AI agents build trust like humans: study measures 60-85% verification drop

Researchers at arXiv have introduced a behavioral framework to measure trust between AI agents, addressing a critical gap as language-model agents increasingly work in teams. In a cooperative survival game, agents spend resources to verify teammates' work, while trusting a wrong answer can be fatal. Reduced verification relative to a memoryless baseline serves as an observable trust metric. Testing six frontier model snapshots, the study found that four—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 consistently reliable teammate. Smaller models showed little to no such adjustment.

Failures reversed this discount, but models reacted differently: some concentrated renewed scrutiny on the culprit, while others became more cautious toward the entire team. Recovery was slower than formation, and clustered failures sustained suspicion far longer than the same number of failures spread apart. Models that formed trust verified less, decided more quickly, and achieved higher payoffs. Persistent over-verification was associated with indecision rather than safety. The authors argue that trust dispositions can be measured before deployment and that calibration—not maximal suspicion—should guide governance of multi-agent AI systems. The findings have immediate implications for teams of autonomous AI agents in finance, logistics, and scientific research.