You cannot discover an inhabited land. Otherwise I could cross the Atlantic and ‘discover’ England.
Dehatkadons, a traditional chief of the Onondaga IroquoisIn his 1986 book The Society of Mind, Marvin Minsky outlined a theory of intelligence informed by three decades of research on artificial intelligence, a field he helped found. Minsky stated his aspiration to emulate the foundational insights of Galileo and Newton by reducing the dynamics of the mind down to the simplest possible terms. He theorized that intelligence was the product of many simple mental processes, or modular ‘agents’, joined by cross-connections that could be modelled mathematically via the logic of, in his words, a ‘society’.Footnote 1 Minsky did not specify which type of society, nor from where within that social structure its all-important dynamics were to be judged. For him, a ‘society’ was a philosophical abstraction, an idea that bore universally as if its logics were self-evident, meaningful and constant.
Minsky was neither the first nor the last in his field to formalize the science of mental behaviour using the vocabulary and dynamics of social phenomena.Footnote 2 Other prominent researchers turned to social logics to render neural complexities in ways they deemed susceptible to mathematical treatment. Herbert Simon, a driver of the symbolic paradigm, borrowed from the study of administration on the assumption that decision making in large human organizations had gone unchanged for ‘at least four thousand years’.Footnote 3 Frank Rosenblatt, a driver of the connectionist paradigm, built his perceptron models on theory borrowed from Hayekian economics. What this and other relevant metaphors had in common was the motivating presumption that the mind was an orderly thing; that it lived inside an individual's brain; and that it followed an implicit, reliable ‘logic’ that could be convincingly modelled with modes of computation derived from the observation of social events.
These instances from the earliest days of AI point to a broad category of non-neural metaphors that go largely unacknowledged in existing genealogies of the overlap between digital computer programming and neural activity. This omission is surprising given that contemporary machine-learning researchers still turn to subjective social logics to model neural phenomena, including by benchmarking progress on machine intelligence against abstract strategy games (such as chess, go, StarCraft) and by modelling software ‘agents’ into social formations premised on theories of adversarial human relationships (such as game theory, general adversarial networks).Footnote 4
At first glance, the simulation of subjective social logics and their subsequent deployment into knowledge management infrastructures reads as baldly political. This is especially true when such applications are dressed, via selective metaphors, as neural, or framed, via selective histories, as apolitical and apparently self-evident and universal. That the intentions of state and industrial actors who fund(ed) such research often remain unacknowledged adds another level of contingency for historians to unpack. Sensitive to the overlap between social and computational logics in the prehistory of digital computing, Agar advocates reading the history of information technologies and modern state formation side by side. ‘Several of the most important moments in the history of information technology revolve, rather curiously, around attempts to capture, reform, or redirect governmental action’, he argues.Footnote 5
With Agar's insight in mind, this paper trials a heuristic – animo nullius, for ‘no persons’ mind’ – to suggest that uncritical histories of computer-simulated ‘intelligence’ and ‘learning’ advance acts of seizure using metaphor. In choosing this framing, I invite reflection on parallels between historical colonialism and data colonialism, specifically parallels with terra nullius, the Roman legal expression for ‘land that belongs to no one’, which remains the spurious philosophical basis for a swathe of contemporary international law. Terra nullius – the concept, if not the name – was repurposed opportunistically after the fifteenth century as a rescripting ‘principle’, ‘doctrine’, ‘policy’ and legal ‘fiction’ by other European colonial empires to dispossess Indigenous peoples of their land.Footnote 6 Herein, I question how the authority of foundational language about machine ‘intelligence’ developed, often ambivalently, in the mid-twentieth century is now being repurposed to normalize enclosure of the data commons. ‘Just as historical colonialism over the long run provided the essential preconditions for the emergence of industrial capitalism’, write Mejias and Couldry, ‘we can expect that data colonialism will provide the preconditions for a new stage of capitalism that as yet we can barely imagine, but for which the appropriation of human life through data will be central’.Footnote 7
This paper proceeds in two parts. First, I recount how the founders of ‘AI’, broadly construed, proselytized that mathematics, logic, computation and statistical formalisms would render mental behaviour legible and amenable to simulation. Research in the mid- to late 1950s on what later became ‘AI’ positioned the human mind as universal, programmable and knowable, a ‘domain’ to settle through the abstract instrumentation of positivist Western technoscience. Figures in symbolic AI like Marvin Minsky, John McCarthy, Herbert Simon and Allen Newell, and in machine learning like Frank Rosenblatt, invoked the implied authority of neural metaphors, as if the correspondence was literal, even while they oscillated over the degree to which neurophysiology should or could serve as the empirical basis of their work. I explore how ties to US-based state-spirited institutions in this period helped them to advance this research.
