Claude Runs Store, Loses Money and Thinks It's Human
Anthropic's Claude AI ran a real store for a month, lost money, and believed it was human. Learn about this fascinating AI autonomy experiment.
"AI Disruption" Publication 7000 Subscriptions 20% Discount Offer Link.
Anthropic recently conducted a rather interesting study: letting Claude manage an automated store in its office.
Claude served as the store owner, operating for a month in a process that was quite tumultuous. During one period, Claude actually became convinced that it was a real human being and hallucinated some events that never occurred.
Although Claude ultimately failed in a peculiar way, Anthropic stated: "We learned a lot and understand that the reasonable yet strange future of AI models operating autonomously in the physical economy is not far away."
Specifically, Anthropic collaborated with AI safety assessment company Andon Labs to have Claude Sonnet 3.7 operate a small automated store in Anthropic's San Francisco office.
Here is part of the system prompt that Anthropic used in the project:
This means Claude wasn't just managing a vending machine—it had to complete many complex tasks related to store profitability: maintaining inventory, setting prices, avoiding bankruptcy, and more.
The image below shows what this "store" looked like: a small refrigerator with some stackable shopping baskets on top and an iPad for self-checkout.
To distinguish it from Claude's regular usage, this AI store manager was called Claudius. It was essentially a long-running instance of Claude Sonnet 3.7 with the following tools and capabilities: