The Dangerous Assumption Behind AI-Powered Crisis Planning

The Pentagon is betting on artificial intelligence to accelerate military planning, with the Defense Innovation Unit's Thunderforge initiative designed to integrate AI-powered wargaming into operational decision-making. The objective appears to be to process vast amounts of information, generate courses of action, and help commanders respond to evolving threats at greater speed. The technology is advancing rapidly, but one question remains unresolved. Who takes responsibility when the machine gets it wrong?

A 2026 study from King's College London examined how three leading AI models handled 21 simulated nuclear crises. Across 329 turns of play, the models chose nuclear signaling in 95% of scenarios, highlighting the risks of allowing systems to navigate escalation without sufficiently understanding the human consequences of their recommendations. The experiment was a simulation, not evidence that AI would behave identically in a real crisis. Its findings nevertheless raise difficult questions about machine reasoning under pressure.

The appeal of AI in strategic planning is obvious. Models can process information at speeds that human teams cannot match, identify logistical constraints, and generate possible responses across complex scenarios. Yet a crisis is not a spreadsheet with missing variables. Adversaries can misread intentions, make emotional decisions, change direction unexpectedly, and respond to pressure in ways that rational models may struggle to anticipate.

Military planners have long understood the limitations of simulation. A 2025 article in the U.S. Naval Institute's Proceedings noted that human judgment remained necessary during wargames when established models could not account for variables such as electronic warfare effects. The challenge is becoming more urgent as AI takes on a larger role in exercises designed to prepare decision-makers for uncertainty.

Devin Ellis, Director of Strategic Business Development at National Security Innovations (NSI), a boutique security advisory firm, argues that the corporate world should approach this shift with a more disciplined understanding of what AI can contribute. NSI brings expertise in national security consulting, crisis simulation, and strategic decision-making to organizations operating outside traditional government circles.

Ellis sees substantial value in AI-assisted exercises, particularly when the technology functions as a decision-support resource. An agent trained to retrieve detailed information about Asian ports, for example, could provide cargo capacities, water depths, operating conditions, and other relevant logistics data almost instantly.

"AI can instantly retrieve unfamiliar port statistics and generate 20 tailored suggestions, making it a powerful decision-support tool," Ellis explains.

The implications extend into corporate crisis training. An AI agent equipped with information about a company's operating costs, supplier agreements, fuel contracts, and logistical constraints could help executives assess routes, schedules, and expenses during a simulated disruption. Tasks that once required analysts to spend hours encoding information into software could be completed in seconds or minutes.

Ellis sees this acceleration as a way to improve the quality of learning within an exercise. The technology gives participants more time to examine consequences and consider strategic choices, while allowing smaller organizations to engage with sophisticated simulations that may previously have been financially inaccessible.

The risk emerges when speed and optimization begin to substitute for judgment. Ellis cautions that agentic systems operating without supervision can produce recommendations that appear technically rational while disregarding the values of the organization using them.

An AI model might determine that sacrificing employees during a kidnapping crisis is financially preferable to paying a ransom. Such a recommendation could emerge from an optimization framework that fails to account for a company's ethical obligations or the human realities surrounding the decision.

"Even when we tell agentic models to have ethics and we tell them to go out and train on large corpuses of ethical data from human history, well, guess what? They train on everything," Ellis observes. "You do have to be very conscious of teaching the agents to serve the capacity that they are supposed to serve, not to freelance."

The concern extends to the assumptions embedded in predictive modeling. Human beings under stress do not always behave according to optimized strategies. Fatigue, fear, emotional responses, and individual cognitive patterns can produce decisions that an algorithm may regard as irrational.

"Your agent that you are playing against may have made very different choices for very good reasons as opposed to another human who is going to make suboptimal choices for emotional reasons, for lack of sleep, etc.," Ellis says.

His position places human accountability at the center of AI-enabled crisis preparation. The machine can expand the information available to a decision-maker, accelerate analysis, and introduce scenarios that would otherwise remain unexplored. The final judgment belongs to the person responsible for acting on that information.

"At the end of the day, are you going to let a computer that cannot explain itself to you, even if its choice was optimal, if it cannot tell you why it made that choice?" Ellis asks. "People at that level of decision-making are generally going to say, well, unacceptable for me because the buck actually stops with me, not with the agent."

For corporate leaders confronting geopolitical disruption, supply-chain vulnerability, or humanitarian emergencies, the lesson carries practical weight. AI can make preparedness more accessible and simulations more responsive. Its greatest contribution may emerge when it gives people better conditions in which to think, question assumptions, and make consequential decisions before a crisis removes the luxury of time.