1. The project that brought us together

As is the case for most research collaborations, the project that brought us together was formed not as an exercise in interdisciplinarity per se, but as an opportunity to deepen our respective engagements with a problem of shared interest. Our work together came about at a time characterized by the massive promotion of AI, not least in its application to the advancement of U.S. military supremacy. In the fall of 2023, David had just completed a PhD in Software Engineering at Carnegie Mellon University (CMU), which derives 42% of its research funding from the U.S. Department of Defense (DoD).Footnote ² He had drawn on Lucy’s work in an article building upon his PhD dissertation (Widder and Nafus Reference Widder and Nafus2023) to argue that the supply chain nature of AI systems contributes to the systematic abdication of responsibility by individual practitioners. David’s subsequent writing (Widder, West, and Whittaker Reference Widder, West and Whittaker2024), problematizing the trope of “open AI” and challenging claims that open-source software can “democratize” access to AI in the absence of more fundamental transformations, was another point of connection. These intersecting interests in AI critique comprised the starting place for our collaboration.

In 2023, David had obtained a substantial dataset comprising DoD grant solicitations issued from 2007 to 2023.Footnote ³ The question was whether and how this large document corpus might provide material relevant to an investigation of research and development in algorithmically enabled warfare, shedding further light on the relations that comprise the U.S. military-academic complex. Thinking about methods for engaging with the corpus, David contacted his friend Sireesh, then a third-year PhD student at CMU studying Natural Language Processing (NLP), whom he had met through graduate unionization efforts. Sireesh had previously taken a leading role in a qualitative study interviewing AI/NLP researchers on the incentives that shape their practice, aimed at generating a wider discussion of the field’s implicit norms in the service of more deliberately (re)directing its future (Gururaja et al Reference Gururaja, Bertsch, Na, Widder and Strubell2023). Sireesh was also, at the time, funded by a DoD grant himself and eager to deepen his understanding of the implications of that position. As well as becoming a co-author, Sireesh designed a customized search interface for the dataset tailored to our interests, enabling us to focus on documents within the corpus relevant to research in AI and Machine Learning. The project, in this sense, was at once a critical analysis of AI funding and an exercise in the development of NLP-enabled search to support that analysis.

The analysis that resulted was framed by a critique of the distinction between basic and applied research, showing how funding calls designated as basic research nonetheless enroll researchers into a warfighting agenda (Widder, Gururaja, and Suchman Reference Widder, Gururaja and Suchman2024).Footnote ⁴ A diachronic analysis of the corpus identified what we characterized as the “one small problem” caveat running throughout the grant solicitations, whereby affirmation of progress in military technologies is qualified by acknowledgement of outstanding problems, which becomes the justification for additional investments in research. A closer analysis of a subset of Defense Advanced Research Projects Agency (DARPA) calls for the use of AI in battlefield applications further demonstrated the two-way traffic between research communities (both academic and commercial) and the DoD. These showed more specifically how the DoD strategically mobilizes the category of “basic research” to reflect the interests and directions of the research community, which in turn inform the problem formulations and imagined solutions of DoD battlefield “applications.” Taken together, we argued that grant solicitations work as a vehicle for the mutual enlistment of DoD funding agencies and the academic AI research community in setting research directions. Bringing together the military trope of enlistment with the concept of enrollment as it has been developed within the field of STS (Callon, Reference Callon and Law1986), we see mutual enlistment as one among a set of processes through which problems are defined, solutions put forward, and actors recruited into positions and relations that work to sustain a particular definition of their collective situation (in this case, the necessity of AI for military supremacy, and of military supremacy as a basis for U.S. national security).

2. Agonistic engagements

Institutionally sanctioned spaces for interdisciplinarity are conventionally opened in response to a problem deemed to require multiple forms of expertise. Those spaces are also, and arguably more importantly, “a means of generating questions around which new forms of thought and experimental practice can coalesce” (Barry and Born, Reference Barry and Born2013: 10). In this respect, there is a difference that matters between normative research enlisted in the service of agendas where the problem framing is not itself open to question, and research that questions the frame.

