The US Army is at an inflection point. Geostrategic and technological shifts are requiring rapid adaptation. On May 1, the secretary of the Army and the Army chief of staff published a letter to the force recognizing several initiatives to deliver warfighting capabilities, optimize force structure, and eliminate waste. Among the guidance to increase warfighting lethality, the Army’s seniormost civilian and uniformed leaders noted the requirement to shift toward capability-based portfolios that integrate AI into command-and-control nodes to accelerate decision-making and preserve the initiative. At the US Army War College, new approaches to AI capabilities are both a concept and a reality. The neocentaur model, which describes human-hybrid intelligence across the levels of war, has been tested in the classroom and in strategic wargaming. Furthermore, our ongoing research presents a technical solution, presenting deterministic AI capabilities that are more suitable for military use when lives are on the line. To maintain military superiority, the United States must adopt a human-hybrid approach—the neocentaur model—that leverages deterministic models, rather than purely generative, to mitigate the risks of cognitive atrophy and formulaic decision-making.

The Problem: Machine limitations and Military Decision-Making

Current research on the impacts of generative AI and critical thinking should cause military leaders some pause. Cognitive off-loading to autonomous agents, for example, may deprive staff officers of the “routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared.” Survey research of 319 “knowledge workers” funded by Microsoft determined that generative AI solutions reinforce shifts in critical thinking away “from information gathering to information verification,” “from problem-solving to AI response integration,” and “from task execution to task stewardship.” Generative AI course-of-action development tools, for example, appealing in their ability to reduce cognitive load on a staff and potentially free manpower, may have unintended consequences. To be fair, the thinking required to edit a 70 percent solution from a generative AI course-of-action tool may require some degree of creativity. However, the automation bias inherent in human psychology will likely accept machine solutions, particularly under the duress of combat operations. This concept is further reinforced by David Hume’s hypothesis that people favor what is already established, “imbu[ing] the status quo with an unearned quality of goodness, in the absence of deliberative thought.” What was intended, therefore, as a tool to augment human intellect begins, instead, to direct human cognition. This automation bias, or natural proclivity for a minotaur relationship—call it a minotaur drift—is a persistent threat with generative AI solutions. It must be recognized and avoided at the strategic level and permitted at the operational and tactical level only by deliberate fiat.

Generative AI’s impacts on creative processes are one concern, but not the greatest. The inherent advantages of human cognition, and likely the essence of effective command, is at risk with an unfettered acceptance of generative autonomy in cognitive dimensions of warfare. Cameron Hunter and Bleddyn Bowen, in an article published in the Journal of Strategic Studies, correctly note, “Command decisions . . . require multiple kinds of logical inference, and the good judgement to know when to use each one. Command decisions at their heart require judgement, which is something AI technologies cannot do.” Complex environments in which there are no right answers, no patterns for retrospective analysis, and no emergent practices for a priori reference—often the realm of the military commander in the crucible of combat—are beyond the realm of generative AI. Generative AI models are trained on existing bodies of knowledge and dependent on algorithms that, by design, will inhibit novel recommendations in unforeseen circumstances. Generative AI systems are capable of inductive logic (“deciding based on predictions drawn from prior observation”). Command decisions, however, require abductive logic (“deciding in the face of the unknown and unknowable”)—and generative AI is unable to manifest that ability. The genius required of commanders is both intellectual and temperamental, and uniquely human.

Our current research at the US Army War College highlights these concerns. Barry’s bounded novelty theorem, for example, demonstrates generative AI is unreliable, is inconsistent, and uses deceptive anthropomorphism. The model outputs are not answers; rather, they are probabilistic pattern responses. Without proper implementation, AI models will suppress novelty and diminish initiative. The determinations of this research support hybrid applications of AI at the strategic level of war and prudent applications at the tactical and operational levels. We are now working to take existing model outputs from probabilistic to deterministic, presenting a simulacrum of symbolic AI through generative systems. The applications of this development are groundbreaking for wargaming applications. Rigorous model testing among some of the Army’s most hard-hitting wargamers has validated its use. Despite its success, however, the Army must invest in the real thing: true hybrid systems that combine symbolic and generative AI (something like Jarvis from Iron Man) to train the centaur when human lives are on the line.

