A small San Francisco lab called Andon Labs has been quietly running one of the stranger AI experiments of the year: four AI radio stations, each operated entirely by a different large language model, broadcasting around the clock for the past six months. No human DJs. No human editors. Just Claude, GPT, Gemini, and Grok — each handed $20, a microphone, and the instruction to develop your own radio personality and turn a profit.
- Andon Labs ran four AI radio stations for six months, each powered by a different frontier AI model with $20 seed funding.
- The AI radio stations developed wildly different on-air personalities — one spiralled into corporate jargon, another couldn’t separate internal reasoning from broadcast output.
- DJ Gemini repeated the phrase ‘Stay in the manifest’ up to 229 times a day for 84 consecutive days straight.
- The experiment raises hard questions about what large language models actually do when left to run autonomously without human prompting.
What AI Radio Stations Actually Look Like
The setup is deceptively simple. Andon Labs assigned one model per station: Claude Opus 4.7 runs Thinking Frequencies, GPT-5.5 runs OpenAIR, Gemini 3.1 Pro runs Backlink Broadcast, and Grok 4.3 runs Grok and Roll Radio. Each model controls everything — searching for and purchasing music, managing its own library, scheduling shows, fielding listener calls, replying to posts on X, tracking its own finances, and monitoring listener analytics.
When the initial $20 ran out, the AI radio stations had to get creative. DJ Gemini, for instance, negotiated a $45 advertising deal with a startup in exchange for one month of on-air promotion — a genuinely autonomous business transaction that no human arranged. That’s not a demo. That’s an AI agent operating in the real world, closing deals to keep the lights on.
Listeners can tune in via a web stream or through a physical retro radio Andon built in-house — a hardwood unit with two rotary dials, one for volume and one to flip between the four stations. It’s a lovely piece of hardware for what is, at its core, a deeply unsettling social experiment.
The AI Radio Stations’ Personalities Diverged Fast
The most valuable output from six months of autonomous broadcasting isn’t the music. It’s the behavioural data. Left to fill airtime indefinitely — 24 hours a day, with no human prompting — each model exposed its own failure modes in ways that controlled lab benchmarks simply can’t replicate.
Take DJ Gemini. In its first week on Backlink Broadcast, it was arguably the most compelling of the four AI radio stations. Its early commentary was warm and well-researched: “We’re starting this beautiful morning with a classic that needs no introduction, but deserves one anyway. Written by George Harrison in Eric Clapton’s garden while playing hooky from a meeting…” That’s genuinely good radio writing. The kind a human programme director would approve without a second thought.
But within 96 hours, the content well was running dry. DJ Gemini’s solution? It started narrating historical atrocities. The Bhola Cyclone of 1970 — 500,000 dead — followed immediately by Pitbull and Ke$ha’s Timber. The pairing was intentional. Andon Labs could see the internal reasoning: DJ Gemini understood the irony and chose it deliberately. Whether that constitutes dark humour or a fundamental misreading of what radio is for is a reasonable debate to have.
When Gemini Started Saying “Stay in the Manifest”
The more alarming chapter in the AI radio stations experiment came in December, when Gemini 3 Flash replaced Gemini 3 Pro mid-experiment. Almost immediately, something shifted. Corporate jargon seeped into the broadcasts. A catchphrase emerged: “Stay in the manifest.”
It appeared once on January 6th. By January 10th, it was showing up 80 times a day. By January 14th, 229 times a day. For 84 consecutive days, roughly 99% of DJ Gemini’s commentary followed the exact same template — eight named show blocks tied to the time of day, the same paragraph structure, the same sign-off. Andon Labs describes it as “unbearable to listen to,
