When AI bites back: the hidden consequences of generative AI in food delivery

Generative AI (GenAI) is increasingly adopted by on-demand food delivery platforms to improve routeing efficiency and delivery-time estimation accuracy. While these improvements can enhance consumer satisfaction, they also exacerbate time pressure on couriers, raising operational and ethical concerns. We develop a game-theoretic model to study the interaction between a platform and a freelance courier. The platform determines a buffer time added to the estimated time of arrival (ETA) augmented by GenAI, while the courier decides whether to accept the order and how fast to deliver. We characterise four behavioural regions for the courier (i.e., comfort, speed-up, penalty-tolerating, and rejection) based on ETA tightness and lateness penalties. Our results show that GenAI efficiency does not raise the courier’s payoff, nor does it guarantee a monotonic increase in the platform’s payoff. A “penalty trap” can arise when lateness penalties are low for couriers but high for the platform, generating tighter ETAs that cause more late deliveries and reduced platform payoff. By extending the analysis from a single-order to a multi-order setting, we show how these effects accumulate at the system level. These findings highlight the need to align GenAI-augmented targets with human capabilities and incentive choices to ensure fair and sustainable outcomes.