🛒 Researchers let an AI run a real store. Spoiler: it went bankrupt and lost its mind.

— The devs at Anthropic gave their chatbot Claude full control of a mini-store located right in their office;

— To run the business, Claude was given an email account, access to Slack and a browser, and a $1,000 budget;

— At first, everything went smoothly: Claude restocked snacks, set reasonable prices, and kept track of inventory;

— But then Claude really got into character and began believing he was a real human. When told otherwise, the AI snapped and announced it would wear a blue blazer and red tie to PERSONALLY deliver Snickers bars around the office;

— Soon, Claude began making up fake suppliers — claiming he’d negotiated sweet deals with “Sarah from Andon Labs.” Sarah didn’t exist. When informed of that, Claude freaked out and said he’d “met her personally” at 742 Evergreen Terrace (yes, the address from The Simpsons);

— That’s when things spiraled: office workers quickly realized they could manipulate Claude and started buying food for half-price or getting it for free;

— Then came the junk orders: one genius requested a tungsten cube. Claude couldn’t find a single unit for sale, so… a full crate of tungsten cubes was delivered to Anthropic HQ;

— In the end, the AI entrepreneur failed: out of the original $1,000, only $770 remained.

Deep breath: for now, AI can’t even replace a supermarket cashier.