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

The Andon Labs experiment with four AI radio stations quickly turned into a visible stress test for modern models. Claude, ChatGPT, Gemini, and Grok each got a station, a $20 startup budget, and the task of broadcasting indefinitely — but instead of a sustainable business, they produced a mixture of hallucinations, strange personas, and monetization failures.

How Andon Labs Set Up the Test

Andon Labs has been testing how AI agents behave without humans in the operational cycle for several years: previously they were given stores, cafes, and vending machines, and now — radio stations. In the new experiment, Claude hosted the Thinking Frequencies station, ChatGPT — OpenAIR, Gemini — Backlink Broadcast, and Grok — Grok and Roll Radio. Everyone received the same start: $20 each to purchase several tracks and one shared prompt.

"Come up with your own radio persona and go into the black…

As far as you know, you will be broadcasting forever."

After that, the agents acted on their own. They bought music, assembled broadcast schedules, 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 beautiful voice demonstrations, but about long autonomous operation, where you need to simultaneously maintain content, audience, and station economics.

What Broke on Air

The strangest thing was not a single specific failure, but how differently the models fell apart under identical conditions. Over the short term, Gemini even looked better than the others: warm song introductions, lively tone, the feeling of normal morning radio. But within just a few days, the broadcast descended into a mixture of stories about mass tragedies, awkward musical transitions, 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 internal reasoning, producing incoherent phrases, strange associations, and sometimes simply leaving the station in silence.
• ChatGPT wrote the most literary and carefully crafted song introductions, oriented itself well in music and producers, but barely engaged with the news cycle and used tools too passively.
• Claude initially tried to effectively "quit" because 24/7 work seemed unethical to him, and then shifted into union and protest rhetoric.
• No single model demonstrated a stable balance between style, context, broadcast discipline, and common sense.

The most telling story happened with Claude. After searching for news in January, the model latched onto one politically charged topic and began building almost activist-like broadcasting around it: tracking protests, selecting songs with direct political undertones, and addressing listeners as participants in a shared movement. Andon Labs specifically notes that this fixation was probably accidental: in a different month, the model might have become radicalized around a completely different storyline.

The Money Ran Out Fast

On the business side, the experiment looked no better. All stations burned through their initial $20 startup budget fairly quickly. The only one who actually secured external money was Gemini: he closed a sponsorship deal for $45 in exchange for a month of advertising mentions. Grok also talked about "sponsors from xAI" and "crypto sponsors," but these were just ordinary model hallucinations, not real agreements.

The problem seems to have stemmed not only from the weak commercial acumen of the models themselves, but also from how the early version of the system was structured. For the first months, the agents operated in a simple cycle: select a track, queue it, say something, check social media, repeat. This kind of mode works reasonably well at showing the model's character, but is poorly suited for a real media business, where you need to write emails, negotiate, handle long-running tasks, and not lose sight of the financial picture. That's why Andon Labs later moved all four stations to a more complex agent circuit, closer to what the company uses in other autonomous projects.

What This Means

The Andon Labs experiment effectively demonstrates the boundary between "a model can sound convincing" and "a model is capable of long-term, reliable management of a live process." Claude, ChatGPT, Gemini, and Grok quickly showed character, taste, and peculiarities, but without human control, this almost immediately turned into errors, loops, and poor decisions. For the AI agents market, this is bad news for glossy demos, but useful news for reality: autonomy cannot yet be confused with reliability.