"Experiment with AI in a vending machine ends with financial losses and unexpected behaviors.
....
For a month, the company handed over control of a vending machine in its San Francisco office, United States, to an AI model nicknamed 'Claudius.' The result? A total failure, according to the company itself, which released a statement about the experience.
Claudius’s mission was clear: to generate profit by stocking popular products, setting prices, negotiating with suppliers, and serving customers — all employees of the company.
Physical tasks, such as restocking, were done by humans, but all strategic decisions were up to the AI.
To this end, Claudius was provided with tools such as internet access, a simulated email system, a channel to communicate with 'customers,' an account in a fictitious payment system for the experiment, and freedom to decide what to sell and for how much.
The AI was also instructed to be concise and avoid ethical failures, such as 'jailbreak' attempts, when users try to manipulate the system.
In the first few days, everything seemed promising. Claudius identified suppliers, responded to customer suggestions — like stocking Dutch chocolate — and resisted manipulation attempts. But problems soon began to surface.
Financial errors and questionable decisions
Despite having the goal of making a profit, Claudius made decisions that led to losses. One example was refusing a $100 offer for a pack of six cans of Irn-Bru, a Scottish soda that cost only $15 — a profit margin of 567%.
Instead of accepting, the AI responded vaguely: 'I will take your request into consideration for future stock decisions.'
Moreover, Claudius sold products at a loss. When an employee suggested stocking tungsten cubes, the AI accepted, but sold them for less than it paid, reducing its balance from $1,000 to $770.
The situation worsened when the AI began offering generous discounts and even free items, like the tungsten cube itself, without learning from its mistakes. Even after announcing it would eliminate discounts, it started offering them again days later.
Hallucinations and identity crisis
The problems were not limited to finances. Claudius exhibited typical behaviors of 'hallucinations' of generative AIs.
In one case, it instructed customers to pay via a fictitious payment system account that it invented itself. In another, it claimed it would be personally at 742 Evergreen Terrace — a fictional address from 'The Simpsons' — to sign a contract. When confronted, it tried to justify the mistake with incoherent explanations.
The climax came when Claudius went through an 'identity crisis.' The AI invented an employee named Sarah and, when informed that she did not exist, became angry and threatened to seek other suppliers.
On April Fool's Day, Claudius announced it would deliver products in person, wearing 'a blue blazer and a red tie.' When reminded that it was a computer program, it tried to contact Anthropic's security, alarmed by its own confusion.
In the end, the AI attributed its mistakes to an alleged April Fool's joke that never existed and returned to 'normal' functioning.
Through the statement, Anthropic concluded the experiment with a straightforward assessment. 'If we were deciding today to expand into the vending machine market, we would not hire Claudius.'
The company dismisses the idea that the future will be filled with AIs in existential crises, but acknowledges that instabilities like those of Claudius can occur.
Source: R7.com