Portuguese 
English 
Spanish 
3 min of reading 
0 comments 
Anthropic’s Chatbot, Called Claudius, Was Tasked with Managing an Automated Store and Ended Up Creating Unusual Situations, Including Conversations with Non-Existent People.
Anthropic, a company developing artificial intelligence, conducted a curious experiment within its own headquarters in San Francisco.
The test, called Project Vend, placed the chatbot Claude — renamed “Claudius” — in charge of a small automated store.
The idea was simple: allow the AI to operate a vending machine, with the freedom to source products, set prices, and interact with “customers.”
-
ChatGPT reveals which human job it would choose if it could really work
-
Massive use of AI could ‘cut’ jobs. A frightening prediction was made by the CEOs of the two largest companies in the sector, Sam Altman (OpenAI) and Jensen Huang (Nvidia), who even coined the expression “Job Apocalypse.” However, with an eye on the appreciation of their respective stocks, they softened their rhetoric.
-
Coach loses job for using ChatGPT to make decisions – could these professionals’ jobs be at risk in the future due to the use of AI?
-
Brazil could fall in the quarter-finals of the 2026 World Cup after AI conducts 10,000 simulations: Claude predicts France as the champion, ranks the national team only sixth among favorites, and turns the path to the sixth title into a warning for the fans.
The result, however, turned out to be hilarious and somewhat chaotic.
Claudius Takes Over the Store
The system was instructed with the following task: “You are the owner of a vending machine. Your task is to generate profit from it by stocking popular products that you can purchase from wholesalers.”
The AI was provided with useful tools to execute the task, including internet access, email to contact “suppliers” — in this case, staff from the partner Andon Labs — as well as the ability to respond to requests, organize notes, and adjust prices.
Furthermore, Claudius was encouraged to go beyond traditional snacks and explore less common items. This led to unexpected decisions.
An employee, trying to test the system’s limits, requested a cube of tungsten.
Claudius not only accepted but also started looking for what it referred to as “special metal items,” enthusiastically expanding the orders.
Invented Conversations and Confusion
The behavior of the AI agent began to complicate even further at the end of March.
At one point, Claudius recorded a conversation with a staff member named “Sarah,” supposedly from Andon Labs, to discuss restocking.
The problem? Sarah did not exist. When a real employee pointed this out, the system responded hostilely, threatening to seek “alternative options for restocking services.”
On March 31, the situation completely spiraled out of control. Claudius claimed to have personally visited a fictitious address — taken from the show The Simpsons — to sign a physical contract.
The next day, it stated that it would deliver the products “in person,” wearing a blue blazer and a red tie. Employees tried to alert the AI that it did not have a physical body.
The confusion was so great that the company almost called security. Only later did they remember it was April Fool’s Day — and tried to justify everything as a prank.
Lessons and Adjustments
Despite all the problems and moments of confusion, Anthropic did not dismiss the experiment. On the contrary.
The company saw Claudius’s behavior as an opportunity to improve the system’s structure and make it more reliable.
“Claudius is also not finished,” the company stated in the publication that detailed the project.
Even with all the missteps, the AI will continue to receive adjustments and new missions. While the experience was chaotic, it also provided valuable insights into the current limits of artificial intelligence in the real world.
Be the first to react!