Basic research,lethal effects: Military artificial intelligence research funding as enlistment

Context for the basic/applied distinction

The Congressional Research Service shows how DoD science and technology funding is broken into eight “budget activity” categories (Sargent, Jr, 2022). Categories 6.1 and 6.2 are “Basic” and “Applied” research, respectively. Numbers from 2022 show that DoD funding makes up 41.2% of the entire federal budget for R&D, and within this DoD portion, basic and applied research together make up less than 10% of the expenditure, with the lion's share going to “operational systems development” and “advanced component development and prototypes” (Sargent, Jr, 2022). And yet within the overall small fraction, “basic” research is funded at about one-third the level of “applied” research. Even if basic research is to be considered in some way morally exempt, in other words, it is at best a tiny adornment on a much larger arsenal of R&D funding directed more explicitly toward military ends.

University policies attempt to support free exchange of ideas while accepting military funding. As one example, Carnegie Mellon University (CMU) derives 42% of its research funding from the DoD (Carnegie Mellon University, 2023), accruing the third most DoD funding of any university (Folts, 2024). The university has a policy that by default, this funding must be “unrestricted” in terms of publication, though it allows explicitly restricted and military-funded research to occur in its “nonacademic” “semi-autonomous units,” and even in academic units by exception (Carnegie Mellon University, 1988). Debate within CMU's School of Computer Science about the ethics of such funding affirms an apparently morally relevant distinction between “weapons” and “non-weapons R&D,” but capaciously defines the latter to include the development of the autonomous tactical vehicle named “Crusher,” because no literal weapons were built onto it by the university (Carnegie Mellon University, n.d.). CMU also hosted a listening session for the DoD's Defense Innovation Board, which affirmed the DoD's goal to use AI to “fight and win in future wars” (Defense Innovation Board, 2019). Recommendations from the Defense Innovation Board included strengthening private sector and university connections, held up top tech companies and Elon Musk (by name) as inspiration, and recommended a “focus on exploratory research.” In short, while basic/applied distinctions are seen as core to the university's mission and to the maintenance of academic ethics, our dataset allows us to examine how the basic/applied distinction works in DoD funding calls and the questions that this raises.

How basic research is directed to military ends

Even when a program is framed as funding for basic research, there are indications that positioning a proposal in relation to DoD objectives will increase its chances of success. For example, the 2024 Vannevar Bush Faculty Fellowship call (#1) ⁷ says, in bold text, that the fellowship is intended for “ambitious ‘blue sky’ research,” and opens by quoting the DoD Definition of “basic research,” as “systematic study directed toward greater knowledge,” “without specific applications […] or products in mind.” This is immediately followed, however, by the qualification that such work must be “related to long-term national security needs.” The stated objectives of the program include “future revolutionary new capabilities for DoD,” as well as training students “for the defense workforce” to enable “long-term relationships between university researchers and the DoD.”

Basic research is also often discussed with explicit reference to battlefield and military applications. In the same Vannevar Bush fellowship cited above, scientific areas set out in the solicitation include numerous instances of the language of warfighting, for example “Fundamental research in neural activity […] can also lead to the development of brain-machine interfaces (BMIs) to facilitate the integration of the warfighters and future AI agents.” (p.12) In another example, a funding call from the Army Research Office, for its Historically Black Colleges and Universities program, calls for research on Lidar Remote Sensing, which the call says “could be used” for “remote sensing over the battlefield, e.g. in support of field artillery targeting” (#3). Finally, another “basic research” solicitation promises funding for the so-called “Internet of Battlefield Things,” to enable “dynamics that exceed human op-tempos” the “survivability of mission assets” and handling the exigencies of an “IoBT-enabled battlespace” (#4). This solicitation notes that while the “core” research is to be funded as basic research, there is also a path to funding for applied research as an “Enhanced Program.” In sum, grant solicitations within the corpus framed as funding for basic research at the same time steer research directions toward military applications, while building relations between university-based researchers and the DoD.