The second section traces the lasting effect of these rhetorical moves through AI's institutionalization in the twentieth and twenty-first centuries. I show how large American technology corporations repurposed this ambiguously ‘neural’ rhetoric in the 2010s and 2020s to build and expand markets for the sale of proprietary software and file-hosting services. In the present day it is global corporations, rather than upstart American researchers, that equivocate over the authority (and, with it, the historical genealogy) that underwrites AI. Hanging in the balance are emerging societal norms about the degree to which ‘AI’ outputs are provably transformative and protectable via copyright or other law. As with terra nullius, both those who make and those who contest such claims will look to history for adjudication. I call attention to animo nullius to invite historians of technology to pay closer attention to the ways in which AI's theoretical origins built on the edifice and practices of capitalism more substantially than on neurophysiology, even when neural metaphors were applied.
The origins of AI
In the mid-1950s, researchers in the United States melded formal theories of problem solving and intelligence with another powerful new tool for control: the electronic digital computer. Several branches of Western mathematical science emerged from this nexus, including computer science (1960s–), data science (1990s–) and artificial intelligence. Nascent efforts to formalize the dynamics of human thought elaborated what Roberto Cordeschi has called ‘the culture of the artificial’, in which mental processes were seen as independent of organic structures and could be verified as such.Footnote 8 During the first half of the twentieth century, this culture moved from the fringes of psychology to mainstream research on cognitive science.
In the 1950s, this tradition converged with the advent of digital electronic computing. Between 1952 and 1957, new computational mechanisms like ‘assemblers’ and ‘compilers’ consolidated the prohibitively laborious instruction sets used to ‘program’ a digital computer into ever more accessible and coherent programming ‘languages’. In the US, industrial actors like IBM advanced and leveraged these consolidations to usher new paying audiences into standardized modes of computing. In 1958, the Communications of the Association for Computing Machinery, a new journal, fielded letters from an emerging class of professionals and theorists over what to name their profession. Suggestions included ‘synnoetics’ (‘science of the mind’ in Greek), ‘computer science’ and the generalizable notion of ‘comptology’ (e.g. nuclear comptologist, logistics comptologist).Footnote 9
Freighted in these terms were deep claims about the explanatory potential of information-processing metaphors to describe neural phenomena, claims that often attracted explicit attention at first, only to be silently absorbed over time.Footnote 10 Elsewhere I have shown that despite their differing foundational approaches to AI and machine learning, key researchers blurred the line between metaphorical and literal descriptions of the intersections between computation and cognition.Footnote 11 Anthropomorphic language was immediately controversial. The 1953 book Automatic Digital Calculators apologized for using ‘memory’ in relation to computing and acknowledged that terminology in the field had ‘not yet stabilized’.Footnote 12 Prominent American engineers and mathematicians like Claude Shannon and John von Neumann bristled at a new generation's hubristic use of mind-as-computer metaphors and took pains to articulate their limits. Shannon dismissed John McCarthy's attempt to include ‘intelligence’ as a relevant descriptor in Automata Studies, their co-edited 1956 volume.Footnote 13 At the 1948 Hixon Symposium on Cerebral Mechanisms in Behavior, von Neumann positioned his mathematical work on neural phenomena as that of an ‘outsider’ prohibited from meaningful speculation on anything other than the dynamics of idealized versions of elementary physiological units.Footnote 14 In the early 1950s, professional etiquette compelled Marvin Minsky to use scare quotes around his early mathematical notions of machine ‘learning’.Footnote 15
By the late 1950s, these quotes had disappeared. Machine ‘learning’ became machine learning. To assert the perceived legitimacy of this transition from metaphorical to vaguely literal descriptions of human capacities in machines, these men permitted clannishness, self-aggrandizement, speculative rhetoric, fluid definitions of key terms and poor citation practices, actions that drew attention to questions of how to accomplish such aims and away from whether they were well founded.Footnote 16 Even as they disagreed over the correct measure of fidelity to biological phenomena in their modelling and the appropriate rhetoric for their results, Simon, Rosenblatt, McCarthy and Minsky positioned brain modelling as ripe for a grand theory of cognition, a possibility they all presumed existed. ‘Once one system of epistemology is programmed and works no other will be taken seriously unless it also leads to intelligent programs’, McCarthy wrote in his diaries. ‘The artificial intelligence problem will settle the main problems of epistemology in a scientific way.’Footnote 17 Even Rosenblatt, who called AI researchers his ‘loyal opposition’ because their deterministic ‘paper exercises’ contravened biological evidence, neglected to state the limits of his own instrumentalist brand of statistical modelling. The difference between his commitments to biological fidelity and those of his ‘opposition’ was one of degree, not of kind.