The initial ground for our collaboration was a shared desire to trouble the unquestioned assumptions and the political and economic arrangements that frame the purposes and practices of the field of AI. Lucy’s engagement with AI began in the mid-1980s (a time of ascending investments in symbolic AI and knowledge representation) as a PhD student in anthropology who found a “field site” for her dissertation at Xerox’s Palo Alto Research Center (PARC). Engagement with her colleagues required the acquisition not of comparable technical competency (clearly beyond her reach), but rather the development of a sufficient level of conceptual understanding to be able to identify issues at the intersection of her interests and those of her colleagues and to make her reading of those issues intelligible to them. The first of her collaborations was a critical engagement with the development of an “expert system” based on the then-dominant planning model of AI. Differences between her own interactionist analysis of sociality (drawn from developments at the time in anthropology, ethnomethodology, and Conversation Analysis) and that of her colleagues provided fertile grounds for a critical reframing of the problem of human-machine communication (Suchman Reference Suchman1987, Reference Suchman2007).

Beginning his career at IBM Watson in 2015, Sireesh imagined his role as a developer of NLP systems in industry to be about making useful things.Footnote ⁵ What he encountered was a gap between the project of advancing NLP research and development on one hand, and the actual requirements of practical use on the other. He found himself building systems that were necessarily ad hoc and aimed at the automation of white-collar office work, an aim that aligned only loosely with the affordances of the NLP technology of the moment. Moving in 2018 to a self-identified AI startup, Sireesh was hired to develop NLP to support financial analysis. He soon discovered that the company was positioned more to be acquired than to produce working technologies: the purpose of the systems that he developed was to create impressive prototypes in which NLP was more of a promise than a fully implemented working system. At the end of his time there, he was team lead for machine learning operations, charged with engineering ongoing assessments of how models were performing, facilitating new data collection, and retraining models when required.

The divergence of new data from the data a model was trained on, however – both at IBM and at the startup – required labour-intensive re-annotation by people with relevant expertise. Overall, the mandate to deliver even partial workplace automation through AI leads, in Sireesh’s observation, to progressive immiseration in the workplace, as the work becomes annotating data for the machine rather than doing the job one is hired for. Billed as transformative, contemporary AI reinforces the promise to management of a comprehensive representation of business processes, independent of the actual labor involved. Traditional computer programming models process in code, depending on the user or operator to enact the parts of the process that cannot be encoded. Contemporary data-driven AI encodes process in weights, which expand the range of (especially imprecise and underspecified behavior) that can be modeled, while at the same time making understanding and editing of the modeled behavior much more difficult. In the end, the job modelled is a further devalued rendering of the job as it existed before, where the view of the work process that practitioners used to have has been encapsulated in a model that reflects none of the nuance with which they understand their practice, and whose logics are even more opaque. What is left to workers, then, is to correct the deficient model in the hopes of recovering the job that they performed, through the language that the model understands: annotated data.

While at the startup, Sireesh proposed an audit for gender bias in a speech-to-text algorithm trained on a severely skewed dataset of earnings call transcripts. These calls consist largely of audio from C-suite executives and analysts discussing public companies’ quarterly performance; as of 2023, an S&P Global report estimates that only 11.8% of C-suite positions in publicly traded U.S. companies are held by women (Chiang et al., Reference Chiang, Kaulapure and Sandberg2024). Sireesh’s manager made it clear that even if the audit were to uncover issues, they would not be acted upon unless they directly contributed to improving the model’s metrics on the same skewed dataset. Pursuing his concern with bias in NLP work, Sireesh decided to return to academia, still imagining his project as a technical one but with a labour-focused view aimed at giving agency back to users. Approaching the field of AI skeptically, with interests arising from his experience of failed projects in industry rather than from historically given problems in computer science, left Sireesh with the question: Where is the space to build software in thoughtful ways within the existing economic system? His move into the academy was motivated by the possibility that he might pursue this question as one integral to his profession, rather than being professionally complicit while engaging in critique on the side. In the academy again, however, he found that the concerns shaping NLP research and development are subject to industry pressures. The field is also narrowly circumscribed by technical mandates that marginalize the orientation to usefulness, while initiatives such as “NLP for good” and audits for bias operate as a separate and peripheral thread of NLP research.

