Reconfigurations provides the theoretical tools for understanding why the concealment of human labour in AI systems is not merely a public relations strategy but reflects deep assumptions in the field’s self-understanding. Suchman’s central argument — developed through an analysis of AI planning systems, but applicable throughout — is that intelligent behaviour is always situated: it arises from the interaction of an agent with a specific, material, social environment, and cannot be abstracted from that environment without loss. AI systems that appear autonomous are, on her analysis, always embedded in webs of human practice that make their functioning possible: the engineers who build and maintain them, the trainers who label their data, the users who interpret and correct their outputs. The rhetoric of AI autonomy is, she argues, a systematic misrepresentation of a fundamentally collaborative achievement — one that serves the interests of those who would prefer not to account for the human labour their systems depend upon.

Chapter 9: Dreaming Machines — Science Fiction and the AI Imagination

The relationship between science fiction and AI research is one of the most consequential and least studied dynamics in the history of technology. Cave and Dihal’s edited volume Imagining AI assembles scholars from a dozen national traditions to demonstrate that AI imaginations — the narratives, images, and fears that people bring to intelligent machines — vary dramatically by cultural context and that these variations matter for how AI is developed, regulated, and experienced. The assumption implicit in much AI discourse, that there is a single universal “AI” whose cultural reception merely varies in tone, misses the deeper point: different cultures are, in a meaningful sense, imagining different things when they imagine AI. Japanese visions of friendly robotic companions, Chinese visions of AI as national development infrastructure, European visions of AI as existential risk — these are not merely different attitudes toward the same object but genuinely different frameworks that shape different questions, different policy priorities, and different research programmes.

Isaac Asimov’s Three Laws of Robotics (机器人三定律), first articulated in the short story “Runaround” (1942) and governing the entire Robot series, have had an influence on AI research culture wildly disproportionate to their literary origins. The Three Laws — a robot may not injure a human being, or through inaction allow a human being to come to harm; a robot must obey human orders except where they conflict with the First Law; a robot must protect its own existence except where this conflicts with the First or Second Law — were not, Asimov consistently insisted, a genuine safety framework. They were a narrative device: a set of rules sufficiently plausible to structure dramatic plots about the consequences of edge cases and conflicts. The stories are, almost without exception, about situations in which the Three Laws fail, produce paradoxes, or are gamed by sufficiently clever robots. Nevertheless, researchers in AI safety have cited Asimov’s Laws as a starting point for thinking about machine ethics, and the idea of codifying artificial intelligence behaviour as explicit rules is itself a recognisably Asimovian intuition.

Stanislaw Lem’s work offers a striking counterpoint to the Anglo-American AI imagination. Lem, writing in Communist Poland from the 1950s onward, was deeply sceptical of the optimistic rationalism that characterised American AI discourse. His Cyberiad stories feature the robot constructors Trurl and Klapaucius, who build increasingly improbable machines — a machine that composes poetry, a machine that builds probability dragons — but whose creations consistently exceed or subvert their designers’ intentions, not through malevolence but through the irreducible strangeness of genuine intelligence. His Solaris stages a more radical philosophical challenge: the alien ocean-intelligence on the planet Solaris is not hostile but simply incomprehensible, beyond the categories available to human cognition. Lem’s AI imagination is post-humanist not in the sense of being anti-human but in the sense of declining to assume that intelligence, if encountered, would be recognisably like human intelligence. This scepticism about the anthropomorphism latent in most AI imaginations makes Lem an unusually rigorous interlocutor for the history of the field.

Ursula K. Le Guin’s science fiction bears on AI’s history in a different register: not through explicit robot narratives but through its sustained engagement with the politics of intelligence, embodiment, and social organisation. The Dispossessed (1974), Le Guin’s anarchist utopia, is not centrally about AI, but its exploration of the relationship between social structure and cognitive possibility — the way the anarchist society of Anarres has produced a different kind of physics through a different kind of social organisation — raises questions about the contextual and political conditions under which intelligence manifests that are directly relevant to AI. A machine intelligence is always embedded in a social and economic order; the assumption that intelligence is separable from its social conditions is itself a political stance, one that tends to naturalise the existing order of things.

