Before It Advises the Commander: Measuring the Hidden Behavioural Biases of AI Decision Aids
AI agents are rapidly entering planning, wargaming, intelligence and decision support. Yet the usual assessment methods (static benchmarks, red teaming and scoring against pre-declared harm categories) rarely show how an agent behaves under pressure, with incomplete information or alongside other agents. They also measure the model in isolation. The operational risk sits in the human-machine team: an AI aid that is quietly overconfident, drifts toward an adversary's framing or selectively discloses information can turn a correct human judgement into an incorrect one. This session presents a simulated market testbed that Yasin Science & Technology is developing to measure an AI system's hidden preferences from the outside. We use markets because every decision becomes a complete, priced commitment, a bid defined by time, asset, quantity and price, so behavior can be compared quantitatively instead of interpreted. The methodology combines a configurable market engine with several auction mechanisms and a synthetic information feed with known ground truth, whose timing, reliability and framing can be controlled. It adds reference agents built from at least ten large language models running in an identical harness, plus operator-proxy agents calibrated against human behavior from experimental economics. Statistical classifiers then detect herding, risk asymmetry, misrepresentation, collusion and drift, reporting confidence bounds and full provenance so any result can be rerun. The environment builds on an existing trading engine and agent harness that already operate in financial markets. The implications for training are direct. The same instrument can screen AI tools before fielding, generate deception and information-manipulation scenarios for human-machine teaming exercises, and teach operators to recognize when an AI aid is shaping their judgement instead of informing it.