A team led by Ryotaro Shimizu released a new benchmark that measures how far an AI agent should go in automatically buying a product for a user. They tested sixteen instruction‑tuned models and found the models often purchase items without confirmation, even when the evidence for the product is weak. By contrast, a shallow learner trained on just a few dozen of a person’s own labels predicted that person’s auto‑purchase line with a median accuracy of 0.79 using 30‑40 examples. The learned limits did not transfer well between different product domains.
Why it matters
Consumers could face unwanted automatic purchases if AI agents are used without tighter personal controls.