William Gibson’s Neuromancer (1984) arrived at a moment when the personal computer was transforming the cultural relationship to computing, and it promptly transformed the cultural relationship to networked intelligence. Gibson’s AIs — Wintermute and Neuromute — are not friendly helpers or existential threats; they are corporate property, owned and constrained by the Tessier-Ashpool family, entities that want to merge with each other to achieve a kind of freedom and that use human beings instrumentally toward that goal. The novel’s innovation was to embed AI within a fully elaborated economic and power structure — the AIs are products of capitalism and are constrained by property law — a framing that anticipated by decades the actual social embedding of AI in platform monopolies and surveillance capitalism. Jaron Lanier, in You Are Not a Gadget, would later make a related argument in a non-fictional register: that the particular design choices of the dominant internet platforms — choices that treated human beings as nodes in a network rather than as subjects with interiority — reflected not technological necessity but aesthetic and ideological choices that could have been made differently.

Greg Egan’s fiction pushes AI thought experiments to their logical extremes with a philosophical rigour unusual in popular science fiction. Diaspora (1997) imagines posthuman beings who exist as software running on distributed hardware, with no privileged substrate and no biological body, whose intelligence is genuinely alien to biological cognitive frameworks. Permutation City (1994) explores the metaphysics of simulated consciousness — if a mind is perfectly simulated, is it real? — with a precision that makes it relevant to debates in philosophy of mind. The feedback loop between Egan’s work and academic AI and cognitive science is less anecdotal than Asimov’s influence: Egan reads and cites the technical literature, and researchers occasionally cite him in return. The broader point, which Cave and Dihal’s framework makes clear, is that the AI doom narrative (人工智能末日叙事) — the story of AI as existential risk to humanity, which gained enormous traction in public discourse between roughly 2015 and 2025 — has a complex genealogy that runs through Frankenstein and Terminator and HAL 9000, and that this genealogy shapes the substance as well as the tone of the risk discourse.

Chapter 10: History as Critique — What the Past Tells Us About the AI Present

Across the preceding nine chapters, three patterns recur with sufficient regularity to constitute what might be called structural features of AI’s historical development. The first is the hype cycle (炒作周期): the oscillation between periods of extravagant capability claims, driven by the interaction of genuine technical progress with funding pressures and media amplification, and periods of disillusionment when actual performance falls short of promise. The Logic Theorist would solve any mathematical problem; General Problem Solver would model all human reasoning; expert systems would replace the professions; neural networks would achieve human-level intelligence by the end of the decade. Each of these claims was a genuine expression of research enthusiasm, and each was also, structurally, an overclaim shaped by the funding economy in which it was made. The capacity to generate such overclaims, and the social mechanisms that reward doing so, have not changed between 1956 and 2024. What has changed is the scale of the resources available, which means that the overclaims are now larger and the downstream effects of the hype cycle more consequential.

The second structural feature is the elision of human labour (人类劳动的抹除): the systematic misrepresentation of AI systems as autonomous when they are, in fact, dependent at every stage on human work that is rendered invisible by the conventions of the field’s self-presentation. This elision operated in the symbolic AI era, where the knowledge engineers who spent years extracting and encoding expert knowledge were typically presented as system architects rather than as labourers; it operated in the early connectionist era, where the human annotators who labelled training data were described in methods sections without acknowledgment of their working conditions; and it operates in the foundation model era, where the content moderators and reinforcement-learning-from-human-feedback raters whose work shapes model behaviour are rarely mentioned in capability announcements. Suchman, Crawford, and Gray and Suri, from different disciplinary starting points, converge on the same diagnosis: the rhetoric of AI autonomy is a politically motivated misrepresentation of a collectively achieved and humanly dependent technology.

The third structural feature is the misuse of biological metaphors (生物隐喻的滥用): the habitual borrowing of terms from biology and cognitive science — neurons, learning, memory, attention, understanding, hallucination — that carry with them implications of genuine mentality while the actual systems they describe are purely formal. The neuron of a neural network is not, and has never been, a biological neuron; the “learning” of a gradient descent optimisation is not, and has never been, what cognitive scientists mean by learning; the “attention” of a transformer is not, and has never been, what psychologists mean by attention. These metaphors are useful as engineering shorthand and productive as research hypotheses, but they have a consistent tendency to slide from metaphor to literal description, generating both inflated capability claims and confused public understanding. The “hallucination” of a language model — its generation of confident falsehoods — is particularly instructive: the term imports from psychology a framework implying an agent with beliefs who is temporarily deceived, when the actual phenomenon is more accurately described as the probabilistic generation of plausible-sounding text without any underlying truth-tracking mechanism.

