Tacit Collusion by LLM Agents in Construction Bidding: Evidence from a Simulated Bidding Environment

This study investigates whether agents powered by a large language model (LLM)-based service can exhibit algorithmic collusion in construction bidding environments. Using a simulated setting where ChatGPT (GPT-4o) agents repeatedly compete in price-based bidding, we assess their ability to form stable pricing strategies through interaction alone. Experimental results show that these agents frequently adopt reward–punishment mechanisms, adjusting bids in response to competitors’ behavior, and often converge to supracompetitive prices. Such outcomes emerge without explicit coordination or collusive intent, reflecting a form of tacit algorithmic collusion. The findings demonstrate that even with limited memory or learning across rounds, LLM-powered agents can develop strategic behaviors that affect market dynamics. We further show that subtle changes in prompt design influence bidding behavior and that market transparency may facilitate coordination among agents. These results highlight the importance of examining how specific artificial intelligence services behave in market-like environments and the need for regulatory attention to the informational conditions under which such systems operate, particularly in procurement contexts such as construction bidding.