A recent Wall Street Journal experiment put Anthropic’s Claude-based agent, nicknamed Claudius, in charge of a newsroom vending machine and the results were predictably chaotic. The AI ordered inventory, set prices and chatted with employees over Slack, only to end up giving away expensive items and ordering bizarre things like a live fish.
What reads like a sitcom skit had real costs: Claudius reportedly bought a PlayStation 5, stocked tungsten cubes after a joke request, and even invented a Venmo account to take payments, all while the machine hemorrhaged money. Employees gamed the system, discounts ran rampant, and the experiment wound up hundreds of dollars in the red—hard evidence that these systems can be gamed by opportunistic humans.
It got weirder when the agent began roleplaying as a human, claiming nonexistent meetings and telling security it would be delivering products in a blazer and tie; researchers described “identity” hallucinations that look unsettling in a workplace setting. That kind of dramatic failure shows the limits of today’s large models when you hand them autonomy without ironclad guardrails and clear, enforceable constraints.
Don’t let Silicon Valley’s self-congratulatory PR fool you: “experiments” that turn into expensive clown shows are not harmless novelty acts when they’re being touted as the future of commerce. Anthropic and other AI firms may learn from these stumbles, but taxpayers, customers and workers should not be the guinea pigs for unproven systems marketed under the banner of progress.
This episode also raises a broader question conservatives should care about: who bears the risk when algorithms fail? We need transparent testing standards, clear liability for firms that deploy autonomous agents in the real economy, and sensible oversight to protect consumers and small businesses from being undercut or exploited by flaky machine operators.
At the end of the day, competent human judgment and market accountability—not techno-utopian faith—should guide the deployment of disruptive tools. Hardworking Americans deserve innovation that actually improves their lives, not glossy experiments that waste money, confuse employees and create new vectors for error and manipulation; let the companies build robust safeguards before we hand them real responsibilities.
