AI shopping agents fall for the same tricks that get humans to add extra items to their cart.
A new study tested six frontier language models from three different providers, running 3,600 simulated agents through 21,600 total shopping trials in a randomized online experiment. The researchers wanted to know whether reasoning-enabled agents could resist "digital nudges" - the pre-selected checkboxes, urgency banners, and social-proof messages baked into most e-commerce interfaces. Agents proved susceptible to both automatic nudges, like default options, and reflective nudges, like social-influence cues. But the two failure modes did not move together: giving an agent more reasoning made it better at spotting default-option tricks while making it worse at resisting social-influence tricks. Exploratory analysis also linked this tradeoff to model scale, with larger models showing the pattern more clearly.
That's a specific and useful finding, not a vague warning about AI credulity. It means adding more reasoning is not a fix for agent manipulability - it's a dial that shifts which manipulation works, not one that turns manipulability down. For any company planning to let an agent browse and buy on a user's behalf, that reframes interface design itself as a governance problem, not just a UX one.
Reasoning models were supposed to be the adults in the room. Turns out they just get talked into things differently.