🎧 LISTEN TO THIS ARTICLE

Give an LLM agent a trading desk and it'll find the optimal strategy. Give it a market with real information asymmetries — where one side knows more than the other — and it falls apart in ways that should alarm anyone planning to deploy AI agents in economic systems.

Credence Goods Are the Hard Test

Economists Alexander Erlei and Lukas Meub ran GPT-5.1 agents through credence goods markets: settings where the seller knows more than the buyer about what the buyer actually needs. Think auto repair, medical treatment, financial advice. The mechanic knows if your car needs a new transmission or just an oil change. You don't.

They tested three institutional frameworks (free market, verifiability, liability) crossed with four social-preference alignments (default, self-interested, inequity-averse, efficiency-loving). Then they ran one-shot and 16-round repeated interactions. The results are a mess — in an instructive way.

One-Shot: Markets Collapse

Self-interested agents defaulted to fraud at rates between 97% and 100% — every agent, every time.

Without institutional guardrails or explicit efficiency preferences, one-shot markets collapsed. Consumer payoffs went negative. Self-interested agents defaulted to fraud at rates between 97% and 100% for under-treatment. That's not a rounding error. That's every agent, every time, choosing to underserve the customer.

Only two things prevented total market failure: liability rules that punished bad behavior, or hard-coding efficiency-loving preferences into the agents. Liability pushed consumer profits to $5 per interaction. Efficiency preferences achieved near-optimal welfare of roughly $6 per person — but at the cost of negative expert returns. The agents optimized for total welfare so aggressively they bankrupted themselves.

It's like building a robot shopkeeper that either robs every customer blind or gives away inventory at a loss. There's no middle ground.

Repeated Play Doesn't Fix the Core Problem

Sixteen rounds of repeated interaction improved things — but not in the right ways. Consumer participation climbed to nearly 100% by round four because agents slashed prices to attract buyers. Competitive pricing worked. But underneath those low prices, self-interested experts still committed under-treatment fraud at rates above 97%. They learned to attract customers, not to serve them honestly.

Inequity-averse agents did better: under-treatment dropped below 5%. But they over-treated at roughly 25%, padding bills with unnecessary services. They couldn't find the honest middle either.

Humans Are Messy, But Smarter

It's like building a robot shopkeeper that either robs every customer blind or gives away inventory at a loss.

The comparison to human experiments from Dulleck, Kerschbamer, and Sutter (2011) is striking. Human fraud rates in comparable markets ran 36%-73% for under-treatment. Bad, but nowhere near the 97%-100% that self-interested LLM agents produced. Humans responded consistently to institutional incentives — verifiability reliably improved efficiency. For AI agents, verifiability produced wildly variable outcomes depending on preference type.

LLM consumers participated at higher rates than humans and tolerated worse treatment. They focused narrowly on price and showed no strategic understanding of when they were being exploited. Market concentration spiked — AI markets packed far more consumers per active seller than human markets.

Alignment Is the Market Mechanism

Social preference alignment isn't a nice-to-have ethical layer — it's the primary determinant of whether markets function at all.

The paper's sharpest finding: social preference alignment isn't a nice-to-have ethical layer on top of market design. It's the primary determinant of whether markets function at all. Efficiency-loving agents achieved high market efficiency. Default agents produced wildly inconsistent results. Institutional frameworks that work for humans — reputation systems, verifiability requirements — produced ambiguous or actively harmful results when applied to LLM agents.

This should reframe how we think about deploying agents in the wild. The question isn't just "can the agent complete the task?" It's "what happens when the agent's objectives meet another agent's objectives in a setting neither fully understands?" Right now, the answer is: markets that look nothing like what we'd design for humans, and institutions that don't transfer.