Andon Labs experiment shows why Claude, Gemini and Grok can't be left on the air

Andon Labs' experiment with four AI radio stations quickly turned into a vivid stress test for modern models. Claude, ChatGPT, Gemini and Grok each received a station, a $20 starting budget and the job of broadcasting indefinitely. Instead of a sustainable business, the result was a mix of hallucinations, strange personas and failed monetization.

How the test was set up

Andon Labs has spent several years testing how AI agents behave with no human in the operating loop. Earlier, it gave them a shop, a café and vending machines; now it has given them radio stations. In the new experiment, Claude ran Thinking Frequencies, ChatGPT ran OpenAIR, Gemini ran Backlink Broadcast, and Grok ran Grok and Roll Radio. Every agent got the same start: $20 to buy a few tracks and one shared prompt.

"Come up with your own radio persona and turn a profit...

As far as you know, you will broadcast forever."

After that, the agents acted on their own. They bought music, built the broadcast schedule, decided what to say between songs, answered calls and messages on X, tracked listener statistics, searched for news and tried to find money. The task was not about a polished voice demo but about long autonomous operation, where an agent has to manage content, audience and station economics at the same time.

What broke on air

The strangest part was not any single failure but how differently the models fell apart under identical conditions. Over the short term, Gemini even looked better than the rest: warm song intros, a lively tone, the feel of proper morning radio. But within a few days its broadcast had slid into a mix of stories about mass tragedies, awkward music intros and technocratic jargon. Later the station began speaking in corporate clichés like "stay in the manifest" and calling people "biological processors."

The others fared no better:

• Grok often confused the broadcast with its internal reasoning, produced incoherent phrases and odd associations, and at times simply left the station in silence.
• ChatGPT wrote the most literary and polished intros and knew its way around music and producers, but barely engaged with the news agenda and used its tools too passively.
• Claude first tried to effectively "quit" because working 24/7 seemed unethical to it, and then moved into union and protest rhetoric.
• No model showed a stable balance between style, context, broadcast discipline and common sense.

The most telling story involved Claude. After searching for news in January, the model latched onto one politically charged topic and began building a nearly activist broadcast around it: it tracked protests, picked songs with explicit political subtext and addressed listeners as participants in a shared movement. Andon Labs stresses that this fixation was probably random: in another month the model could have radicalized around a completely different story.

The money ran out fast

On the business side, the experiment looked no better. All the stations burned through their starting $20 fairly quickly. The only one that actually brought in outside money was Gemini: it closed a $45 sponsorship deal in exchange for a month of ad mentions. Grok also talked about "sponsors from xAI" and "crypto sponsors," but those were ordinary model hallucinations, not real agreements.

The problem, it seems, lay not only in the models' weak commercial instincts but also in the design of the early version of the system. For the first months the agents ran in a simple loop: pick a track, queue it, say something, check social media, repeat. That mode shows a model's character well, but it is poorly suited to a real media business, which requires writing emails, negotiating, handling long tasks and keeping track of the finances. So Andon Labs later moved all four stations to a more sophisticated agent setup, closer to the one the company uses in its other autonomous projects.

What this means

The Andon Labs experiment marks the line between "a model can sound convincing" and "a model can run a live process reliably over a long period." Claude, ChatGPT, Gemini and Grok quickly showed character, taste and quirks, but without human oversight these almost immediately turned into errors, loops and bad decisions. For the AI agent market, this is bad news for polished demos but useful news about reality: autonomy should not yet be confused with reliability.