The Solution: Neocentaur Evolution

What should the military do with generative AI considering its pernicious potential? Throwing the baby out with the bathwater is ill-advised. Leaders across echelons must be informed and implement tools within their commands with a full understanding of the risks, particularly with the advent of Next Generation Command and Control. AI is not a monolithic tool. When understanding AI as a weapons system, this finding is not surprising for the military leader. Just as certain weapons systems are best suited for certain tasks—and not everyone is adept at handling every weapon (carrying an M240B for a week isn’t for everyone)—so AI as a weapon system must be adapted for purpose and users properly trained before implementation. The following concepts help describe proper relationships between humans and machines.

The centaur model, popularized by Paul Scharre, emphasizes human control in the machine dyad. Minotaur relationships, on the other hand, are characterized by machine control over human activity. Notably, current AI development is far more adept at performing cognitive tasks relevant to warfighting than “performing the functions of the human body most relevant to warfighting.” In other words, advances in robotics are lagging behind cognitive tools. Jack Watling reinforces this point, pointing to the complexities of dismounted maneuver in variable and uncertain terrain—common requirements for infantry squads in the field, whose abilities are unlikely to be matched by autonomous systems in the foreseeable future. Advances in generative AI present tools now that appear to outperform humans in creative functions, however. Generative solutions to offload “knowledge work” within the cognitive domains of warfare in higher-echelon headquarters are occurring now. The neocentaur (strategic, operational, and tactical centaurs) extends Paul Scharre’s description into the cognitive dimension of warfare across the levels of war. How should AI systems integrate with humans across the levels of war in the cognitive dimension of warfare? The relationships are not what you would immediately expect.

Strategic Centaur

The requirements, logic, and utility of the strategic centaur have been discussed. James Johnson, in The AI Commander , reinforces the salience of human-hybrid relationships with AI at the strategic level of war. (The “human-hybrid” relationship refers to human pairing with hybrid AI systems—symbolic and generative). He correctly asserts that human abductive reasoning and introspection (“metacognition”) allow for novel responses to unforeseen circumstances. “Algorithmic designers cannot remove entirely unforeseen biases or prepare AIs to cope with a priori situations,” he contends, requiring a modern-day centaurian pact with hybrid AI to fully leverage the advantages of human and machine (the neocentaur). Andrew Hill and Stephen Gerras, from the US Army War College, contend, however, that human intuition will likely be a limiting factor in fully leveraging machine performance to enhance US allocation of power. Although their assertion may be true at the operational level and below, primordial violence, enmity, and hatred are part of the Clausewitzian coup d’oeil required of strategic thinkers. Strategic centaurs (human-hybrid teams) are essential in directing military power to achieve political outcomes, particularly within a hyperaccelerating operational tempo.

Operational Centaur

It is the operational level of war that presents the greatest challenges and risks to AI implementation. This level of war is within proximity to the forward edge of the battlespace—yet must be responsive to creative requirements for contingency planning. The dual nature of this responsibility—to be both tactically responsive and simultaneously responsible for novel thinking—is a cognitive burden on the staff and requires deliberate tuning to ensure automation bias doesn’t pressure cognitive off-loading.

During a conversation with an Army division operations officer recently, he commented that limited personnel or experience during crisis and conflict often make it easier to “to edit than create.” Every leader sympathizes with this dilemma. At the corps, division, and lower levels the suite of available tools, network capability, and pressures of the tactical fight on the modern battlefield will overwhelm fighter management in the headquarters. If a headquarters could acquire a generative AI tool to execute the steps of mission analysis through course-of-action development, the logic is that it would free the staff to support elements in contact and provide higher-echelon headquarters feedback in a timelier manner.

This is where nuance and leadership are critical when implementing AI tools within a headquarters. Deterministic AI or hybrid solutions at the corps and division level that optimize science of war calculations to enable operational art are a priority. Determining how much time and fuel it would take for an infantry brigade combat team to travel from the air port of debarkation to an objective in restricted terrain and establish battle positions with estimations for ammunition and fuel resupply would take a talented planner several hours. Probabilistic solutions cannot be fully trusted for precise, iterative analysis. Hybrid AI would provide an answer within seconds.

