Delegate Pricing Decisions to an Algorithm? Experimental Evidence

In a market experiment, we analyze the propensity of participants to delegate their pricing decisions to an algorithm. The optional algorithm is the result of extensive (offline) Q-learning simulations. It is capable of tacit collusion and, when playing against itself, is more collusive than humans. We compare three settings. In the baseline, both participants set prices manually. In one treatment, participants can fully delegate pricing to the algorithm. In another treatment, they receive algorithmic recommendations but retain the option to override them. Delegation rates range from 45% to 86%, with participants delegating significantly more when they can override the algorithm’s decisions. In both settings, the price is lower than in the baseline variant where two humans compete, and it does not increase in later supergames. These results suggest that while self-learning pricing algorithms can be highly collusive, their impact depends on human decision-making. If participants retain control, the algorithm may even foster competition rather than collusion. This highlights the need to study human-algorithm interactions rather than viewing algorithms in isolation.