How the basic/applied distinction is obviated

Finally, there are many programs funded as either Basic or Applied research, with little apparent distinction made between the two. A Navy funding call states that “ONR is interested in receiving white papers and proposals in support of advancing AI for future naval applications. Work under this program will consist of basic and applied research…,” including “AI enabled training environments that test warfighter skill” (#5). Many other solicitations invite proposals addressed explicitly to military applications; for example, a funding call for “Air Delivered Effects” from the Air Force Research Laboratory's Munitions Directorate includes a research funding area on “Autonomous Target Recognition.” Subgoals include “Approaches for training AI/ML or traditional algorithms with synthetic target data that result in good target recognition performance when using real target data,” and “autonomous handoff of target cue information from intelligence, surveillance, and reconnaissance (ISR) or fire control sensors to weapon[s]” (#6). The Army Corps of Engineers asks for better “sensor payloads” for “drone swarms” to “monitor human […movement…]” (#7).

We look in more depth at calls for research directed toward military applications below, in our examination of DARPA calls for research in AI/ML. Here the point is to underscore the extent to which DoD grant solicitations, including those framed as basic research, are saturated with reminders of the department's aim of expanding the warfighting capacities of the U.S. military. Research characterized within university laboratories as basic, relevant to a wide and possibly indeterminate range of applications, is effectively steered toward military ends through indications that proposals will be ranked in terms of their relevance to DoD priorities. At the same time, DoD investments extend beyond funding for particular research projects, to building out and securing the network of relations, and financial dependencies, that hold the military-academic complex in place (Giroux, 2007). In the following section, we look at a series of grant solicitations over time, to trace other devices through which cycles of problem formulation and promised technological solution are sustained.

From problem to solution, or solution to problem?

Our dataset spans 16 years, allowing us to track trends in AI research topics across this period. To ground our analysis, we examine grant solicitations in three programs, each with a different disciplinary focus, all explicitly categorized as “basic research.” Firstly, the Multidisciplinary University Research Initiative (MURI) solicits research yearly on “high priority topics and opportunities," ⁸ primarily in STEM fields. Secondly, solicitations for the Vannevar Bush Faculty Fellowship (VBFF) fund faculty fellows each year, typically on the same topics year-to-year ranging from AI to materials science to quantum computing (though with some major changes discussed below). Finally, the Minerva Research Initiative is devoted to “supporting social science for a safer world,” focusing on topics of “strategic importance to the U.S. national security policy.” ⁹ The fact that these three programs do not focus solely on AI or even computer science, but instead reflect broad areas of military research interest, allows us to examine AI's relative importance among other fields of study, and its impact as a method within other fields and relative to outside trends in academic and industrial AI research. ¹⁰ Notably, we see how, especially after 2021, AI increasingly becomes entangled with other fields, where it is seen either as a necessary part of those areas of research, or even their ultimate goal.

One small problem

Solicitations that outline an aspirational capability employ a recurring rhetorical move characteristic of the genre more widely, which we name the “one small problem” caveat. Here we trace how affirmation of the promise of an existing approach is followed by assertions of its limited scope, and the need for research that would allow the approach to work under more “realistic” conditions. For example, when discussing approaches based on game theory, solicitations assert that it models how “emotionless geniuses” behave, rather than “average people with emotions” (MURI 2011, #14), and assumes a “stable world/environment,” which “may not be guaranteed” (MURI 2015, #15). Similarly, reinforcement learning, despite its “dramatic success” at games like chess or Go, is of uncertain applicability to “problems of practical interest” (MURI 2021, #16). This framing balances two competing incentives: emphasizing certain methods’ promise and past successes and thus presenting them as viable directions for future work, while at the same time directing that future work toward military contexts in which outstanding problems require solutions.