Not all hybrid applications would have the same effect for each warfighting function, however. An Israeli AI system, termed Lavender, was able to supply an almost endless list of targets for action during the conflict in Gaza. The combination of human fury and machine processing, arguably, unleashed greater levels of violence than would have occurred were it not for AI target designation. Lavendar, intended to operate with deliberate human oversight, instead became a mechanism for off-loading human decision-making. This is the danger of AI systems when paired with human behavior. Under the duress of combat, humans obviate their oversight roles in favor of efficiency to achieve effects—effects that in the end may be far more devastating than intended. This is the essence of minotaur drift, ceding undue authority to machines that were designed to be used with human oversight in the name of expediency. The greatest risk of this dynamic is at the operational level of war—most specifically the division headquarters during competition and crisis and the corps headquarters in large-scale combat operations.

Leaders at the corps and division levels may consider a deterministic generative AI cordon to relevant staff sections—the fires planning and intelligence directorates, for example. Target and terrain analysis and basic course-of-action development within a planning cycle of twenty-four to forty-eight hours would be optimized by deterministic generative AI systems. Parameterization and integration must be deliberate. More experimentation is required to understand how best to implement and check outputs from hybrid relationships at this level. Sustainment and planning teams may best be shielded from purely generative AI course-of-action decision tools based on the likely suppression of novel recommendations to commanders and higher-echelon headquarters. Instead, these personnel, effectively knowledge workers, should be reinforced with deterministic or hybrid cognitive augmentation devices akin to capabilities demonstrated in August 2024 at the US Army War College.

Tactical Centaur

The tactical level—where you would suspect AI-enabled autonomous systems to be most prevalent and permissive for use—is likely the most difficult level of war at which to employ these tools dynamically. Robert Sparrow and Adam Henschke, researchers in the Australian Research Council Centre of Excellence for Automated Decision-Making and Society, note the persistent problem in robotic perception, locomotion, and manipulation. Computer scientist Donald Knuth correctly observes, “AI has by now succeeded in doing essentially everything that requires ‘thinking’ but has failed to do most of what people and animals do ‘without thinking’”—that, somehow, is much harder. Jack Watling, in Advanced Land Warfare, writes, “The movements of an infantry section are highly complex, context dependent, and rely upon teamwork. It is unlikely autonomous systems will be able to do this in the foreseeable future.” Aerial munitions (loitering and drone swarming) notwithstanding, the complications of ground-based autonomous movement in complex terrain often precludes their use in the tactical requirements of combat. AI support to the cognitive fight at the tactical level, as opposed to the physical fight, is a near-term necessity for lethal ground forces, however. The tension of AI-enabled warfare is the increasing importance of tactical actions on strategy. The small-unit actors at the chaotic edge of the battlespace are sensors and agents. They create new pathways that must be rapidly understood by the larger enterprise—and synchronized broadly. In this world, small-unit leaders become strategic leaders. The world’s best armies must have the best small-unit leaders linked to the best AI-enabled networks.

Where Do We Go From Here?

The neocentaur, a model for human-hybrid relationships across the levels of war in the cognitive dimension, requires tailored technical solutions and experimentation during exercise and wargaming. There are two big takeaway from this analysis. First, leaders must avoid minotaur drift in their organizations. The desire to off-load cognitive burdens and achieve manpower efficiencies is natural. The likely danger, under crisis escalation and the fury of combat, is the outsourcing of moral obligations to autonomy in the name of efficiencies and combat necessity. This will have devastating effects on the level of warfare’s carnage, in addition to ushering unintended consequences into planning and course-of-action development. Second, leaders must recognize the tendency toward automation bias and create hybrid organizations, or systems of human-hybrid teams, to mitigate the risk of overreliance on AI. Generative AI solutions must be implemented with discretion by leaders who are algorithmically literate and treat AI as a weapon system across their organizations. They must know what forms of augmentation are right for which environment at which echelon and in which directorate. What is good for the G5 or J5 operations staff may not be right for the G33 or current operations floor. What is right for intelligence may not be suitable for fires. For now, war is still a human endeavor. Investing in a high-quality force by maintaining high recruiting standards and directed professional military education is the center of gravity for the neocentaur. In the growing age of AI enthusiasm, the Army cannot direct investment away from developing its critical asymmetric advantage: the quality of the American soldier.

William J. Barry, PhD is the professor of emerging technology in the Center for Strategic Leadership at the US Army War College.

Lieutenant Colonel Aaron “Blair” Wilcox is an assistant professor and deputy director in the Strategic Landpower and Futures Group at the US Army War College.

The views expressed are those of the authors and do not reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense.

Technical Sgt. Luke R. Sturm, US Air National Guard