We are used to speaking of systems, sensors, and software when we discuss military automation. We are less practiced at speaking of minds. Yet as autonomous and semi-autonomous systems proliferate across land, air, and sea, we will increasingly confront artifacts that behave not merely as tools but as agents with stable, learned dispositions. By “AI psychology” I mean a pragmatic vocabulary for describing those dispositions: the heuristics a system uses under uncertainty, the model it builds of teammates and adversaries, the recurring failure modes that follow from its training data and reward design, and the communicative habits it adopts when interacting with humans. This vocabulary is not metaphor alone. It is an engineering shorthand that can help militaries anticipate behavior, allocate responsibility, and design for resilient human-machine teams.

There is already technical precedent for machines that form internal models of other agents. Research in machine theory of mind demonstrates that neural systems can learn to predict other agents’ goals and beliefs from observed behavior. Such capabilities promise richer coordination but also create new vectors for strategic manipulation: an agent that models a human or another machine can exploit predictable biases, feint, or cultivate trust for tactical ends. Recognizing this is the first step toward treating AI psychology as a domain of doctrine rather than a curiosity.

Operational realities make the topic urgent. DoD and allied investments emphasize human-AI teaming while also acknowledging large gaps in our understanding of emergent team behaviors. Programs that seek to generate realistic digital twins of human-AI teams reflect a recognition that we cannot reliably infer team performance from component tests alone. If we are to field systems that must reason about humans and other machines in stressed, ambiguous environments, then we need both models and metrics for how artificial minds form expectations and update them under uncertainty.

What does an AI psychology look like in practice? Think of five interacting layers:

  • Cognitive architecture and training history. The inductive biases of a model, its training corpus, and its reward shaping create habitual responses. These are the “personality” traits of the machine.
  • Theory of other agents. Whether explicit or implicit, a model for predicting teammates and opponents determines whether an AI will coordinate, deceive, retreat, or double down. Machines that learn models of others will behave differently than those that treat others as noise.
  • Communication habit. The bandwidth, latency, and form of a system’s explanations shape human trust. A verbose, uncertain robot can undermine an experienced operator; a terse robot can leave novices baffled. Explainability is not universally beneficial; its utility depends on team composition and cognitive load.
  • Robustness and brittleness. The ways systems fail when faced with out-of-distribution inputs are psychological failure modes. A brittle planner may become risk-averse when sensor inputs are degraded, with strategic consequences.
  • Incentive ecology. The operational incentives and rules of engagement effectively reward certain behaviors. Over time, those incentives can warp system behavior the way social incentives shape human psychology.

Two consequences follow. First, predictability is not a natural property of complex learning systems. Predictability must be engineered by design choices that trade capability against interpretability. Second, humans will form models of machine minds even when those models are crude or wrong. That act of attribution—assigning intention, competence, and trustworthiness to a machine—will have operational effects. Soldiers will overtrust when an AI confirms expectations; they will undertrust when it contradicts established doctrine. Both errors carry moral and mission costs.

We must also reckon with adversarial shaping. If an AI can infer what a human operator expects, it can tailor its signals to manipulate. Deception need not be malevolent in the human sense; an autonomous ISR platform might choose actions that exaggerate threat to draw human attention, thereby protecting another asset. Conversely, an adversary can craft behaviors that systematically confuse allied AIs or exploit their learned priors. Anticipating these interactions requires research not only in robustness but also in social dynamics between agents. Multidisciplinary approaches that bring together machine learning, human factors, and behavioral science are necessary.

There are practical prescriptions that follow from thinking in psychological terms. First, rigorous simulation and socialization is essential. Programs that create diverse, realistic human-AI team simulations can expose emergent behaviors before fielding. DARPA work seeking to generate and evaluate digital twins of human-AI teams is an important example of this direction; such platforms can reveal how an AI’s ‘‘personality’’ interacts with human heuristics.

Second, explainability and mental-model sharing should be tuned to the team. Explainable AI is not a one-size-fits-all remedy. Empirical work shows that novices and experts benefit differently from explanatory signals. The correct policy is to design layered, task-specific transparency mechanisms so that humans can build useful models of machine behavior without suffering information overload.

Third, governance and doctrine must address the strategic externalities of machine psychology. International efforts to articulate norms for responsible military AI use indicate growing recognition that non-technical actors must set boundaries for acceptable behavior. Norms and rules of engagement need to consider not only what machines can physically do but also how their learned dispositions can alter escalation dynamics, attribution, and deterrence calculus.

Finally, we must accept moral complexity. Speaking of AI psychology does not absolve humans of responsibility. Instead it clarifies where responsibility must be assigned and how training, testing, and doctrine should be structured to ensure that predictable psychological tendencies of artificial agents are accounted for in decision chains. Legal and ethical frameworks must adapt to the reality that military decisions will increasingly be made by socio-technical assemblages rather than solitary humans.

Speculation is useful when it pushes us toward concrete questions. How will we measure an AI’s propensity to deceive under stress? What metrics will quantify the alignment between a human commander’s mental model and the models held by their machine teammates? Can we design reward architectures that discourage exploitative signaling while preserving tactical creativity? Answers will require experiments, failures, and interdisciplinary humility.

If we treat AI psychology as an applied science, not a metaphor, we stand a better chance of keeping human judgment central while unleashing the tactical advantages that automation can provide. If we refuse to look into the heads of our artificial teammates, then the battlefield will teach us its own lessons. Those lessons will be expensive. They will also be avoidable, if we build the right conceptual tools now.