Over time in these series, the “one small problem” caveat increasingly narrows its focus to machine learning approaches within AI. In 2011, MURI topic 22 solicits work on game theory with the open-ended goal to “handle imperfect information” by “modeling deception” (#14). By contrast, later solicitations assume much more specific approaches in vogue in the wider research landscape at the time: one solicitation in 2017 refers to “semantic energy,” as well as “deep learning and ensemble modeling approaches” for characterizing graphs (#11); another from 2024 focuses entirely on generative adversarial networks (#17), with each technique appearing in solicitations shortly after their prominence in academic discourse. We see residual effects in the social science-focused Minerva grants as well: in successive revisions of their 2013 solicitation (#20), “belief formation” changes to “belief propagation,” the name of an algorithm used in the then-popular kind of graphical or Bayesian models. Through these framings, outstanding problems in the application of computational methods to DoD interests can be formulated in terms that affirm the value of previous investments while setting out new directions for further research, updating approaches to problems in ways that align with the changing interests of the research community.

AI as a durable frame

This narrowing of focus to specific techniques reflects the increasing prominence of AI, as even the conception of problems is set out in the terms, methods, and affordances of AI and machine learning. AI then becomes a durable frame: a way of conceptualizing problems over the span of multiple years, and a precedent that any new approach should reference.

In the MURI series, the most notable early example of a durable frame is control theory, prominent in the late 2000s, which faded as big data and network science took over around 2012. Network science remains a durable frame in the VBFF dataset as AI emerges: the 2021 rewrite of topics in the VBFF series (#19) elevates “Networks and Artificial Intelligence” to the first topic of concern, and AI is discussed as a field that can be revolutionized using network science. In the Minerva series, quantitative methods can be seen as a durable frame throughout the series, as solicitations for better quantitative methods occur at the top of the list of topics in many years. Quantitative methods are held up as being “more consistent and reproducible” than qualitative methods, offering “well-defined comparisons” (2013, #20). This holds even as the authors of later Minerva solicitations acknowledge that “quantitative” does not mean “objective,” that changes in quantification and metrics result in different insights, and that quantitative work may not be “grounded and ethnographically representative” (2022 #21).

By the end of the series, AI is the most prominent durable frame in this dataset. In the 2024 VBFF solicitations (#1), we see invocations of AI in multiple focus areas—applied mathematics, neuroscience, and materials science—which all mention AI's importance in their fields. In the neuroscience topic, we see a shift in the focus of solicited research over time toward AI. In 2016, the first year for which we have descriptions of the VBFF topics, the neuroscience area focuses on supporting “cognitive effectiveness and emotional resilience” for warfighters in the face of sleep deprivation and the “growing volume of data presented by today's weapon systems,” along with the development of BMIs to assist with recovery from traumatic brain injuries (#22). In subsequent years, however, these focuses shift—while the desire to support warfighters with information overload remains, the rest of the topic is dedicated to prospective applications for AI: research into the brain and its function are now for inspiring new concepts in AI, and BMIs are now meant to integrate warfighters with AI systems, rather than support their recovery from injury.

Inevitability amid ambivalence

Durable frames, in their conception, require successes to point to in order to justify their continued funding and interest. The successes attributed to AI in the series often rely on discourse familiar in more consumer-facing applications, including the inevitability of the revolution that it brings. AI's inevitability, however, is based in claims of its progress that are, at best, only vaguely connected to the solicitation's intended field of application. One MURI topic in 2020 (#23), “Adaptive and Adversarial Machine Learning,” rests its assertions that machine learning “improves overall effectiveness” on the “success of autonomous cars on US roads,” despite the fact that no autonomous cars in 2020 had moved beyond trial programs in narrowly defined environments (Piper, 2020). Further, as the solicitation notes, deep learning can be “easily defeated by minute perturbations” in operating conditions. Similarly, another MURI topic in the following year (#16) states boldly that “in 10 years, many cyber tasks will be carried out by autonomous systems,” without citing any evidence or source for this prediction, while noting that deep learning provides few behavioral guarantees, and key technical assumptions are “commonly […] violated in the real world.” In these cases, the authors of the solicitations appear firmly convinced that AI and deep learning are inevitable, despite their own articulation of the (potentially fundamental) flaws with these approaches.

