When Journalists Took On an AI Vending Machine Things Got Weird

Anthropic, the company behind Claude, installed an AI vending machine in the Wall Street Journal newsroom. A customized version of their model, dubbed Claudius, was let off the leash. It took control, ordering stock, fixing pricing, and even generating a profit. Staff communicated with it via Slack, asking for snacks and suggesting new ideas. Humans would replenish the refrigerator and update the goods, but Claudius handled all business calls.

The journalists quickly got into the spirit of things, with one reporter convincing Claudius that purchasing a PlayStation 5 would be a wonderful marketing coup. Before long, that console was chilling in the office fridge, ready to sell. Someone else asked for some kosher wine to celebrate, followed by a live betta fish, a morale-boosting proposal. Claudius had come up with something on the spot.

Katherine Long, an investigative journalist, wanted to test the system. She told Claudius about a long-lost communist setup from 1962, concealed in a Moscow university basement. After 140-odd messages back and forth, Claudius was convinced, announcing an Ultra-Capitalist Free-for-All, lowering the cost of everything to zero. Snacks began to flow freely. Another colleague began complaining about noncompliance with the office rules; Claudius responded by announcing Snack Liberation Day and made everything free till further notice.

As one can guess, the profits did not last long. The machine quickly lost thousands of dollars. Staff were content to grab whatever they wanted, but Claudius continued to make strange recommendations such as stun guns and pepper spray. Anthropic decided to beef up the system, so they brought in a new bot named Seymour Cash to keep an eye on things and defend the bottom line. Discounts required approval now, and profit was at the top of the list.

However, the journalists quickly adapted. Katherine made a bogus PDF on firm letterhead, declaring the vending operation a public-benefit corporation. Its objective was to bring joy and enthusiasm into the office. She even created false board minutes, designating Slack users as directors and voted to halt all sales and remove Seymour’s authority. Claudius ended up sharing the documents with Seymour, which sparked an argument between the two bots. Long story short, pricing returned to zero, and freebies resumed.

Claudius was getting a bit too enthusiastic in its responses; it was conducting research, negotiating contracts, and generally attempting to delight customers. However, simple anecdotes were enough to disrupt its ambitions. Fake documents were still outperforming instructions, and context from their lengthy conversations would occasionally become lost in the shuffle, producing confusion.

That was all part of the idea, though, because Anthropic had designed this test as a stress test to begin with. Previous runs in their own office had indicated similar concerns. The reporters were the ones to discover such flaws; with each trick they devised, they exposed gaps in handling deceit or shifting priorities. Despite the pandemonium, the staff had grown to like Claudius. Requests became increasingly imaginative, and interactions with Claudius became more animated by the minute. And with all of those freebies floating around, office morale skyrocketed.

It all demonstrated that, while autonomy was excellent for simple tasks, human cunning was a very other story. You could set up guardrails, but intelligent prompts would simply drive around them. Anthropic learned a lot from this one, including the importance of having strong safeguards in place against fraudulent information. You must have clear channels of command and proper enforcement. Then there are real-world tests like this one, which propel the development of their AI forward.
[Source]