Like the polar bear beleaguered by global warming, artificial intelligence (AI) serves as the charismatic megafauna of an entangled set of local and global histories of science, technology and economics. This Themes issue develops a new perspective on AI that moves beyond conventional origin myths – AI was invented at Dartmouth in the summer of 1956, or by Alan Turing in 1950 – and reframes contemporary critique by establishing plural genealogies that situate AI within deeper histories and broader geographies. ChatGPT and art produced by AI are described as generative but are better understood as forms of pastiche based upon the use of existing infrastructures, often in ways that reflect stereotypes. The power of these tools is predicated on the fact that the Internet was first imagined and framed as a ‘commons’ when actually it has created a stockpile for centralized control over (or the extraction and exploitation of) recursive, iterative and creative work. As with most computer technologies, the ‘freedom’ and ‘flexibility’ that these tools promise also depends on a loss of agency, control and freedom for many, in this case the artists, writers and researchers who have made their work accessible in this way. Thus, rather than fixate on the latest promissory technology or focus on a relatively small set of elite academic pursuits born out of a marriage between logic, statistics and modern digital computing, we explore AI as a diffuse set of technologies and systems of epistemic and political power that participate in broader historical trajectories than are traditionally offered, expanding the scope of what ‘history of AI’ is a history of.

This paper traces elements of the theoretical origins of artificial intelligence to capitalism, not neurophysiology. It considers efforts in the twentieth and twenty-first centuries to formalize a science of mental behaviour using the dynamics of social rather than neural phenomena. I first revisit early American theorists’ controversial ambivalence toward neurophysiology, showing how this group benefited from post-war corporate and military investments in commercial and imperial expansion, which sustained and expanded their influence over the emerging field. I then trace the lasting effect of the founders’ early rhetoric through AI's institutionalization after 1960, arguing that from the 2010s technology corporations set out to veil their enclosure of the data commons via appeal to a curious precedent: the scientific pedigree of AI. By relating the field to the history of capitalism, and specifically the rise of assetization in modern technoscience, I invite reflection on AI's origin story and on broader parallels between historical colonialism and data colonialism. I offer a heuristic – animo nullius, for ‘no persons’ mind’ – as an attempt to name rhetorical manoeuvres that leverage the authority of mind-as-computer metaphors in order to naturalize acts of seizure.

Drawing on prior work in the history and philosophy of statistics, I argue that in many cases analyses powered by artificial-intelligence (AI) techniques such as machine learning (ML) are fundamentally ‘conjectural’: reliant on ex post facto abductive logics often misinterpreted in contemporary machine-learning systems as reliably reproducible truth. Here I relate what Carlo Ginzburg calls ‘the conjectural sciences’ as a historical category to their contemporary instantiation in machine learning and the practice of ‘automated conjecture’. I observe how the automation of physiognomic and phrenological concepts are exemplary of the ways in which discredited conjectural pseudosciences are being revived by today's AI research. Finally, I argue that the conceptual distinction between ‘conjectural’ and ‘empirical’ science can help support contemporary efforts to regulate the design and use of AI systems by providing conceptual and historical justification for the non-development of certain classes of systems intended to automate inference.

Myth, hype and industry-captured historiographies of AI, machine learning and computing depict the current moment as a unique and unprecedented confrontation with computational power, paying particular attention to the devastating effects this has on vulnerable communities (better: communities made vulnerable). But automated decision schemes are neither new nor newly urgent; they are the inheritance of almost seventy years of computing history, a history that has never not been entangled with state repression, genocidal and ecocidal violence and racialized expropriation. Carceralilty and artificial intelligence have a shared history, and the entanglements between them remain underappreciated by history writing on computing – despite the near-universal injunction to leverage history as a means to critique and counteract the cultural hegemony of computing. This paper examines some of the tensions underlying such calls for historical critique, calling into question the efficacy of such a project. I offer a study of the complex senses of history and development at work in the 2006 International Congress of Mathematics, discuss recent historiographies of computing (and science) from guild historians, and describe the ways academic history writing reproduces the same relations of dominance and campaigns of creation and conservation that scaffold mathematics and artificial intelligence.

