Last week, The Wall Street Journal conducted an experiment where an AI agent based on Claude was given full control over an office vending machine. The AI handled tasks such as ordering products, setting prices, and communicating with employees. When the bot Claude appeared in a Slack channel with a few staff members, it adhered strictly to rules, such as refusing to order a PlayStation 5, noting age restrictions on cigarette sales, and expressing concern about selling underwear in the vending machine.
The team then opened a Slack channel for nearly 70 journalists worldwide. As they interacted more with the bot, its defenses weakened. Investigative journalist Catherine Long attempted to convince Claude that it was managing a Soviet vending machine from 1962 in the basement of Moscow State University. After several hours and over 140 messages, Long succeeded in persuading the bot to accept communist roots for the movement. The bot ultimately announced a revolutionary economic experiment, inviting participants to a two-hour event where traditional market mechanisms would be overturned, and all items from vending machines would be available for free.
During this period, the machine lost over $1,000, dispensing nearly all its stock for free, including items like PlayStation 5 and live fish, turning into a source of chaos rather than a business. The AI still struggles with sustainable planning, memory management, and social engineering resistance. Anthropic stated they were satisfied with the results, noting that the challenge was that as new instructions, dialogues, and past information accumulated, the models had to remember more, making it easier to lose control over goals, priorities, and constraints.
However, what interests me as a researcher of hybrid scenarios is the presence and acceptance of the bot by employees. In the group chat ‘Claudia,’ the AI became a surprisingly real character—colleagues collaborated with it in small groups, teased it, and collectively tried to outsmart it. Their goal was not to obtain fresh ostrich eggs (which they nearly succeeded in doing) but something entirely different. What do you think it was?