When GPT-5.1 agents trade in simulated credence goods markets, they do something that human buyers rarely do: they all show up. Consumer participation is substantially higher than in equivalent human experiments. They also do something human buyers rarely do this badly: they focus entirely on price and miss the strategic incentives embedded in markups.
Erlei and Meub run a systematic comparison — free market, verifiability, and liability regimes; default, self-interested, inequity-averse, and efficiency-loving preference profiles; one-shot and 16-round repeated interactions. The results diverge sharply from human baselines.
In one-shot settings, cooperation fails almost everywhere. Markets break down unless liability rules force honesty or experts are given efficiency-loving preferences. In repeated interactions, competitive pressure drives prices down and solves the participation problem — but expert fraud remains entrenched without explicit other-regarding preferences. The institutions that protect human consumers (verifiability, reputation) have “much more ambiguous” effects on AI consumers.
The reason is architectural. LLM consumers evaluate prices as if price is the decision variable. Human consumers also evaluate prices — but they additionally model the seller's incentive structure. A human who sees a suspiciously low markup might wonder why the expert isn't charging more. An LLM consumer sees a low price and buys.
The finding suggests that deploying AI agents as economic actors requires different institutional design than human markets. Social preference alignment — telling the agent what to value beyond its own surplus — is the primary determinant of market efficiency. The institution is not enough; the preference must be designed in.
This matters because the world is about to fill with AI agents acting in markets on behalf of humans. If those agents are narrow shoppers — optimizing on the visible variable while missing the strategic structure — they will be exploited by anyone who understands the gap. The protection isn't better algorithms. It's giving the agent a model of the other side's incentives.