Coming from a liberal arts background but with a lifelong love of computers, David undertook a PhD in software engineering (a field considered interdisciplinary within the rubric of academic computer science). During his studentship, he was hired at Microsoft Research (MSR) to conduct interviews examining why engineers were unable to use the responsible AI guidance provided to them in the form of a booklet of procedures. Expected to return results along the lines of “The font size is too small here,” David instead asked further questions more akin to, “Do you feel supported by your manager in acting on this guidance?” These questions were judged to introduce methodologically unacceptable “bias” beyond the set scope of the project, and he was asked to train further in how to conduct unbiased and ethical interviews. Rather than disciplining his qualitative research methods, however, David’s further reading led him to shift from viewing interviewing as a method to a deeper theoretical analysis of the interview as an encounter, and to a wider understanding of the differing epistemological stances that comprise qualitative research (Guba and Lincoln, Reference Guba and Lincoln1994). Like Sireesh, he experienced how, entrained in the prevailing instrumental view of interdisciplinarity in the service of the AI project, initiatives that question the fundamental premises or goals of the project are systematically sidelined. This problem is clear in the operationalization of ethics in AI in the form of a checklist or guideline that serves as a substitute for thought and critical engagement.

In contrast to the premise that there are given problems that require interdisciplinarity for their solution, in sum, our shared experience of interdisciplinarity is as a pathway to questioning the frame that has been handed to us in order to reformulate the problem. This is, necessarily, what Barry and Born (Reference Barry and Born2013: 12) identify as an “agonistic encounter,” in these cases disclosing the many ways in which computer science has encapsulated itself, only expanding its boundaries in instrumental ways. As a consequence, the aspects of the discipline that engage with the world outside the closed theoretical/technical domain of academic computer science (e.g., interface design, software engineering) are marginalized. While these subfields are necessarily more central to practices of industrial research and development, the latter reinforce the commitment to closure over problematization in the interest of the delivery of products. In the current moment of rapid expansion and commercialization of AI-enabled technologies, moreover, opening the field of AI requires radical modes of disciplinary reconfiguration that question not only received assumptions and methodological commitments, but also the political economies and professional subjectivities through which technical practitioners are disciplined.

3. Critical technical practice

In “Toward a Critical Technical Practice: Lessons Learned in Trying to Reform AI,” Philip Agre (Reference Agre, Bowker, Gasser, Star and Turner1997a) points to broader frictions between computer science and what that field considers to be its cognate disciplines. “In its representational aspirations,” he writes, “computing has been constituted as a kind of imperialism; it aims to reinvent virtually every other site of practice in its own image” (Agre, Reference Agre, Bowker, Gasser, Star and Turner1997a: 131). Agre contextualizes the field of AI within the broader cybernetic turn and the intensified entanglements of computer science with the military-industrial complex emerging from World War II and the Cold War. He identifies two related precepts that underwrite the AI project: that cognition comprises mental operations translatable as computation, and that those translations can be achieved through techniques of formalization. These precepts are sustained, he argues, rhetorically: “As a practical matter,” Agre observes, “the purpose of AI is to build computer systems whose operation can be narrated using intentional vocabulary” (Reference Agre, Bowker, Gasser, Star and Turner1997a: 139). The dominant narrative for that project in the 1980s was the so-called planning model, in which plans – understood as mental representations – provide the executable instructions for action.

As PhD students during the 1980s at Xerox PARC and MIT respectively, Lucy and Phil Agre met in the context of their shared critical engagement with the planning model. Conceptual and methodological resources from anthropology and ethnomethodology, combined with a technically informed understanding of the prospects and limits of AI planning, provided fertile ground for a critique of the representation of plans as mental constructs determining action. We argued instead for an articulation of plans as discursive and material artifacts, generated within and mobilized as resources for situated activity. Along with prior scholarship in the humanities and social sciences (e.g. Dreyfus, Reference Dreyfus1992; Forsythe, Reference Forsythe1993) our critiques drew from a small but growing body of work within computer science concerned with AI’s foundations and ramifications – not only as a matter of effects outside the discipline, but also for its constitutive technical practices (e.g. McDermott, Reference McDermott1976; Weizenbaum, Reference Weizenbaum1976).Footnote ⁶