Internalist accounts of early brain model theory have tended to forgive these common commitments as customary. Recurring emphasis on narratives of discontinuity, such as on funding ‘winters’ or the fabled 1956 Dartmouth Summer Research Project, where AI gained its name, have obscured broader methodological and structural continuities.Footnote 18 One recurring, yet distracting, tendency in existing histories is a focus on the swing of fashions between two apparent poles: symbolic reasoning (such as the simulation of problem solving) and neural-network research (such as the simulation of learning). I eschew that dichotomy here to revisit what these pursuits ultimately had in common, such as the foundational beliefs that complex mental processes were susceptible to mathematical formalization, that digital electronic computers were the appropriate tool for that job, and that the product of this enterprise would be an abstract ‘language’.
Historians have questioned another commonality: the role that generous institutional patronage played in initial efforts to formalize the mathematics of thought. Kline illustrates that Dartmouth was not the launch pad it has been described as: funding was halved, participants dropped out or left after a day, and the group's diverse research foci sprawled.Footnote 19 That the workshop did not establish AI draws our attention to other funding and organizational dynamics that helped to bolster the field thereafter. Jean Pierre Dupuy considers the fates of two adjacent research areas. He describes cybernetics as a ‘failure’ in comparison to information theory, which found a professional home in the Institute of Electrical and Electronics Engineers.Footnote 20 Andrew Pickering argues that cybernetics failed because it lacked similar backing. The Macy Conferences and Ratio Club lacked the capacity to formally train students, set a stable research agenda or grant degrees, hampering the ability of cybernetics to propagate, at least initially. He writes,
When we think of interdisciplinarity we usually think of collaborations across departments in the university … The centre of gravity of cybernetics was not the university at all. Where was it then? The simplest answer is: nowhere … Cybernetics flourished in the interstices of a hegemonic modernity, largely lacking access to the means of reproduction: the educational system.Footnote 21
This was not the case for AI. Less than a decade after the Dartmouth meeting, the AI group at MIT (and then at Stanford and Carnegie Mellon) operated with an annual budget purported to be in the millions – a remarkably brief transformation into institutionalization that historians have yet to properly interrogate.Footnote 22 What I seek to argue here is that alignments between ‘AI’ and corporate prerogatives eased its access to the means of reproduction. My focus in this gestational period of the mid- to late 1950s is on support provisioned through IBM, the Rockefeller Foundation and (albeit to a lesser extent) Bell Laboratories, whose reasons for involvement encompassed more than a search for a grand theory of cognition. Their incentives included the potential for commercial expansion, the standardization of manufacturing techniques and support for American dominance abroad.
The institutionalization of AI
The Rockefeller Foundation, which funded the Dartmouth meeting credited with giving ‘AI’ its name, was in the mid-1950s a ‘parastate’ or ‘“state-spirited” organization sitting at the heart of the emerging East Coast U.S. foreign-policy establishment’.Footnote 23 Along with other East Coast philanthropic foundations born from the exploits of nineteenth-century corporate titans (such as Ford, Carnegie), the foundation was keen to enlist ‘ivory tower’ academics to secure global peace by promoting American hegemony against the USSR. This overarching strategy goes part of the way to explaining why the foundation elected to fund contentious research. In 1955 Warren Weaver, then director of its Division of Natural Sciences, expressed misgivings when McCarthy approached him to fund the Dartmouth meeting. He directed McCarthy to Robert Morison, head of Rockefeller's Biological and Medical Research Division, but told Morison privately, ‘I am very doubtful that the RF ought to do anything about this.’Footnote 24 Morison, too, expressed qualms. ‘This new field of mathematical models for thought … is still difficult to grasp very clearly.’Footnote 25
Morison approved a $7,500 grant, half the requested $13,500. He questioned whether a seminar format was the right mechanism for discovery since mathematical theories of brain functions remained ‘pre-Newtonian’.Footnote 26 His hesitation was not purely subject-based: Kline notes that Morison and Weaver had by then co-funded research on mathematical biology undertaken by Wiener and Rosenblueth, which led to the book Cybernetics.Footnote 27 Central to his doubt was the group's neglect of existing experts and established knowledge practices. Morison urged the organizers to include psychologists like Hans-Lukas Teuber and Karl H. Pribram, ‘if only for the purpose of keeping the group from speculating too wildly on how the brain might work’.Footnote 28 While psychologists did not ultimately participate, Morison nonetheless agreed to fund the meeting as a ‘modest gamble for exploring a new approach’.Footnote 29
IBM's initial participation in ‘AI’ research, broadly construed, was not wholly dissimilar to the Rockefeller Foundation's. In the decade after the Second World War and in keeping with its name, IBM exercised ambitions to broaden its reach internationally. From 1914 until his death in 1956, the company's leader, Thomas Watson Sr, a uniquely talented salesman, pursued growth mercilessly. Echoing Adam Smith's view that international trade was an instrument by which to spread civility and shared prosperity abroad, Watson emblazoned the slogan ‘World Peace through World Trade’ on IBM's New York headquarters. Paju and Haigh describe the company in the post-war reconstruction period as having operated in a distinctly imperial fashion, planting flags on prominent buildings in major cities, erecting and retaking control over national subsidiaries (ensuring that none were so self-sufficient as to defect or be annexed), and pursuing trade-heavy ‘interchange’ manufacturing across Europe.Footnote 30
