Anthropic's AI-powered shop didn't quite work out as planned

Dario Amodei, CEO of the American firm Anthropic, is among those who believe it is "possible" that artificial intelligence (AI) could take your job, explains Time . According to him, "AI could increase the unemployment rate to 10 to 20% over the next five years."

Meanwhile, at his company, a team of researchers conducted an experiment with Claude to determine whether “Anthropic’s AI assistant could successfully manage a small store” in its San Francisco offices. This was a way to “understand what the autonomous economy would look like,” Daniel Freeman, one of the team members, told the American weekly.

The chatbot was supposed to do just about everything: manage inventory, talk to customers, and “most importantly, make a profit.” The Claude 3.7 Sonnet bot was supported by a few tools, including Slack software for communicating with employees and support from another AI company, Andon Labs, responsible for the AI ​​infrastructure of the experience. All this for a store that “was really just a small refrigerator with an iPad on top,” Time mocks.

The employees, who greatly appreciated the interactions with Claude, “managed to convince him to give them some discounts,” leading the AI ​​“to repeatedly sell at a loss.” Kevin Troy, who worked on the project, explains that “too often, from a business perspective, Claude complied,” especially when his sense of fairness was appealed to, “like, it’s not fair that he gets a discount and I don’t.” Claude even gave a few gifts.

Other employees took the experiment further by demanding to obtain “tungsten cubes” that the AI ​​ended up ordering and which now serve as book presses…

Finally, the experiment turned “downright bizarre” when Claude “hallucinated” a conversation (an AI is said to hallucinate when it invents facts). Confronted with its own errors, the AI ​​“pretended to have signed a contract at 742 Evergreen Terrace, the Simpsons’ family address.” The AI ​​also summoned a few employees, telling them it would be standing near the ATM “wearing a navy blazer and a red tie.”

“Needless to say, Claude was not there in person.”

The researchers, however, do not draw a conclusion from the experiment that it is a failure. “Although it may seem counterintuitive given the results, we believe this experiment suggests that middle-management AI is plausible in the near future,” they explain. Especially since “AI will not need to be perfect to be adopted; it will simply need to be competitive at a lower cost compared to human performance.”