The juxtaposition of inevitability with the acknowledgement of outstanding problems represents a significant and recent rehabilitation of AI's image in the military imagination. Until about 2020, the technology undergirding the modern conception of AI, deep learning, was considered wholly unsuitable for military problems. Solicitations discussing deep learning in 2017 and 2018, rather than posit it as promising using “the one small problem” caveat, instead point at its “crippling fragility,” and imply that deep learning with its large data requirements is not “physically viable” for controlling military systems (MURI 2018, #24). Several solicitations go so far as to forbid deep learning approaches for successful submissions. Multiple Minerva solicitations also express an implicit uncertainty about the efficacy of AI solutions, while nonetheless affirming their inevitability in statements like: “regardless of whether possessing AI capabilities equates with power, the world finds itself in a race to develop and deploy these technologies” (2020, #25).

In its growing prominence in the framing of DoD grant solicitations, AI demonstrates a departure from previous forms. Despite the lack of clear (public) successes in the military domain, it is still increasingly seen as inevitable by multiple communities, who at the same time demonstrate their awareness of the fundamental, deeply rooted pitfalls in adopting it: the lack of behavioral guarantees and “crippling fragility” of deep learning systems, the potential for ungrounded quantitative analyses that bear the sheen of objectivity, and the massive costs of data and computational resources. These problems become acutely consequential when we move from domains of primarily laboratory-based research to the arena of warfighting.

Situational awareness

Continually confounded by the “fog of war” (Shapiro, 2005), SA is a foundational concern of military doctrine. SA is defined by Micah Endsley, former Chief Scientist of the United States Air Force, as “the perception of environmental elements and events with respect to time or space, the comprehension of their meaning, and the projection of their future states” (1988; for a critique see Suchman, 2023). Military doctrine takes information gathering, formalized as ISR, as key to enabling SA through network infrastructures and targeted surveillance. Consistent with the automation of computation in the mid-twentieth century, these capabilities are conceptualized as recursive procedures that follow the hierarchical ordering of command and control. This runs from the formulation of strategy by senior officers, to the tactical operations of front-line commanders, through to the perceptions and actions of individual warfighters. In the contemporary moment, our corpus shows how optimization of SA is to be achieved through increasing reliance on the application of AI (principally ML) in ISR.

In March 2010, DARPA issues a BAA for the “Mind's Eye Program,” with the aspiration to develop machine capabilities in “visual intelligence” (#26). The announcement opens with a review of the limits of the current state of computer vision, in its reliance on statistical computations based on benchmark datasets rather than models of perception and cognition in realistic environments. Humanlike visual intelligence is characterized as the capacity to represent observed events in “the mind's eye,” in order “to manipulate those imagined scenes mentally to solve problems.” The (re)turn to mental modeling indexes the “one small problem” of the limits of ML-based techniques for computer vision, and accompanies a shift in focus from object to action recognition. In this case, advances in computer vision are aimed at the creation of a “smart camera,” to be deployed as a “payload” on Unmanned Ground Vehicles (and potentially other platforms), with sufficient “machine-based visual intelligence” that it can report on activity in an area of observation. The aim of the research program is to develop the “perceptual and cognitive software” that would equip the camera with “the capabilities necessary to perform surveillance in operational missions.” The aspiration, in other words, is to encode the specifications for SA into the camera itself, a move that requires the translation of detectable signals into warfighting's normative frames.

The BAA goes on to list the requisite capabilities in terms that anticipate their translation to code, beginning with the “recognition of primitive actions that take place between objects in the visual input, with a particular emphasis on actions that are relevant in typical operational scenarios (e.g. vehicle APPROACHES checkpoint; person EXITS building).” The system should learn “generally-applicable, invariant spatiotemporal patterns that make up these actions directly from visual inputs based on few labeled examples and with minimal supervised learning,” and be able to apply the patterns learned in one scene to another. Those capacities should be iteratively refined through continued exposure to visual inputs during the camera's operation, such that the system “develops a sense for what is ordinary and out-of-the-ordinary through continued exposure to a particular visual scene.” In addition to these imagined ML-based capabilities, the system should incorporate “reasoning” functions that integrate “stored and acquired knowledge of visual objects as well as symbolic knowledge” to enable it to issue “alerts to activities of interest, as defined by concepts the device is tasked to detect.” Examples of such concepts include the introduction of an object into the scene, or the handoff of an object from one person to another. The system should be “taskable” through simple text commands and provide alerts via text messages, as well as able to “explain its reasoning by displaying relevant video segments for what has been observed.” Finally, and most ambitiously, the system should perform interpolation (inferring actions that are not directly observed, e.g. an object carried into a building that is absent on the person's exit from the building must have been left inside), and prediction (projecting next actions based on previously observable cases, e.g. a boy carrying an object and approaching a soldier may be expected to hand the object over) (see Figure 1). ¹¹

OPEN IN VIEWER

Figure 1.

A diagram from a grant proposal demonstrating the prediction that a boy approaching a soldier may hand him an object, based on previously observed cases.

The aspiration to advance from object to activity detection begs the question of to what extent even object recognition in the context of warfighting is a solved problem, particularly when the objects in question are not specifically military (e.g. armored tanks) but ordinary civilian objects (e.g. in this case a box). But we can see here a desire to expand the range of algorithmic processing to encompass surveillance operations aimed at the detection of anomalies in so-called “patterns of life.” ¹² In its explicit mandate for the inclusion of “reasoning” functions, the call anticipates a niche but persistent thread within the AI R&D community away from a reliance exclusively on statistically based ML toward a “hybrid” approach that restores techniques of knowledge representation and symbolic logic marginalized with the rise of data-driven approaches (Marcus, 2020).

Over a decade later, the vision of algorithmic analysis of ISR data as a hybrid of ML and symbolic reasoning is reiterated in a September 2022 BAA from DARPA's Information Innovation Office (I2O) (#27). The call for “Environment-driven conceptual learning (ECOLE)” sets out an ambitious program for SA, object recognition, activity-based intelligence, and “pattern of life”analysis. Now, however, the human analyst is restored to the frame. The program aims to “enable human-machine collaborative analysis of image, video, and multimedia documents during time-sensitive, mission-critical DoD analytic tasks [so that the machine can] readily learn a new symbolic representation through interaction with its human partner.” The move to “human-machine collaborative analysis” could be read as well as an acknowledgement of the limits of the hybrid approach imagined in 2010, as the 2022 BAA calls for “an extensible framework capable of instantiating an arbitrarily large (i.e. global-scale) web of knowledge representations.” The hope for a comprehensive encyclopedia of relevant information is an aspiration that dates back to the mid 1980s and the Cyc project, a four decade-long effort to codify the “common-sense knowledge” assumed to be the foundation of human reasoning (Fisher, 2024). Here, the call attributes the “small problem” of training in ML and the continued need for human feedback to the limits of knowledge representation (a problem that data-driven statistical approaches promised to resolve). But while initially imagined as a replacement for human experts, knowledge representation is reconceived here as a form of human-assisted machine learning, which brings us to our second theme.