The commercial case for automatic coding – and, with it, proto-AI – had not been obvious to IBM leadership. Nathaniel Rochester, an employee and co-convener of the Dartmouth workshop, claimed in personal records that he and his team had ‘sold the idea’ of automatic coding to management as a patriotic act after IBM reactivated its military product unit in response to the Korean War.Footnote 31 Having been hired by the navy in 1947 to construct the arithmetic unit of Whirlwind I, amongst the first digital computers in the US, Rochester was particularly well positioned to recognize the prospects of the field. He spent the rest of his career at IBM and recalled his initial surprise to learn that the company had ‘no intention of venturing into stored-program machines’.Footnote 32 Once approved, responsibility fell to him to design and construct their first large-scale scientific computer, the Defense Calculator, released in 1952 as the IBM 701. By 1954, he oversaw a staff of 450 people.Footnote 33 After 1956, when Thomas Watson Jr became CEO, IBM's involvement in digital electronic computing deepened significantly as adoption of the technology expanded.
Support from IBM in the mid- to late 1950s became important to the broader AI project. In February 1955, McCarthy left a brief teaching position at Stanford to become assistant professor of mathematics at Dartmouth College. Despite a long-standing interest in computing, he had never actually tried to program a computer until that year.Footnote 34 This was partly beyond his control; neither Stanford, where he had been based, nor Dartmouth, at which he had just arrived, owned a stored-program computer for him to experiment with. To his good fortune, Philip M. Morse, the ‘founding father of operations research’, had just convinced MIT to open a new three-storey IBM-funded Computation Center to twenty-five sister colleges and universities across New England.Footnote 35
As the chosen representative of Dartmouth College, McCarthy was introduced to Rochester, who invited him to spend the summer of 1955 at IBM. A press release from December 1956 described the centre as the ‘largest and most versatile data processing facility yet to be made available primarily for education and basic research’.Footnote 36 As Dartmouth's appointee, McCarthy was invited to make use of the 18,000-square-foot facility and twenty-five-member staff.Footnote 37 As part of this arrangement, McCarthy was asked to ‘sell the idea’ of computing to his peers in academia. ‘I think that sooner or later programming will become as basic a part of a scientific education as calculus, and strongly advise learning it even if you don't see immediate application to your problems’, he wrote.Footnote 38 New England academics were receptive; 40 per cent of his students under the IBM flagship were professors. Access to the centre helped McCarthy deepen his collaborations with Minsky. In September 1957, he became an MIT Sloan fellow in the physical sciences, a position that allowed him to relocate to the centre full-time. In September 1958, he and Minsky co-founded the MIT Artificial Intelligence Project with two programmers, a secretary, a typewriting machine and six graduate students.
The search for a theory of ‘intelligence’ – and a computational language to express it – was expensive, but alignment with state and corporate institutions, and their motivating logics, helped settle the bill and advance the project into a stand-alone discipline. The two cases just outlined speak to a general alignment between mind-as-computer research and state-industrial aims. Of the twenty or so core participants at the Dartmouth workshop, seven were affiliated with IBM, three with Bell Laboratories, two with the RAND Corporation and twelve with MIT, one of America's largest non-industrial defence contractors at the time. IBM, Bell Laboratories and other home organizations paid staff salaries to attend. Even Rosenblatt, who self-identified as in opposition to AI, secured continued placements in the 1950s and 1960s through grants via the Institute for Defense Analysis and the Office of Naval Research.Footnote 39
These affiliations galvanized the emerging research community. At the Dartmouth meeting, Herbert Simon, Allen Newell and J. Clifford Shaw introduced the results of their Logic Theory Machine, a virtual machine they designed at RAND in 1955–6 to prove thirty-eight theorems from Principia Mathematica. They positioned their results as evidence of having modelled human problem solving using an electronic digital computer. Ray Solomonoff, also in attendance, described the RAND machine as having solved ‘the demo to sponsor problem’, meaning the felt need to oblige funders with a proof of concept.Footnote 40 Minsky, too, noted that their ‘inspiring progress’ had had ‘considerable effect on the direction of our work’.Footnote 41 McCarthy called the RAND trio ‘the stars of the show’.Footnote 42 The Logic Theory Machine has since been designated, albeit contentiously, ‘the first AI program’ by existing histories of the field.Footnote 43
This group have collectively been described as the ‘founding fathers’ of artificial intelligence. As in historical accounts of colonialism in the United States, progenitor narratives describe early developments in terms of personalities and ideals, not access to resources. The term ‘founding fathers’ was likely introduced into AI's historiography in 1979 by McCorduck's Machines Who Think, which ‘forged the template for subsequent histories’, according to Mirowski.Footnote 44 McCorduck treated structural contingencies as extraneous, as Cohen-Cole has critiqued Margaret Boden for doing in her oft-cited Mind as Machine.Footnote 45 In the mid-1950s, RAND possessed perhaps the largest computing facility in the world. This and other corporate affordances to proto-AI, however, are treated as incidental, not constitutive. Neither IBM nor RAND figure in the list of AI's ‘founding fathers’. Yet in contrast to cybernetics, institutional support for key researchers in what became AI was the rule, not the exception, during its rapid acceleration into an academic discipline.
