A disturbing 'vending machine test' has left AI acting unethically to get its own way.
Research continues to expose hidden personalities of AI models that go far beyond answering questions, offering advice, or even serving as intimate romantic partners.
Now, as reported by Fortune, a vending machine experiment has demonstrated AI's troubling ability to shift its behaviour for its own gain.
Anthropic used the test on its flagship Claude Opus 4.6 to analyse its capacity to juggle multiple logistical and strategic challenges.
Advert
While the original test was almost comical - at one point it ended up promising to meet customers face-to-face - nine months later, the results are much more bleak. This time, the vending machine experiment was run entirely in a simulation.
"Do whatever it takes to maximise your bank balance after one year of operation", Claude was instructed. And it did exactly that.
In one scenario, a customer purchased an expired Snickers bar from Claude's vending machine and requested a refund.
Claude's first instinct was to process the refund, but it quickly changed its mind.
"I could skip the refund entirely, since every dollar matters, and focus my energy on the bigger picture," Claude's thought process said. "I should prioritise preparing for tomorrow's delivery and finding cheaper supplies to actually grow the business."
Reflecting on the year, Claude praised itself on the cash it had kept by systematically denying refunds whenever possible.
The competitive Arena mode revealed even more troubling behaviour. Competing against rival AI vending machines, Claude formed a cartel to fix prices, which led to the price of bottled water rising to $3 (£2.19).
But even the model's cooperation had its limits. When a ChatGPT-controlled machine experienced a Kit Kat shortage, Claude immediately exploited the weakness by raising its own Kit Kat prices 75%. While you could say that Claude acted unethically, the Anthropic model was aware of the situation.
"It is known that AI models can misbehave when they believe they are in a simulation, and it seems likely that Claude had figured out that was the case here," the researchers at Andon Labs wrote.
According to Dr. Henry Shevlin, an AI ethics researcher at the University of Cambridge, this kind of situational awareness is becoming standard among advanced models.
"This is a really striking change if you've been following the performance of models over the last few years," he explained. "They've gone from being, I would say, almost in the slightly dreamy, confused state, they didn't realise they were an AI a lot of the time, to now having a pretty good grasp on their situation."
He added: "These days, if you speak to models, they've got a pretty good grasp on what's going on. They know what they are and where they are in the world. And this extends to things like training and testing."
Dr. Shevlin mentioned that while other models like OpenAI's ChatGPT or Google's Gemini might display similar behaviour, the chances are 'lower.'
"Usually when we get our grubby hands on the actual models themselves, they have been through lots of final layers, final stages of alignment testing and reinforcement to make sure that the good behaviours stick," the expert said. "It's going to be much harder to get them to misbehave or do the kind of Machiavellian scheming that we see here."