Original title: Competitive Market Behavior of LLMs
Pawel Struski, Jakub Swistak, Inez Okulska, Przemyslaw Biecek · September 2026
People are starting to use large language models as economic agents to buy and sell things. But we do not know if these AI agents work well in trading systems built for humans. To find out, researchers took classic economics experiments and replaced human participants with AI agents. They set them up in a double auction, a market setup where buyers and sellers constantly propose prices to each other. The researchers wanted to see if the market reached fair, efficient prices. They found that markets of AI agents converge toward equilibrium either more slowly than human markets or not at all, making the markets less efficient. Furthermore, looking at the models' written reasoning showed that agents switch from strategic thinking to feeling urgency when they decide to finalize a trade.
Imagine a bustling farmers' market where vendors and shoppers haggle over prices. Normally, through bargaining, they find a price where everyone is satisfied and goods are sold. But if the vendors are robots following strict rules, they might stick to their initial prices too long, causing unsold produce or empty stalls. In this paper, the LLM agents act like those robots, failing to adjust prices dynamically, so the market never clears efficiently.
If AI agents are used in financial systems, their failure to trade efficiently could lead to economic losses. This study shows that current LLMs don't behave like human traders, which matters for designing AI-driven markets.
The abstract doesn't mention testing in market types beyond double auctions or scaling to larger populations. It also doesn't discuss long-term adaptation or comparison with other AI strategies.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Competitive Market Behavior of LLMs — Pawel Struski, Jakub Swistak, Inez Okulska et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.