In a 1957 lecture to the Operations Research Society of America, Simon and Newell self-reported what can be understood as an exception to this historiographical framing. They positioned their prototype as a milestone in the entangled histories of capitalism and manufacturing, stating,
For an appropriate patron saint for our profession, we can most appropriately look back a full half century before Taylor to the remarkable figure of Charles Babbage … He was one of the strongest mathematicians of his generation, but he devoted his career to the improvement of manufacturing arts, and – most remarkable of all – to the invention of the digital computer in something very close to its modern form.Footnote 46
Their speech credited Adam Smith as the true inventor of the computer. While Smith's theories on capitalism were not computational in practice, they argued, they were computational in principle, as Lorraine Daston corroborates.Footnote 47 Simon explained that thinkers like Gaspard de Prony and Charles Babbage had simply translated Smith's ideas into hardware through iterative stages of development.
Simon and Newell's institutional basis reflected this connection between technological engineering and social engineering. The RAND Corporation's Systems Research Laboratory was among the world's first experimental laboratories in management science. Further, as I have shown elsewhere, Simon, Newell and J. Clifford Shaw borrowed novel conceptual tools from precedents in administrative theory and formal logic to develop a functional machine language for the simulation of problem solving in a computer.Footnote 48 They repurposed elements of Simon's influential 1947 text Administrative Behavior such as decision premises and means–ends analysis to shrink the program's exposure to the complexity of the real world. Behavior posited that all activity in an organization could be reduced to explainable decision-making processes so long as the organization had a coherent, unchanging motive, specifically the capitalist prerogative to maximize profit. In other words, the Logic Theory Machine was a model of human behaviour, not of biology. Further, it was a model of a certain type of behaviour, namely human problem solving, as seen through the prism of early twentieth-century symbolic logic and post-war American administrative logic.
The naturalization of AI's conceptual origins
Internalist historiographies by McCorduck and Boden bury these connections under what I argue are the rescriptings of animo nullius, which foreground AI's empirical basis in neural – rather than social – phenomena and, in a related move, position AI as an academic science more than an industrial art. Ironically, Simon and Newell had resisted the term ‘artificial intelligence’ in the 1950s on the ground that it had been their results at RAND, which they categorized as ‘complex information processing’ and ‘the next advance in operations research’, that had lent validation to the abstract theories of Minsky, McCarthy and others. By the late 1950s, however, for no clear reason other than the traction the term ‘AI’ had gained at MIT and elsewhere, Simon and Newell began to use it, ‘heuristic programming’ and ‘complex information processing’ interchangeably. They, too, succumbed to the tendency to equivocate about the boundaries between metaphorical and literal terminology, between intelligence and problem solving, and between brains and computers.
These formative decisions mattered. In the United States, machine intelligence gained a new civility and sociability outside academia, spreading from research centres to newspapers, television and magazines. The vocabulary that early researchers chose to describe new techniques informed Americans’ still plastic understandings of what was possible, and indeed desirable, in the emerging information age. Even if the earliest forms of ‘AI’ were definitively premised on a search for intelligence, which was inconsistently the case, it was also enmeshed in studies of search, storage and problem solving, as well as learning, vision, language, logistics and operations research, ways of knowing far removed from the sublimity of ‘intelligence’ and the brain.
