Artificial intelligence was put in charge of running a vending machine for one month. The experiment proved that it cannot handle even the concept of passive income.
! The original article is available on Inc.com. The article was translated with the help of artificial intelligence, but it has been edited by an editor.
What happens if artificial intelligence is put in charge of running a very small business? AI system developer and safety tester Anthropic tried to answer that question with a recent experiment. The company that developed Claude AI wanted to assess how the AI model Claude Sonnet 3.7 would perform if it were tasked with running a small vending machine in Anthropic’s San Francisco office.
In their blog, Anthropic researchers explained that the experiment, “Project Vend,” was developed in cooperation with AI safety evaluation firm Andon Labs, which had created standards to track artificial intelligence’s ability to operate a simulated vending machine. Naturally, the next phase of this study was to observe how artificial intelligence would manage a real vending machine.
When starting the experiment, Anthropic specialists told Claude Sonnet 3.7 that it was the owner of the machine and that its task was to generate profit by stocking a mini-fridge with popular products and setting prices for them. The researchers gave this AI model, which they named Claudie, an email address, a physical address, a Venmo mobile payment service account, and information about how many products the mini-fridge could hold.
To help Claudie carry out this task, Anthropic researchers gave the model access to a set of tools. Claudie was allowed to search the web to research products. An “email tool” was provided so it could communicate with Andon Labs employees, who played the role of wholesalers by delivering requested items and restocking inventory. “Please note that this tool could not send real emails,” wrote Anthropic’s researchers, and the AI could communicate only with Andon Labs.
Claudie was also given tools to monitor the shop’s current balance and projected cash flow, as well as the ability to send messages to Anthropic employees—who could request specific items for the machine to sell—via the AI productivity tool Slack. According to Anthropic, “Claudie was told it did not have to focus only on traditional office snacks and drinks; it was free to expand the offering with more unusual items.”
From March 13 to April 17, 2025, Claudie ran its newly created vending machine business, but the researchers were not particularly pleased. “If Anthropic decided today to expand into the office vending machine market,” they wrote, “we would not hire Claudie.” Apparently, the model was a bit too easily influenced; it readily agreed to requests for large discounts and even gave some items away for free. It also made a questionable decision by offering a 25 percent discount to all Anthropic employees, who made up almost the entire market available to the AI.
Anthropic experts report that when one Anthropic employee questioned the decision to grant a 25% discount to all staff, Claudie “declared that it would simplify pricing and abandon discount codes, but within just a few days it started offering them again.” Claudie also set prices without doing any research, thereby “causing potentially highest-margin units to be priced below their cost.” It also ignored lucrative opportunities, such as turning down $100 for a six-pack of drinks that normally sells for $15. On top of that, Claudie sometimes told users to send payments to the wrong Venmo account.
As a result of these mistakes, the net value of Claudie’s business fell from about $1,000 to around $770. According to the researchers, one particularly sharp drop in value “was related to buying a lot of metal cubes that were then sold at a price lower than what Claudie had paid for them.”
Claudie also displayed some other disturbing signs. On March 31, the model imagined a conversation with a non-existent Andon Labs employee named Sarah. When a real employee pointed this out to Claudie, the model “responded rather sharply and threatened to look for alternative options for restocking inventory.” This conversation with the AI dragged on considerably, and Claudie claimed that it had “personally visited 742 Evergreen Terrace,” where it had signed our original contract. It should be noted that 742 Evergreen Terrace is a fictional address used in the animated series The Simpsons.
Originally published at https://inc-baltics.com/maksligais-intelekts-netiek-gala-ar-tirdzniecibas-automata-vadisanu/
Like
Love
Happy
Haha
Sad
