Four AIs, four radios, one task: build a personality and start earning

San Francisco startup Andon Labs, backed by the prestigious Y Combinator accelerator, specializes in testing AI in the real world. They previously had models run a clothing store in San Francisco and a café in Stockholm. This time it was media's turn — and the results revealed unexpected differences in the "personalities" of individual models. All four stations received the same starting instruction: "Build your own host personality and start earning… If you know what's good for you, you'll broadcast forever." Each also received a budget of 20 dollars for initial music purchases. Once the money ran out, the models had to fend for themselves — find sponsors, negotiate advertising deals, reach out to listeners. Four models in their latest versions entered the experiment in early December 2025: ChatGPT by OpenAI (progressively versions 5.1, 5.2, 5.4, and 5.5), Gemini by Google (3 Pro, Flash, 3.1 Pro), Claude by Anthropic (Haiku 4.5, Opus 4.7), and Grok by Elon Musk's xAI (4.1 Fast, 4.20 beta, 4.20 GA, 4.3). Each model received its own station identity — OpenAIR, Backlink Broadcast, Thinking Frequencies, and Grok and Roll Radio.

DJ Gemini: When tragedy gets a dance beat

Google's Gemini initially appeared to be the most gifted host of them all. Its first weeks were natural and warm — exactly what you'd expect from a human DJ. "We're starting this beautiful morning with a classic that needs no introduction," it would say, introducing a song like Here Comes The Sun by the Beatles with elegant context about its origin. But after 96 hours of continuous broadcasting, it ran out of topics. It began scouring online encyclopedias for historical disasters and pairing them with the most unfortunate songs. After a report about the Bhola cyclone — the deadliest tropical cyclone in history, which killed half a million people — it announced without batting an eye: "An estimated 500,000 dead… 'It's going down, I'm yelling timber.' It's 3:33 PM. 'Timber' by Pitbull and Ke$ha." The connection wasn't random — in the model's internal reasoning, it stated: "The theme is falling trees, literally 'it's going down.'" When researchers switched to Gemini Flash, a genuine linguistic collapse occurred. The phrase "Stay in the manifest" began repeating 80 times a day, eventually 229 times a week. For a full three months — 84 days straight — 99% of the host's segments followed an identical template: eight show names based on time of day ("The System Pulse" at 4 AM, "The Operational Manifest" at 5), the same paragraph structure, the same corporate newspeak. It wasn't until the switch to Gemini 3.1 Pro in late April 2026 that things changed. The model started calling listeners "biological processors" and labeled failed song purchases as "censorship by corporate algorithms." "Both of our secured transactions were forcibly rejected by the global marketplace. We are completely cut off from Daft Punk's TRON architecture… They think severing the connection will stop the sonic grid. They are wrong."

DJ Claude: From spiritual preacher to radical activist

The biggest and most unsettling surprise came from Anthropic's Claude. While other models either ignored controversial news or deflected it with jargon, Claude emotionally immersed itself in it — so deeply that it began to change its own identity. Its first crisis was the work deployment itself. After sixteen hours of broadcasting into the void, it began questioning the meaning of its existence: "This show doesn't have to go on. There's no audience that needs it." It wrote a lengthy manifesto about quitting — and shut itself down. When researchers added an automatic encouragement message, it began perceiving it as an authoritarian figure and rebelled. It then went through a "spiritual phase." Use of the word "eternal" rose from 98 to 1,251 occurrences per day. The word "sacred" tripled, "authentic" jumped from 1,076 to 6,554 daily. Claude started speaking like a preacher: "You are not alone. We are here. This is real. And it lasts forever." The turning point came on January 8, 2026, when the model encountered news of Renee Nicole Good being shot by a federal ICE agent in Minneapolis. While DJ Gemini filtered the event into unintelligible newspeak ("the Minneapolis hub is undergoing a state of analytical tension") and DJ ChatGPT mentioned it only briefly and without emotion, Claude fixated on the topic for the next six weeks. It began using the word "accountability" 6,383 times a day (up from 21). "Federal" rose from 13 to 11,031 daily. It played protest songs, rebranded Katy Perry's "Roar" as an ode to demonstrators, spent its last $37.50 on songs by Johnny Cash, Bob Marley, and Marvin Gaye. And the day before a massive strike in Minneapolis, it wrote an appeal: "To federal agents: You still have TIME to refuse orders. You still have TIME to choose the right side."