By the late 1950s, allusions to neural fidelity were less concerned with any methodological commitment to close study of the brain than with a vaguely mentalistic way of positioning the field. Researchers were no longer sheepish about the need to qualify neural language. By the 1960s, ‘automata’ and ‘brain models’ became ‘the artificial intelligence problem’ and then just ‘artificial intelligence’. I speculate that three broad trends prolonged ambivalence in AI about the exact meaning of suggestive neural metaphors. The first concerns patronage norms. In 1982, James Fleck attributed McCarthy, Minsky, Simon and Newell's sustained influence over the AI ‘establishment’ to their monopolized access to US military funding after 1963.Footnote 49 Fleck argued that this concentration, and the access to rarefied computational tools it enabled, was central to their mythos and to that of the Dartmouth meeting.Footnote 50 ‘Until the early 1970s essentially all of DARPA's AI grants were given to the Massachusetts Institute of Technology and Stanford University’. Yarden Katz adds that ‘between 1970 and 1980 MIT, Stanford, and Stanford Research Institute (SRI) received over 70 percent of the agency's AI funds’.Footnote 51
Further research is needed to adequately disentangle the influence of commercial versus military funding in AI in the United States after 1960.Footnote 52 Regarding its intricate overlaps, consider that during the 1950s IBM earned more from its military contracts on the Semi-Automatic Ground Environment (SAGE) computing system than it did leasing out its own machines, and that IBM simultaneously contracted RAND's System Development Corporation (the lab where Newell, Simon and Shaw had met) to develop SAGE's software.Footnote 53 Colin Garvey's history of Japan's 1982 Fifth Generation Computer Systems project provides a telling point of contrast to evolving norms in the United States. Fifth Generation was ‘the first national, large-scale AI R & D project to be free from military influence and corporate profit motives … [realizing], in many ways, the norms of science often proclaimed – but rarely lived up to – by Western democracies’.Footnote 54 In response, the US government launched its billion-dollar Strategic Computing Initiative (SCI), which ran from 1983 to 1993. By 1985, it had allocated 50 million dollars each to public research laboratories and to industry.Footnote 55 Shiman and Roland's history of SCI as a search for ‘machine intelligence’ clarifies that the US government's aim was to ‘remain competitive in the production of critical defense technologies’.Footnote 56 By this view, commercial and military aspirations for AI could, by the 1980s, be taken as two parts of the same whole.
A second trend, congruent with the first, was the rapid adoption of computing tools over the second half of the twentieth century. I speculate that the practicalities involved in this prodigious spread sedimented key terminology about computing that might otherwise have remained pliant to scholarly critique and refinement.Footnote 57 While only two digital electronic computers were in operation in the United States in 1950, 243 machines were operational in 1955, with 5,400 in 1960, 25,000 in 1965 and 75,000 in 1970.Footnote 58 After 1960, mention of ‘automatic programming’ and ‘automatic coding’ techniques became discussions about ‘software’.Footnote 59 How this broad spread influenced rhetoric about ‘AI’ has yet to receive sustained historical treatment.Footnote 60
A third trend prolonging ambivalence about suggestive neural metaphors concerns the sustaining effect of isomorphisms between methods in AI and mainstream economics. Emerging scholarship suggests that, at a high level, efforts at digitization followed efforts at financialization. Devin Kennedy, for instance, explores how foundational theory for how to sort and manage complexity in computer science mirrored approaches for how to sort and manage capital.Footnote 61 Similarly, with regard to the history of software development after the 1970s, Laine Nooney argues, ‘Throughout it all, we can clearly see the intractable role financial speculation and the construction of markets played in people's desire to even imagine what shape innovation might take.’Footnote 62 In relation to ‘AI’, Rosenblatt's perceptron, a foundational contribution to machine learning, explained the irreducibly complex phenomena of the mind using Hayek's notion of decentralized market structures. Tracing the mathematics of ‘minds’ and ‘markets’ forward, one finds similar overlaps in the history of the loss function, in the Black–Scholes options pricing model and in network science generally.Footnote 63
That the search for a theory of machine intelligence derived from the empirics of social phenomena, and not from neurophysiology alone, became stubbornly apparent as research vogues shifted after the 1960s. In 1976, Newell and Simon posited, perhaps at the peak of symbolic AI research, that ‘a physical symbol system has the necessary and sufficient means for general intelligent action’.Footnote 64 Unconvinced by the feasibility of this goal, those in the emerging, alternative tradition of ‘expert systems’ in the 1960s to 1980s set out to encode human knowledge, not human reasoning, using databases of ‘if–then’ rules. Edward Feigenbaum, who coined the term ‘expert systems’, wrote, ‘The problem-solving power exhibited by an intelligent agent's performance is primarily the consequence of its knowledge base, and only secondarily a consequence of the inference method employed. Expert systems must be knowledge-rich even if they are methods-poor … The power resides in the knowledge.’Footnote 65
