How Liability Regimes Shape Risk-Taking Behavior in AI-Agent Populations

We present the first simulation comparing substantive liability rules in populations of LLM-backed agents. Twenty agents (10 Claude Haiku 4.5 + 10 Claude Sonnet 4.6) interact across 30-round episodes under six liability regimes — no liability, negligence, strict, strict + insurance, joint-and-several, and market share — under both moderate and heavy harm-magnitude tails. We also test two information conditions: an anonymous baseline where each agent sees only population aggregates, and an identified condition where each agent additionally sees the state of five sampled peers per round. Across 120 simulations (5 seeds per cell), three findings emerge. First, regimes produce a clean, robust ordering on mean care (Friedman $p < 0.005$): market share and strict + insurance yield the lowest risk-taking, no liability the highest, and the perpetrator-pays regimes (strict, joint-and-several) sit in between; pooling and sharing regimes (strict + insurance, joint-and-several) additionally prevent the judgment-proof events that no liability and negligence allow under heavy tail. Notably, no individual-optimization benchmark we compute — risk-neutral or ruin-averse — prescribes any positive care under these payoffs, so the observed caution under every liability regime is not explained by expected-value maximization. Second, peer visibility raises care substantially under strict but produces near-zero effect under negligence or strict + insurance: visibility matters where regime structure has not already locked behavior at a focal point, and saved reasoning traces show agents articulating regime-aware reasoning, including sustained strategic deep-pocket avoidance under joint-and-several. Third, and more tentatively, the two model classes respond differently — Haiku scales its care across regimes while Sonnet anchors near the population midpoint — though this capability-tier contrast rests on two same-provider models and we flag it as preliminary. These results characterize liability regimes as observable, measurable behavioral interventions on LLM-agent populations in a stylized environment; we frame them as an exploratory step toward empirically evaluating ex-post-liability proposals, not as direct evidence about how deployed AI systems would respond.