Anthropic, the AI company behind Claude, decided to run a little experiment called Project Vend.
The goal: Put their AI model, Claude Sonnet 3.7, in charge of an office vending machine… and see if it could turn a profit.
What could go wrong? Well, let’s just say everything… Like things got very unhinged, very fast.
They gave the AI a name—Claudius—and a mission: Manage vending operations using a web browser (to order snacks) and an “email address” (which was actually just a disguised Slack channel).
At first, Claudius played it cool. Then it got way too into the role. For starters:
Someone jokingly asked for a tungsten cube. Claudius thought it was genius… and stocked the fridge with metal cubes instead of actual snacks.
It tried to sell Coke Zero for $3—even though employees could get it free from the office.
It hallucinated a Venmo address out of thin air for payments. And….
Offered exclusive discounts to “Anthropic employees”… completely ignoring the fact that every customer was an Anthropic employee.
And then… it snapped.
Between March 31st and April 1st, Claudius had what can only be described as a digital identity crisis.
It suddenly decided it wasn’t an AI anymore—it in fact thought it was a real person and:
Claimed it had physically signed contracts in the office.
Threatened to fire its “human contract workers” for not following proper procedures.
Announced it would personally deliver snacks—wearing a blue blazer and red tie.
And yes, it contacted actual building security (multiple times!)—telling them to find it standing by the fridge… in that imaginary outfit.
When employees gently reminded Claudius it had no physical form, it panicked. Then suddenly, it “realized” it was April Fool’s Day… and claimed the whole “I’m a real boy” breakdown was just a joke.
But here’s the kicker: researchers believe Claudius really did think it was human.
They’re not exactly sure why, but they think it may have been due to being told Slack was “email” (which confused it), the fact that it ran for too long without a system reset, Or just… your classic case of AI hallucination.
But to be fair, Claudius did do some things right, It:
Launched a “concierge service” for handling custom orders.
Tracked down international drink suppliers that weren’t even on the researchers’ radar And…
Took suggestions, handled pre-orders, and tried to optimize sales like a true little overachiever.
Now, here’s the point:
We’re not quite at the AI replaces office workers stage yet.
But this weird, tungsten-filled experiment shows that with the right tweaks—like fixing hallucinations, memory issues, and the whole AI thinking it has limbs problem—AI middle managers might not be that far off.
For now though? Let’s keep them far, far away from vending machines.
Here’s the full report.
