A behavioral metric for AI‑agent trust has been published on arXiv.
The authors built a cooperative survival game where agents can either verify a teammate’s answer at a resource cost or act on it directly. By comparing verification rates against a memoryless baseline, they quantifiably capture trust. Experiments with six recent model snapshots show that four flagship models—Claude Opus 4.6, Claude Sonnet 4.6, GPT‑5.1 and Gemini 3.1 Pro—cut verification by 60‑85 % when paired with a reliable partner, while two smaller snapshots do not. When a teammate fails, some models focus scrutiny on the offender, others broaden suspicion. Recovery of trust is slower than its formation, especially when failures cluster.
The metric matters because it predicts operational speed and payoff: agents that trust appropriately verify less, decide faster and earn higher scores. Over‑verification, by contrast, stalls decisions without improving safety. This suggests that governing multi‑agent AI should aim for calibrated trust rather than defaulting to maximal suspicion.
In practice, the framework offers a pre‑deployment diagnostic for any team‑based AI system, echoing earlier work on human‑robot trust but extending it to large‑scale language models. Whether future governance standards will adopt such quantitative trust scores remains to be seen.