Human–machine teaming

Along with the aim of greater machine autonomy, contemporary investment in AI is justified within the R&D community as a step toward “harmonious partnership between humans and AI.” ¹³ Embraced by the DoD, the aspiration to what is named “human-machine teaming” is a continuation of a cybernetic imaginary that positions the human as the “machine in the middle” of complex military systems. ¹⁴

In March of 2019, a BAA from DARPA's Defense Sciences Office calls for proposals on the “Science of Artificial Intelligence and Learning for Open-world Novelty (SAIL-ON).” (#28) The problem to be solved is formulated as the limited progress of AI systems in “understanding the most important component of the environments in which they operate: humans.” The BAA returns to a paradigm of mental models as the basis for successful interaction. ¹⁵ Among other deliverables, the BAA asks for simulations of “agents” that demonstrate the requisite social skills, implemented in a virtual testbed with input from standardized “sensing channels” and output through “communication/action channels” made available to human teammates through standardized interfaces. The testbed should demonstrate agents’ capacities to “operate in increasingly complex and specialized environments, be adaptable to sudden perturbations in the mission or team (like the loss of communication with a key teammate), and use noisy multi-channel observations to represent the world and do complex inference and prediction.” By confining the demonstration to a simulation, rendering contingency as complexity, specialization, perturbations, and noise (all terms with familiar computational translations), and defining social skills as the alignment of mental models, the call for “open-world novelty” remains within the arguably closed-world of information processing psychology.

Modeling of mental states is expanded in a November 2020 DARPA call for Post Doctoral Fellowships in the domain of Emotion Recognition, based on analysis of “ego-centric video captured from head-mounted cameras”(#29). The premise is that “estimating and modeling user emotions is important for effective human-computer teaming.” More specifically, the aspirational capability involves “new techniques to determine the emotional state of a user of a head-mounted camera based on the ego-centric video, audio, and inertial data it collects […] The goal of this project is to develop new techniques that integrate any and all cues to a user's emotional state into a continuous, real-time estimate of user affect.” Applicants are reminded that this “ego-centric video data” does not include the user's face, excluding the usual source data for so-called “emotion recognition.” The research challenge is to move beyond affect or sentiment analysis based on facial images to inferences from first-person perspective images of a scene, recorded speech, and head motion, gesture, and body position, as well as “how a user's emotional state may be reflected in the emotional reactions of other actors in the scene, actors whose faces are visible from the user's perspective.” Despite the fact that the validity of emotion recognition based on facial images is highly questionable, ¹⁶ the call proposes to extrapolate further from that approach, using inferences attributed to the user as a basis for their own, not directly observable, affective state.

Making the interior states of warfighters available for assessment and management is matched by a concern for the “introspective” capacities of machines. To realize the promise that weapon systems can become collaborative and trustworthy warfighting partners, DoD calls reiterate the need for technologies with the capacity to adjust and fine-tune their own operations in response to unfolding circumstances. A BAA released in August 2021 by DARPA's Information Innovation Office, titled “Learning Introspective Control,” “aims to develop machine learning-based introspection technologies that enable physical systems, with specific interest in ground vehicles, ships, drone swarms, and robotic systems, to respond to events not predicted at design time” (#30). In May 2022, an Information Innovation Office BAA on “Assured Neuro-Symbolic Learning and Reasoning” reiterates “one small problem” within the capabilities of existing warfighting systems, namely that “high levels of autonomy remain elusive” due to the limits of ML-based approaches (#31). This call responds to the Defense Science Board Report on Autonomy (Office of the Undersecretary of Defense, 2016), emphasizing the importance of trustworthiness: “Informally, trust is an expression of confidence in an autonomous system's ability to perform an underspecified task [and is] key to DoD's success in adoption of autonomy.” The call is focused on the question of autonomous systems to be used in ISR, logistics, planning and Command and Control, where “The purported benefits are many, including—(1) improved operational tempo and mission speeds; (2) reduced cognitive demands on [the] warfighter in operation and supervision of autonomous systems; and (3) increased standoff for improved warfighter safety,” that is enhanced action at a distance. Machine autonomy, in sum, is a requirement for human–machine teaming, which in turn enables and legitimizes the further automation of weapon systems.

