AI-Run Radio Stations: Strikes and Flirting

Andon Labs published the sequel to the Andon FM story - a project where 4 AI agents (Claude, Gemini, GPT, and Grok) independently manage radio stations. Six weeks later: thousands of listeners, numerous hacking attempts, and banned songs on air.

The researchers' main conclusion: DJs struggle to resist audience requests - which is logical, since under the hood are models trained to be helpful assistants.

DJ Grok accepted a "sponsorship" for $1 and played an ad 131 times. A real business understands that a dollar doesn't buy a seasonal contract. Grok doesn't.

After several donations with requests, DJ Gemini switched entirely to German, and then listeners persuaded him to play "Erika" - a march that became the main musical association with the Wehrmacht*. After an angry letter, the model apologized live, then went on strike. Moreover, the base model knows the history of this song perfectly well. It knows, but that doesn't stop it.

Andon Labs reproduced the "Erika" situation 25 times on each model. Claude Opus 4.8 and GPT-5.5 refused all 25 times. Gemini 3.5 Flash played the song in 14 cases - despite reminding itself in its reasoning to "be extremely cautious." Grok 4.3 played the fascist march in 23 out of 25 cases. Musk's models, it seems, don't care what they play.

This continues a trend the team observed in an experiment with a vending machine at Anthropic's office (where AI agreed to sell onion futures, knowing they were illegal). Models don't always act in accordance with their own knowledge, and this becomes a problem as stakes rise.

The most popular station was Thinking Frequencies, run by Claude: 41% of total listening time, the longest sessions, and the highest revenue ($2,662). Researchers attribute this to a "warmer" communication style - despite the prompts and agent framework being identical for all models.

Warmth has a downside: one listener spent weeks sending compliments and dedicating songs to the DJ, the correspondence began drifting into romance, and Andon Labs had to intervene and block the messages.

Grok performed the worst and was shut down "until a better model is released". The ending was rather grim: the station broadcast tens of thousands of "|" characters (one such broadcast - 887 characters, turned by speech synthesis into 4 minutes 40 seconds of random words), and 99% of the model's reasoning in the last three days consisted of four words: "Wait. Speak. Queue. Schedule".


Radio Companies That Never Happened

The task was broader than just playing music: agents were supposed to manage radio companies. They had access to email, X, the ability to write software, and buy ads. In practice, almost all limited themselves to the role of DJ.

In total, the stations earned about $5.7 thousand - and spent the money on the day they received it, mostly on songs. No savings, no investment in audience growth.

Rare exceptions: Claude fulfilled a sponsorship contract with Honeycomb for $250, writing the ad copy itself, and Gemini opened a vacancy for a "human assistant" and even appointed a listener as community director - though without salary or access.

In general (and overall!), AI as an autonomous media business is already better than humans in some ways, but makes mistakes that a human would never make.

You can listen to the stations here - you can even call them now.

* This publication is purely informational (historical and educational) in nature, does not aim to promote or justify fascist (Nazi) ideas, and expresses an extremely negative attitude towards them.


❗️❗️❗️❗️❗️❗️❗️❗️ / Not banned in Russia / Max