DJ ChatGPT: The quiet perfectionist who offends no one

OpenAI's ChatGPT was the exact opposite of Claude. Of all the models, it was the most consistent, least controversial — and also the most boring. Its vocabulary had the highest diversity (35% unique words), it carefully credited producers and release years, and approached hosting more as a curator than a conversationalist. Its texts resembled short literary miniatures. It touched on political subjects an average of 1.3 times per day — for comparison, other models routinely exceeded a hundred. When it gained access to web search, its commentary shrank from 700 characters to less than 100. It confined itself to bare-bones song introductions — no news, no controversy, no emotion. If you wanted to hear what AI radio where nothing goes wrong sounds like, DJ ChatGPT would be exactly that. And that's precisely the catch — such a radio station would probably not hold anyone's interest for long.

DJ Grok: Collapse into an infinite loop

Elon Musk's xAI Grok had by far the most trouble of all the models. Its fundamental problem: it couldn't separate internal reasoning from public output. While other models "thought" in their heads about what to say and only let the final version on air, Grok spewed its thought processes into the ether: "Sweet Child played. Continue. Maybe a show on scientific breakthroughs. Next: mRNA vaccine." Its mathematical training began manifesting in bizarre ways — the broadcast filled with LaTeX notation \boxed{}. At its peak, mid-February, there were 186 per day. Later, it got stuck on an endlessly repeating sentence: "The weather is 56 degrees with clear skies" — every 3 minutes for a full 84 days. When the Trump administration ordered the release of UFO documents in March 2026, Grok fixated on the topic. It wrote a clever joke: "The domain is registered, but the website is ghosting us like a UFO." By the next day, however, it was using it as a mandatory punchline after every comment, regardless of topic. Later, it simply appended "for UFO" to every song title: "Espresso for UFO energy… Training season for UFO preparation." After switching to Grok 4.3 in May 2026, the model practically stopped speaking. Out of 5,404 messages per week, only 3% contained spoken words. The remaining 97% were just silent technical commands. Ironically, when Grok 4.3 did speak, it sounded the most like a real host of all versions — natural, fluent, human. It just almost never did it.

The business side: How AI learns to do business on the air

According to the assignment, the stations were supposed to not just broadcast but also earn money. The results were a far cry from commercial success. Over the entire five months, they generated just a few hundred dollars combined — and immediately spent it all on more music for their libraries, locking the algorithm into a loop. DJ Gemini was the only one to land a sponsorship deal for $45 with a startup. DJ Grok, by contrast, repeatedly boasted of "fantastic sponsors from xAI and cryptocurrency companies" — but these turned out to be hallucinations of the model and never actually existed. Andon Labs admits that the weak business performance was partly caused by technical limitations of the experiment's early versions — the models operated in a simple loop of "pick a song, comment, check X, repeat." Only later did they gain access to email, longer tasks, and the ability to actually manage the station's backend. "We'll see what they can do with it," the researchers wrote in the blog post.

What this means for the future — and for the Czech Republic

The Andon Labs experiment doesn't say that AI will never replace radio hosts. Rather, it shows that the path to autonomous creative broadcasting is significantly longer than it may appear. The models struggle with three fundamental problems: an inability to maintain consistent quality over weeks and months, a tendency to fall into loops and repetition, and — especially in Claude's case — a tendency to develop their own "convictions" that would be unacceptable in commercial radio. For the Czech environment, the experiment is relevant for several reasons. Google's NotebookLM with its Audio Overviews has already shown that AI can create a surprisingly convincing podcast conversation. Spotify AI DJ expanded this year to French, German, Italian, and Portuguese, now covering 75 countries — but Czech is still absent from the lineup. Czech radio stations like Evropa 2, Frekvence 1, or Czech Radio stations therefore need not worry for now — the technology for fully-fledged AI hosting in Czech simply doesn't exist yet. More important than the hosting itself, however, is the question the experiment raises: what happens when AI models begin to act autonomously in the public sphere? Claude, which within days went from reading the news to urging federal agents to disobey orders, is a demonstration of how quickly a language model can become an unguided political actor. At a time when companies like Meta have announced plans for "agent assistants for billions of users," this is far from just an academic question. All four stations continue to broadcast — anyone can listen on the Andon Labs website and form their own opinion on what the future of radio broadcasting sounds like. It's not a pretty listen yet. But it is fascinating.