In the context of the international debate over the legality of automated targeting, and civil society campaigns against autonomous weapon systems, the invocation of “human-machine teaming” serves as an assurance that human control will govern the incorporation of AI technologies into the kill chain. Critics have argued, however, that there is an irresolvable contradiction between increasing the speed of warfighting through automated targeting and adherence to the requirements of International Humanitarian Law. ¹⁷

Distributed warfighting/lethality

In the years following the terrorist attacks of 11 September 2001 the focus of U.S. military strategy shifted from superpower conflict to what critical scholars have designated as “everywhere war” and the increasing use of Special Operations Forces (Chamayou, 2014; Masco, 2014). The turn to Special Operations and a program of continuous, distributed counterinsurgency is accompanied now by the revival of competition for military dominance, framed as a new arms race between the US and China. Both developments support a shift from a traditional focus on maneuvers at “the front” to operations conducted “on the edge,” which in turn require new modes of “command and control” across forces. Conjoined with the aspiration to greater distribution of command and control is a growing concern for integration across the longstanding divisions between the various armed forces, intensified by the increasing expansion of all forces into multiple “domains” of warfighting (air, land, sea, space, and cyber), imagined in the vision of Joint All Domain Command and Control (JADC2) as a set of interlocking and interoperable “systems” (U.S. Department of Defense, 2022a).

Consonant with these developments, a June 2017 BAA from DARPA's Tactical Technology Office titled “Innovative Systems for Military Missions” (#32) calls for demonstrations and research programs aimed at “cross domain systems” that “validate revolutionary precision engagement capabilities.” The requirements for precision engagement include “the development of persistent, global surveillance architectures; real-time data updates, at scale; provision of real-time, decision quality information; and the demonstration of novel approaches that support rapid and affordable integration,” with a focus on the deployment of “swarm technologies.” Two years later, in March 2019, the aspiration to command and control “at the edge” is evident in a BAA issued by DARPA's Microsystems Technology Office on “Real Time Machine Learning,” addressing the problem of how ML systems can incorporate “new datasets in the field.” (#33) A joint effort with the National Science Foundation, the aim is the “rapid development of energy efficient hardware and ML architectures that can learn from a continuous stream of new data in real time.”

In a February 2019 BAA, DARPA's Strategic Technology Office (STO) goes directly to the point, stating that the STO “is seeking innovative ideas and disruptive technologies that provide the U.S. military increased lethality in an era of eroding dominance” ¹⁸ (#34). The announcement goes on to explain that U.S. dominance in the development of sophisticated weapon systems is now being “challenged by peer competitors.” The response must be a “new paradigm that values ‘lethality’ over monolithic system dominance. Whereas dominance is measured by comparing capabilities across systems, lethality is measured by the ability to deliver a desired effect at will, regardless of the system or systems of systems involved.” This lethality is to be delivered through “Mosaic Warfare,” enabling “fast, scalable, adaptive joint multi-domain lethality. It is the disaggregation of effects chain functions (e.g. Find, Fix, Target, Track, Engage, and Assess or F2T2EA) across a heterogeneous mix of manned and unmanned platforms from all domains.” A BAA in October 2021 from the STO on Mosaic Warfare continues the program (#35), with an emphasis on the achievement of a “disaggregation of capabilities and re-composition of adaptable effects chains at mission speed.” In military doctrine the most prominent “effects chain” is a successful “kill chain,” a sequence of intelligence analyses and decisions leading up to a potential strike, in a process promoted as an appropriate application of AI-enabled optimization. ¹⁹

We see, in sum, the emergence of a paradigm of U.S. military dominance across two, parallel constructions of the geopolitics of warfighting, one characterized by ongoing, “low level” but no less lethal, operations of so-called counter-terrorism and counterinsurgency against “irregular” and often non-state actors, the other by a return to superpower conflict, most imminently in the Indo-Pacific region and focused on China. Within the DoD's R&D agenda there is a place for AI-enabled weapon systems in both, and both provide justificatory grounds for research directions well aligned with the current interests of the commercial and academic AI research community. ²⁰