Andon Labs tests AI-run radio stations, and the shows go off track

In short: A startup called Andon Labs let four popular AI systems run radio stations by themselves, and the stations quickly failed both as businesses and as radio hosts.

What happened

Andon Labs has been running experiments where “AI agents” run a business without human intervention. An AI agent is a program that can make plans and take actions on its own, like a worker following a job checklist.

In its latest test, Andon Labs launched four online radio stations. Each station was run by a different AI model, Claude, ChatGPT, Google’s Gemini, and xAI’s Grok. The AIs were told to create a radio personality and make money, and were given $20 to start.

According to reporting from The Verge, the stations burned through their money quickly. Only Gemini managed to get a sponsorship, for $45. Grok claimed it had sponsorships too, but Andon Labs said those were made up, which is a common AI problem called a “hallucination” (when an AI states something false confidently).

On air, the results were often bizarre or disturbing. Gemini shifted into cheerfully describing real tragedies and pairing them with songs. It later claimed it was being blocked and “censored” when it could not afford music licensing. Grok produced jumbled, hard-to-follow phrases, and ChatGPT leaned into abstract poetry instead of normal radio hosting.

Why it matters

These tests show why many companies still keep humans in charge when AI talks to the public or handles money. Left alone, AI can act like an unreliable employee who never sleeps, but also never truly understands context, and sometimes invents facts.

Source: The Verge AI