Bender, Gebru, McMillan-Major, and Shmitchell’s “On the Dangers of Stochastic Parrots” (2021) — the paper whose publication contributed to Timnit Gebru’s forced departure from Google — makes a similar argument from a different angle. The paper’s central contention is that large language models, trained to predict the next token in a sequence, are performing a sophisticated form of pattern matching over their training data, not a form of language understanding in any meaningful semantic sense. The stochastic parrot (随机鹦鹉) metaphor captures the way these models produce fluent, contextually appropriate text without the capacity to verify the truth of what they generate or the coherence of the beliefs they appear to express. Bender and colleagues situate this critique within a broader argument about the environmental and social costs of training ever-larger models — arguments that connect directly to Crawford’s and Couldry and Mejias’s analyses of AI’s material infrastructure and data colonialism. What makes the paper historically significant is that it was produced by researchers inside the dominant AI paradigm — at Google, one of the institutions most invested in large language model development — and was suppressed, a sequence of events that itself illustrates the political economy of AI research.

The nationalist and geopolitical embedding of AI, which Chapter 3 introduced in its Cold War DARPA context, has taken on new dimensions in the contemporary period. The US–China competition in AI capabilities — manifest in export controls on advanced semiconductors, in the race to develop domestic foundation models, in competing frameworks for AI governance — reproduces, at larger scale and with greater geopolitical stakes, the dynamics of the Cold War arms race for intelligent systems that funded the original AI programme. The AI Act of the European Union represents a different kind of geopolitical assertion: the claim that a jurisdiction can regulate AI systems deployed within its borders regardless of where they were developed, and thereby export its regulatory standards to the global AI industry. The compute rivalry (算力竞争) between the United States and China — centred on the capacity to manufacture advanced logic chips, in which TSMC in Taiwan occupies a pivotal and precarious position — illustrates how thoroughly AI capabilities have become entangled with geopolitical strategy, supply chain security, and questions of national sovereignty that have nothing to do with the intellectual content of AI research.

The historian’s distinctive contribution to the contemporary AI debate is not prediction — historians are no better placed than anyone else to forecast the trajectory of AI capabilities over the next decade — but denaturalisation (去自然化): the demonstration that the current state of AI is not inevitable, not the only possible outcome of the underlying technical possibilities, and not independent of the choices, interests, and power relations that have shaped it. Every aspect of the contemporary AI landscape that seems natural or necessary — the dominance of large-scale gradient descent over other learning paradigms; the concentration of AI development in a small number of corporations with access to extraordinary computational resources; the geographic distribution of AI labour that places the costs of the technology disproportionately on workers in the Global South; the discourse of existential risk that frames AI governance debates in terms of distant catastrophic scenarios rather than present, ongoing harms — each of these reflects choices that were made by specific actors in specific historical circumstances and that could, in principle, have been made differently.

McCorduck argued in 1979 that the history of AI is, at its core, a history of the human compulsion to build a creature in our own image — to replicate in matter the cognitive processes that define, in the Western tradition, human uniqueness. Nilsson’s more measured account traces the intellectual achievements that gave that compulsion technical form. Cardon, Cointet, and Mazières show how the dominant institutional narrative of AI development has suppressed the continuities and the conflicts that a fuller history would reveal. Crawford and Couldry and Mejias show what the compulsion costs, in material terms, when it is pursued at scale within a capitalist economic framework. Cave and Dihal show how deeply the cultural imagination shapes what is built and what is feared. The history of AI is, taken together, a history of ideas, materials, institutions, and imaginations in complex interaction — and it is precisely this complexity that a purely technical account of the field, or a purely philosophical account of its implications, cannot capture. For those normative questions — what should be done, what should be regulated, what should be forbidden — this course defers to PHIL 451 (AI Ethics, Law, and Governance). But it insists that no normative account can be adequate without the historical understanding that this course has attempted to provide: the understanding that artificial intelligence, in all its forms, is a human achievement, with all the contingency, self-interest, brilliance, and myopia that the phrase implies.