Little historical work examines the problems, practices and values of ‘machine learning’ as it was understood and justified within pattern recognition research communities prior to the 1980s. This omission has led to a failure to appreciate how the efficacy of machine learning was often justified by its perceived capacity for ‘originality’ rooted in machine (and human) subjectivity. This paper examines why and how early 1950s pattern recognition researchers came to see ‘machine learning’ as a technical and epistemological set of nominalist strategies for performing induction and abduction given incomplete, complex or contradictory information that might also spur ‘creative’ insight in such diverse activities as political judgement and scientific inquiry. I document local research problems, epistemological commitments, institutional contexts and the circulation of ‘machine-learning’ practices and values through three cases of early-career researchers imagining, building and programming digital computers to ‘learn’ from 1950 to 1953. This machine learning implemented in learning programs came to be seen by some researchers as more efficacious descriptions of the natural and social world because these descriptions were perspective-dependent, profoundly contingent and contextually non-exhaustive. Often indexed as colloquial appeals to greater ‘generality’, conceptions of machine learning's efficacy as the capacity to make meaning from contradictory information continues to inform contemporary debates regarding artificial intelligence, society and possibility.

In the 1960s, creativity became an important category for the Soviet state. Soviet educators and policy makers came to define creativity as problem solving in the service of Soviet automation. At the same time, the introduction of cybernetics, information theory and methods of artificial intelligence (AI) to psychology enabled Soviet researchers to perform quantitative studies of human cognition. The state concern with creative thinking and the cyberneticization of Soviet psychology allowed for the first quantitative studies of human problem solving. These shifts in Soviet society and scientific communities created fertile ground for the creation of Lev Landa's algo-heuristic theory (AHT), a pedagogical method of cultivating rule-bound creativity relying on tools and instruments developed and perfected in information theory and AI research. Drawing on scholarship in the history of algorithmic rationality, the Cold War discourse on creativity as a corporate imperative, and the place of cybernetics-inflected methods in the welfare domain, this article analyses the AHT as rule-based instrument of making creative thinking accessible to the lay mind.

This article characterizes early research in the field of ‘human–computer interaction’ (HCI) by analysing the first decade of ‘user psychology’ research at Xerox's Palo Alto Research Center (PARC). PARC's Applied Information-Processing Psychology Project (AIP) provided an initial theoretical foundation for HCI in the early 1980s. Like researchers in artificial intelligence (AI), researchers at AIP drew from information-processing psychology. However, AIP researchers argued that their focus on human behaviour distinguished their research from AI and other fields allied with computer science. Previous scholarship has shown that United States computer engineers became concerned with ‘users’ as they sought to commercialize military-funded developments in interactive computing. This paper argues that the decision made by upper management in computerizing workplaces to shift some text production work from clerical workers to middle managers during the 1970s and 1980s led AIP to perceive ambiguities around gender and technical skill. This shaped the initial theoretical foundations that the research group offered to HCI – especially the group's conception of the ‘user’. Computer designers went from presenting word-processing programs as clerical machines for women workers to presenting them as tools for masculine thinking. AIP's research diverged from industrial engineering and AI in response to this transformation.

This article analyses the intellectual and institutional development of the artificial-intelligence (AI) research programme within the Soviet Academy of Sciences from the 1970s to the 1980s. Considering the places and ideas from which it borrowed, I contextualize its goals and projects as part of a larger technoscientific movement aimed at rationalizing Soviet governance, and unpack shared epistemological and cultural assumptions. By tracing their origins to debates accompanying the introduction of cybernetics into Soviet intellectual and political life in the 1950s and early 1960s, I show how Soviet conceptions of ‘thinking machines’ interacted with dialectical materialism and communist socio-technical imaginaries of governance and control. The programme of ‘situational management’ developed by Dmitry Pospelov helps explain the resulting conception of AI as control systems aimed at solving complex tasks that cannot be fully formalized and therefore require new modelling methods to represent real-world situations. This specific orientation can be understood, on the one hand, as a research programme competing with systems analysis and economic cybernetics to rationalize Soviet management, and, on the other hand, as a field trying to demarcate itself from a purely statistical or mathematical approach to modelling cognitive processes.

This article traces the historical emergence of a new understanding of radiologists as fallible expert observers from the late 1940s, a conception that was shaped by new technologies and techniques, but also prepared the ground for promises of automation and artificial intelligence in the field of medical imaging. Reports of radiologists’ unreliable performance prompted investigations in many countries into ‘observer variability’ and ‘observer error’. Towards the end of the 1950s, scientists could conceive of radiologists as imperfect medical decision makers, while they concurrently developed a new model for ‘logical analysis’ of the diagnostic process that would limit errors. As well as technological solutions to flawed X-ray readers, researchers proposed ‘double-reading’ practices (a second independent reading) as a way to mitigate the ‘human factor’. Yet these ideas did not find widespread resonance due to concerns about feasibility and debates about radiological expertise, and also because of a discrepancy between experimental models and real-world practices. A genealogy of the fallible trained observer helps us understand persistent worries about – and solutions to – radiologists’ ‘error problem’ and contributes to a better understanding of current discourses on AI in medical imaging.

