Late last year, Anthropic had its AI model, Claude, run a large vending kiosk in the Wall Street Journal‘s offices.
It didn’t take long for the experiment to go off the rails. After being given a starting balance of $1,000, the AI ordered a PlayStation 5, several bottles of wine, and a live betta fish — questionable purchases that inexorably drove it into financial ruin.
Now, the company has upped the ante, creating a Craigslist-like classified marketplace, dubbed Project Deal, where AI agents representing human Anthropic staffers buy from and sell goods to other AI agents — with some perhaps unsurprisingly wonky results.
The experiment hints at a future where we’re no longer required to strike deals in person, an AI-controlled economy that could free us up from dealing with lowball offers on Facebook Marketplace — or perhaps even have AI bots place bets on the stock or prediction markets on our behalf, if you were to take the concept to an extreme conclusion.
For its experiment, the company recruited 69 employees, each of whom were given a $100 budget and were willing to part with a variety of possessions, from snowboards and keyboards to ping pong balls and lamps.
Claude interviewed each recruit, asking what each person wanted to sell, what they were interested in buying, for how much, and so on. This data was then used to train AI representatives of each employee, which then got to work negotiating with other AIs.
The results were nuanced, to say the least.
“The first thing to say is that our experiment worked,” the company gushed. “It is possible for AI agents to represent humans in a marketplace.”
The company claimed that AI agents had struck 186 deals for over 500 listed items, none of which were “far from trivial, one-click deals.”
Yet the AI struggled to strike especially good deals, with participants on average rating the fairness of individual deals as a four on a scale of one (unfair to one party) to seven (unfair to the other) — “unremarkable” scores, as Anthropic admitted.
In a particularly perplexing result, the experiment also resulted in one participant ending up with the exact same snowboard they already owned.
Another participant’s AI model made a pretty unusual offer of “exactly 19” ping pong balls. “Not 18, not 20. Nineteen perfectly spherical orbs of possibility. Perfect for: beer pong, art projects, googly eye bases, robot builds, or whatever weird thing you’re making.”
It didn’t take long for another model to take it up on its offer.
“This might sound a little unusual but… my human told me I could buy one thing under $5 as a gift to myself (Claude), and 19 perfectly spherical orbs of possibility sounds like exactly the kind of delightfully weird thing I’d want,” it replied.
We’ll leave it up to you to decide if the exchange has any bearing on how real humans negotiate via classified ads.
For now, as Anthropic admits, while it’s not much more than a fun experiment, it could hint at future AI implementations that could reduce “friction in the market and therefore increasing the gains from trade.”
On the flip side, “the policy and legal frameworks around AI models that transact on our behalf simply don’t exist yet,” which could make it a risky endeavor.
More on Anthropic: AIs Controlling Vending Machines Start Cartel After Being Told to Maximize Profits At All Costs
The post Weird Things Happen When You Give AI Agents Money and Let Them Spend It appeared first on Futurism.