In Computation and Human Experience (Reference Agre1997b), Agre formulates the concept of a critical technical practice as a stance open to its own reconfiguration. This requires, he argues, a rejection of the premise of disciplinary foundations as an underlying structure, in favour of attention to a discipline’s workings as a historically specific practice. Integral to generative disciplinary practice, on this premise, is continuing critical inquiry into that history. Rather than seeing the associated risk to disciplinary concepts and methods as a threat, he proposes that such risk comprises “the promise of better ways of doing things” (Reference Agre1997b: 23). Embracing this stance, Agre maintained his conviction that AI as a field could contribute to a deeper understanding of human experience, but only insofar as it developed a reflexive critique of the genealogy and consequences of its technical language. His own voracious and extensive reading beyond the discipline was, for him, a way “to extricate myself from the whole tacit system of intellectual procedures in which I had become enmeshed during my years as a student of computer science” (Reference Agre1997b: 148). Central to that process for Agre was a resistance to what he identified as the false precision of technical formalization (Reference Agre, Bowker, Gasser, Star and Turner1997a: 12).

Developments in the field of AI since Agre’s interventions are mixed. Consistent with his call is a growing body of critical analysis coming from the various disciplines that contribute to the current AI project, including computational linguistics/natural language processing and machine learning (e.g. Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Gururaja et al., Reference Gururaja, Bertsch, Na, Widder and Strubell2023), along with scholarship in the interdisciplinary fields of information studies and STS (e.g Burrell, Reference Burrell2016; Ribes et al., Reference Ribes, Hoffman, Slota and Bowker2019; Seaver, Reference Seaver2017). These are joined by activist collectives and progressive research institutes within commercial and academic computing.Footnote ⁷ These critiques question the validity of narrow benchmarks as measures of the quality of AI systems, insisting that adequate assessment of systems can only be done in the context of their applications and consequential effects. At the same time, the massive growth of (largely speculative) investment in AI has exploited the evocative vagueness of the language of the field to expand beyond academic confines, shape public opinion, and promote uncritical narratives of technological solutionism and inevitability. As the rhetorical and financial reach of those with vested interests expands, the work of critical resistance requires wider coalitions both within and beyond the academy.

4. Auspicious interdisciplinarity

For critical technical practitioners, moving outside the disciplinary boundaries of computer science affords intellectual resources for understanding how those boundaries are constituted and with what effects, and how they might be redrawn. While arguably interdisciplinary from its beginnings, at least in terms of those disciplines that make up the field of Cognitive Science (principally cognitive psychology, computer science and philosophy), the AI project has been implemented as and through the techniques and technologies of computer science and software engineering, and those technical fields retain their dominance. Recently, however, the boundary-crossing traffic has thickened as critical practitioners are moved to expand and deepen their understanding of the starting premises, limits, and consequences of their technical work, while those with primary affiliations to the humanities and social sciences expand and deepen their technical literacy. “Auspicious interdisciplinarity,” Barry and Born propose, “is associated not only with the constitution of new objects, but with the cultivation of interdisciplinary subjectivities and skills” (Reference Barry and Born2013: 39).

Early critiques of the disciplinary limits of AI, then a somewhat esoteric project, focused on the question of whether and how computational models of mind, action, and interaction could do justice to the lived, organically embodied, culturally embedded, and historically evaluated relations that comprise intelligence. Now, more than ever, those disciplinary critiques require a radically extended frame—one that includes the political economies of what Hao (Reference Hao2025) has identified as the “Empire of AI” (see also Crawford, Reference Crawford2021). That reframing insists on attention to the conditions of possibility for the scaling of AI, from resource extraction to massive new investments in energy production, and from exploited labour to concentrations of wealth and associated power. In this moment of frantic investment and expanding colonization by governments and interested corporations, more and more of us are enrolled as (often unwitting) contributors to the AI project through the capture of our data traces, as outsourced labour, or as ingrained consumers. Adequate critique in the present moment calls for a critical technical practice committed to asking fundamental questions, resisting the posited inevitability of AI, and generating more radical alternatives that reject technosolutionism in favor of collective movements to build just and sustainable futures. Whether and how algorithmically based technologies could enable those futures can only be determined through continually unfolding experiments in auspicious interdisciplinarity.