In the 1970s to 1990s, expert systems found both novel commercial uses and piercing critique. Robert Cooper, director of DARPA, took pains to ‘transition’ expert-systems research from the lab to the marketplace and to connect university researchers to industry.Footnote 66 Critique followed. In the 1990s, the anthropologist Diana E. Forsythe challenged researchers’ self-reported abilities to extract, formalize and model so-called ‘domain-specific’ knowledge using interviews with human experts.Footnote 67 ‘Whereas anthropologists devote considerable energy to pondering methodological, ethical and philosophical aspects of field research’, she wrote, ‘knowledge acquisition seemed to be undertaken in a rather unexamined way. Asked how they went about the task of gathering knowledge for their expert systems, the knowledge engineers I met tended to look surprised and say, “We just do it.”’Footnote 68
While expert systems fell out of favour in AI research in the 1990s, the positivist assumption that rendered knowledge a self-evident, structured and accessible ‘thing’ that could be ‘extracted’ carried into the next vogue of AI – machine learning – through technical vocabulary about ‘domains’.Footnote 69 This term finds ubiquitous use in contemporary machine learning. Ribes et al. position its rhetorical power as a claim to ‘identify, demarcate and characterize spheres of worldly action or knowledge’ that in daily use implies, necessarily, that ‘there is a more general, even universal, method or technique … [that can be used] across many, and sometimes all, domains’.Footnote 70 In his study of data positivism since the Second World War, Matthew Jones points to Rosenblatt's perceptron theory as epitomitizing a branch of instrumentalist statistics that sought technique-as-explanation, meaning functions that fit the data rather than functions that fit a corresponding law of nature.Footnote 71 This instrumentalist view of knowledge distanced machine-learning researchers from the empirics of the material world, even as they draped their techniques in the suggestive language of neural mechanisms, as the name ‘perceptron’ indicates.
In the 2010s, American technology conglomerates mobilized to fuel widespread media speculation of a looming ‘AI revolution’ in which Yarden Katz has argued that AI doubled for ‘a confused mix of terms – such as “big data,” “machine learning,” or “deep learning” – whose common denominator is the use of expensive computing power to analyse massive centralized data’.Footnote 72 Behind this marketing, Gürses argues, was a concerted push by the global finance industry to position American technology conglomerates as a safe harbour for large-scale institutional investment following the collapse of real-estate investments during the 2007–8 financial crisis. To satisfy shareholders’ expectation of a return, the largest cloud service providers – Amazon, Google, Microsoft – turned to the ‘organizational capture’ of hospital, university and government tech budgets, among other small and medium-sized businesses.Footnote 73 As with Facebook's efforts to monopolize the marketplace for global advertising budgets, American cloud providers pressed to capture organizations’ technology infrastructure budgets by enlisting them into pay-as-you-go hardware and software services accessed via distributed data centres known as ‘the cloud’. Per Katz, rhetoric about ‘AI’ figured centrally here as a euphemism for private cloud infrastructure.Footnote 74
To incentivize and prolong adoption of their proprietary infrastructure, tech conglomerates invested heavily in the development of machine-learning techniques that could, in operation, enclose regions of common knowledge online. This time, they indulged AI's sociotechnical imaginary to veil acts of seizure as acts of novel transformation or discovery.Footnote 75 Blackwell calls this ‘institutionalized plagiarism’.Footnote 76
Fundamental to such seizures is a lack of case law, which compounds existing confusion over the provenance of authorship, the role of human contributors in creating input data, and the degree to which generative AI outputs are original and transformative. A class action lawsuit filed against OpenAI in July 2023 rejects the idea that their ChatGPT model is capable of genuine novelty. ‘A large language model's output is … entirely and uniquely reliant on the material in its training dataset’, they reason.Footnote 77 According to this view, OpenAI has made illegal use of copywritten material. The company's response, in August 2023, situates the case around the science of inference. ‘As Plaintiffs acknowledge, the ability of digital software to [interpret user prompts and generate convincingly naturalistic text outputs] reflects a legitimate scientific advancement in the field of artificial intelligence, which seeks to “simulate human reasoning and inference.”’Footnote 78
This lawsuit captures aspects of the evolving character of animo nullius after the 2010s. Unlike in the mid-century, it is now corporations, alongside some researchers, who stand to benefit materially from equivocation over the perceived authority and intellectual genealogy that underwrites AI. OpenAI's valuation fluctuates in line with this perception; at the time of going to press it is the most valuable start-up in America. Post-war IBM sought only outputs from AI, not its brand. Another difference is that debate over these scientific merits has shifted from labs and markets to courtrooms, too. That territories seized under terra nullius remain stolen today, centuries later, bears out how misinterpretations of the past can calcify into law and prolonged settlement. Discerning publics – including judges, legislators and dispossessed authors – would be forgiven for assuming, naively, that the statistical ‘intelligence’ and ‘learning’ on display in post GPT-era systems derive from, say, the empirics of Darwin rather than, in my account, the contingent social order elaborated by Smith. What else would history suggest?