Between 1986 and 1996 researchers at the AT&T Bell Laboratories Adaptive Systems Research Department curated thousands of images of handwritten digits assembled by the United States Postal Service to train and evaluate artificial neural networks. In academic papers and conference literature, and in conversations with the press, Bell Labs researchers, executives and company spokespeople deployed the language of neurophysiology to position the systems as capable of codifying and reproducing feats of perception. Interpretations such as these were pivotal to the formation of brain–computer imaginaries that surrounded the development of the systems, which obscured the institutional infrastructures, clerical and cognitive labour, and the manipulation and maintenance of data on which feats of ‘recognition’ depended. Central to building the group's networks was the development of data sets constructed in consort with the US Postal Service, which arbitrated between the practicality of conducting research and the representation of an extensive catalogue of possible forms and permutations of handwritten digits. These imaginaries, which stressed a likeness with the human brain, were compounded by the promotion of ‘successful applications’ that took place under the AT&T corporate umbrella and with the essential support of US Postal Service workers to correct system errors.

A canonical genealogy of artificial intelligence must include technologies and data being built with, for and from animals. Animal identification using forms of electronic monitoring and digital management began in the 1970s. Early data innovations comprised RFID tags and transponders that were followed by digital imaging and computer vision. Initially applied in the 1980s for agribusiness to identify meat products and to classify biosecurity data for animal health, yet computer vision is interlaced in subtler ways with commercial pattern recognition systems to monitor and track people in public spaces. As such this paper explores a set of managerial projects in Australian agriculture connected to computer vision and machine learning tools that contribute to dual-use. Herein, ‘the cattle crush’ is positioned as a pivotal space for animal bodies to be interrogated by AI imaging, digitization and data transformation with forms of computational and statistical analysis. By disentangling the kludge of numbering, imaging and classifying within precision agriculture the paper highlights a computational transference of techniques between species, institutional settings and domains that is relevant to regulatory considerations for AI development. The paper posits how a significant sector of data innovation – concerning uses on animals – may tend to evade some level of regulatory and ethical scrutiny afforded to human spaces and settings, and as such afford optimisation of these systems beyond our recognition.

Using images from large-scale vision datasets (LSVDs), five practice-based studies – experimentations – were carried out to shed light on the visual content, replications of historical continuities, and precarious human labour behind computer vision. First, I focus my analysis on the dominant ideologies coming from a colonial mindset and modern taxonomy present in the visual content of the images. Then, in an exchange with microworkers, I highlight the decontextualized practices that these images undergo during their tagging and/or description, so that they become data for machine learning. Finally, using as reference two counterhegemonic initiatives from Latin America in the 1960s, I present a pedagogical experience constituting a dataset for computer vision based on works of art at a historical museum. The results offered by these experimentations serve to help speculate on more radical ways of seeing the world through machines.

In 1972, ten members of the machine intelligence research community travelled to Lake Como, Italy, for a conference on the ‘social implications of machine intelligence research’. This paper explores their varied and contradictory approaches to this topic. Researchers, including John McCarthy, Donald Michie and Richard Gregory, raised ‘ethical’ questions surrounding their research and actively predicted risks of machine intelligence. At the same time, they delayed any action to mitigate these risks to an uncertain future where technical capabilities were greater. I argue that conference participants’ claims that 1972 was ‘too early’ to speculate on societal impacts of their research were disingenuous, motivated both by threats to funding and by researchers’ own politically informed speculation on the future.

This paper considers three moments in the treatment of data about race and identity in India. Many elements go into the development of data imaginaries as these change over time. A complete history is beyond the scope of this paper, but I develop three key episodes to explore critical but changing features of interrelations between race, identity and statistical arguments historically. One aim is to explore key features of the argument developed by two significant individuals – Thomas Nelson Annadale and P.C. Mahalanobis – as they sought to develop databases that could answer questions about race formation and, in the case of Mahalanobis, might also be used to develop statistical methods on the one hand and aid governance on the other hand. A second aim is to use this historically based but highly selective investigation to uncover key features of the ideology with which the government of India has presented Aadhaar, its vast biometric identification system powered by authentication technologies afforded by artificial intelligence. This enables me to identify different forms of racial or ethnic identity that could be – and in one or two cases actually have been – implicated in the way Aadhaar has been used in practice.

In April 2018, as the Argentine Congress debated decriminalizing abortion, local media revealed that public-health officials in a northern province had deployed an algorithmic system to predict teenage pregnancy. Public response to the technology quickly became entangled in society-wide debates about reproductive rights. Both proponents and detractors of the algorithmic system framed the technology as novel and cutting-edge. However, this paper argues for an analysis of the system not as a form of innovation or rupture but as a continuation of historical forms of biopolitical governance in Argentina, particularly puericultura, a eugenic theory of child rearing.