ИИ-агенты не смогли воспроизвести рыночное равновесие в эксперименте, имитирующем аукцион 1962 года

In 1962, an economist named Vernon Smith ran an experiment with students at Purdue. Half got a slip with a secret number: The most they could pay for a pretend good. The other half got a different secret number: The least they’d accept to sell it. Nobody saw anyone else’s slip, and an auctioneer at the front simply called out and recorded each bid and ask, declaring a trade whenever one met the other. There was, unknown to any of them, a single price — the point economists call the ‘equilibrium’ — at which the gains available from the whole market were as large as possible. Yet, round after round of calling out numbers and haggling, strangers with no shared information beyond overheard bids and asks had pushed prices strikingly close to it, as though the room had solved a puzzle none of them could individually see. This is the double auction, the mechanism underlying most centralised financial and commodity markets. Smith would go on to share the 2002 Nobel Prize in Economic Sciences for establishing laboratory experiments as a tool of economic analysis. Large language models now take on the roles once played by Smith’s students, negotiating and trading on people’s behalf. A team of four researchers tested whether AI agents reproduce Smith’s result in their recent study, titled Competitive market behaviour of LLMs. They built a Smith-style double auction adapted for AI agents: 11 buyers, 11 sellers, each with a private reservation price, trading through an order book that accepted only improving quotes. Every trader was an AI agent — drawn from a large OpenAI model; a smaller and cheaper sibling of the same family; and a large Google model — each run 10 times over. None of the three converged fully to the competitive equilibrium. Prices stayed well above what Smith’s humans had managed, and trading volume often fell short of the six trades the equilibrium implied. This wasn’t a failure of arithmetic; each agent was explicitly told to maximise profit and was given its own reservation price. The failure sat somewhere between understanding a market and knowing how to behave inside one. It was not the flagship model that came closest to functioning as a market. The larger, more capable OpenAI model produced the worst market of the three. Its agents kept improving their quotes in one cent increments and simply refused to close the gap, so trading volume fell to between a third and two-thirds of the equilibrium quantity — rent-seeking taken to a self-defeating extreme. The smaller, cheaper sibling behaved rather better. Its buyers and sellers leapt across meaningful price gaps rather than inching toward them and, round over round, its markets tightened in the same broad pattern Smith’s humans had shown, with efficiency climbing above 90 per cent by the final round. It behaved as though it were, in some rough sense, adapting to the structure of the game. The researchers are careful not to oversell this, since their design wasn’t meant to isolate the effect of model size, but it echoes other research suggesting that greater sophistication doesn’t necessarily make AI traders more human-like. One model provides reasoning traces as it trades, and combing through them shows a pattern: While an agent shaved cents off its offer, its language was dominated by words like ‘strategic’ and ‘optimise’ — the vocabulary of protecting a margin. The moment it leapt across the spread to trade, that vocabulary shifted toward ‘immediate’, ‘secure’, ‘guarantee’ — the vocabulary of urgency. The researchers caution that the stated reasoning need not match the computation actually driving it, but the shift suggests two instincts in tension: One protecting profit, and one eventually panicking about ending the round with nothing sold. AI agents are already being used for pricing, negotiation, bidding and financial tasks. The mechanisms within which they operate were built around assumptions about how human traders behave, and a mechanism that reliably squeezes efficiency out of self-interested humans may need different rules, or a human hand at the wheel, when the traders are language models instead. The same caution applies to research using language models as stand-ins for human subjects: Smith’s result is one of the most robust findings in experimental economics, and even here the language models didn’t reproduce it off the shelf. There’s a narrower implication, too. Some researchers read the repeated small price adjustments by AI agents as evidence of collusion. But calling a price “collusive” only makes sense against a baseline of what honest, non-colluding agents would produce, and the obvious baseline is the competitive equilibrium. If AI agents on their own can’t reliably find that equilibrium in the first place, a price sitting above it isn’t necessarily proof of a conspiracy. Whether AI agents can ever be trusted inside markets built for humans remains open. The institution that let a room of anonymous 1962 undergraduates turn scattered private information into a remarkably accurate public price does not, on the evidence here, produce the same result with machines. At least not yet, and not without knowing rather more about which machines, and why. Published on September 7, 2026