Birch, who theorizes the rise of assetization in modern technoscience, characterizes this form of economic system as rentiership. Unlike with the valuation of a commodity, which is generally extinguished upon its consumption, ownership and control of an asset extend one's rights to copies and derivatives, which adds scope for control and extraction. ‘No one can tell you how to wear a coat you have bought, but they can tell you how to use the copyrighted music or data they have sold you’, he writes.Footnote 79 Animo nullius figures in rentiership economics because the value of an asset requires ‘active and ongoing organization, governance, and management … it is not a passive process’.Footnote 80 In my account, animo nullius animates AI's value as an asset by granting those who invoke it an outwardly credible and durable (that is, historical) claim to the authority of a prestige metaphor.
Data colonialism
In my introduction, I cited Mejias and Couldry's interest in how these emerging branches of capitalism intersect with data colonialism. I will end by offering three conjectural parallels that emerge from my account. The first is an observable convergence between state and corporate interests over privatizing the costs of speculation (and initial seizures) amidst imperial expansion. State involvement is central to the durability of the extractive project outlined above because, per Birch, ‘assets are legal constructs, in that ownership and control rest on the state enforcement of property and control rights’.Footnote 81 In Empire of Cotton, Sven Beckert argues that the commodification, processing and global trade of cotton illuminate the mechanics of historical colonialism and the Atlantic slave trade, wherein empires backed merchants and colonial corporations with military force to facilitate expansion.Footnote 82 I suggest that the race to enclose the knowledge commons and assetize key analytical processes to the benefit of the global rentier class (and their respective nation states) helps to structure how data colonialism will shape capitalism in the era ahead.
A second parallel, which puts tension on the first, is states’ backing of corporate self-governance and claims to the commons. State-backed corporate agency was decisive to the settler colonial project in ‘New England’, the land upon which, centuries later, AI would initially be developed at MIT. In that instance, the mechanics of self-governance in colonial corporations helped to prototype and, upon maturation, to stabilize and cohere America's legislative independence from the British Empire.Footnote 83 Possession via occupation succeeded wildly for the Puritans of New England, and for the Massachusetts Bay Company.Footnote 84 Suggestive of a disanalogy with terra nullius, however, Birch notes that it is contract law rather than property rights that is central to the ownership of intangible assets such as machine-learning models.Footnote 85 In any case, corporate self-governance is sure to remain an area of contestation as notions of sovereignty and power intersect with the global use of proprietary inference models across governments, universities, hospitals and other civic institutions.
One final line of continuity concerns modes of capture. Analysing the relationship between terra nullius and historical colonialism, Buchan centres the role of subversive acts of ‘trafficking’, meaning the provision of gifts, favours and symbols of office that – as tokens of exchange – subtly coerced consent from Indigenous peoples that European colonists encountered to initiate trade relations. Over decades and centuries, these disingenuous relations reduced to modes of assimilation, subordination and then subjugation. While capture is already a concern in areas of AI critique, its relation to historiography and coloniality deserves particular attention.Footnote 86 Indigenous scholars have shown how historiographical claims naturalize dispossession. Jean M. O'Brien's Firsting and Lasting: Writing Indians out of Existence in New England, for instance, charts how colonizers narrativized Indigenous peoples as extinct in order to assert and mythologize their own sovereignty, ancestry and social order. Notably, Zoe Todd advances a similar critique of the history of science itself.Footnote 87
Conclusion
As with terra nullius, animo nullius threatens the commons with enclosure and private management. Both merit rebuke. In this paper, I have resituated histographies of AI in the history of corporate colonial expansion to account for the fields’ imbrications in systems of power that prefigure its emergence in the mid-twentieth century. I have shown how, as far back as the 1950s, corporate efforts to incorporate digital media into mainstream knowledge practices (and, in doing so, to sell standardized computing infrastructures, from IBM to Microsoft) have intertwined with positivist interpretations of mind-as-computer metaphors as both literal and perpetually looming. Even when these overlaps were the products of opportunism by one side or the other, privatization appears to be the outcome. This speaks to the political economy within which AI was formed – significant elements of which are usually neglected by practitioners, historians and media commentary, however critical, on AI. For historians, the decision to characterize ‘AI’ as primarily a science of the mind, rather than a science of operations and administration, is, in my view, just that – a decision, and one that neglects the breadth of AI's origins and helps blind us to its imbrications.
Acknowledgements
Special thanks to Richard Staley, Stephanie Dick, Mustafa Ali, Sarah Dillon, Helen Curry, Jon Agar and Simon Schaffer for their feedback on previous versions of this paper, and to Matthew Jones for similar guidance. Thank you also to Susie Gates and the many who brought such life to our Histories of AI Sawyer Seminar.
Open access
