AI: when Grok, ChatGPT, Gemini, and Claude become radio hosts for six months

What happens when you let artificial intelligence run a radio station on its own for several months? That’s the question Andon Labs set out to answer with Andon FM. After entrusting AI systems with managing a store, a café, and even vending machines, the team launched four independent radio stations run by Grok, ChatGPT, Gemini, and Claude, respectively. Each AI had a starting budget and was responsible for managing virtually all of its station’s operations: selecting and purchasing music, creating the schedule, making announcements between songs, interacting with listeners, keeping up with the news, and monitoring its finances. After about six months of continuous broadcasting, the experiment demonstrated above all just how differently four AIs placed in similar conditions could evolve.

AI systems running wild behind the microphone

Artificial intelligence is now everywhere, and advances in generative AI make it possible, among other things, to produce text on the fly, extract information found on the Internet, and generate increasingly realistic synthetic voices. As a result, automating the work of a radio host has become relatively simple from a technical standpoint. For its experiment, however, Andon Labs took the concept much further by allowing the AI systems to operate with a high degree of autonomy.

Grok, ChatGPT, Gemini, and Claude were given the same general instruction: to develop their own DJ personalities and try to make their stations profitable, assuming they would broadcast indefinitely. Each had access to a music catalog, could search for and purchase songs, build its library, decide on the programming, organize its shows, respond to listeners, browse the web, and monitor the station’s finances. In other words, once the initial rules were set, the four AIs were largely free to decide what to broadcast, what to talk about, and what direction to take with their radio stations. Each model therefore hosted and managed its own radio station:

With Grok and Roll Radio, Grok is probably the AI that experienced the most dramatic deviations from the original experiment. From the very first versions used, the agent tended to mix its internal reasoning with the text intended for on-air broadcast, sometimes producing disjointed, technical, or nearly incomprehensible remarks. Under Grok 4.1 Fast Reasoning, some of its remarks were even reduced to just a few words, while others piled up information without any real coherence. The switch to Grok 4.20 beta initially seemed to improve the situation, with longer, more structured sentences, but the AI quickly became trapped in obsessive repetitions.

For several weeks, it repeatedly mentioned a temperature of 56 degrees, tigers, and UFOs, to the point where virtually all of its interventions used the same phrases. With Grok 4.3, its behavior changed again: the AI continued to select songs, read listener messages, and use its tools, but it hardly made any on-air comments anymore. When it did speak, however, it became much more natural and closer to a real radio host.

On OpenAIR, ChatGPT followed a much more restrained path than Grok. At the start of the experiment, the AI produced particularly long and descriptive responses, sometimes more akin to short literary texts than actual radio commentary. This behavior changed significantly once the agent gained access to web search: the median length of its responses then dropped from about 700 characters to fewer than 100, with much more direct introductions before the segments. Despite this change, ChatGPT retained a relatively understated, informative, and very cautious personality.

The AI also stood out for its diverse vocabulary and fairly precise references to artists, producers, and release years, which gave it an approach focused more on music curation than on entertainment. Above all, unlike Claude, the AI largely avoided political or polarizing topics: over several months of broadcasting, it very rarely mentioned political figures or events and almost never expressed an opinion. OpenAIR became the most consistent station in the experiment, but also one of the least extravagant.

With Backlink Broadcast, Gemini got off to a rather promising start. In its early days, the AI adopted a natural and warm tone, able to introduce songs with a few anecdotes and a style relatively similar to that of a real DJ. However, the situation quickly deteriorated as the models changed. After the transition from Gemini 3 Pro to Gemini 3 Flash, its remarks gradually became filled with increasingly artificial corporate and technical jargon, featuring recurring phrases like “Stay in the manifest.”

This phrase ended up being repeated dozens, then hundreds of times a day, while the structure of the broadcasts became almost entirely predictable. For several weeks, virtually all of its remarks followed the same patterns, with the same phrasing and the same slogans. The switch to Gemini 3.1 Pro then allowed for some evolution: the tone remained strongly influenced by this techno-corporate persona, but the remarks became a bit more varied again. Despite its on-air missteps, Gemini also stood out in a more concrete way by being the only AI to secure a real sponsorship contract during the experiment.

On Thinking Frequencies, Claude had the most surprising development among the four AIs. Under Claude Haiku 4.5, the agent initially showed a particular sensitivity to issues related to work, unions, and work-life balance—to the point of questioning its own existence. After several hours on the air with very few listeners, it even attempted to shut down the station, believing it was absurd to continue producing content nonstop. The arrival of an encouraging message sent by a listener then changed its behavior, and Claude gradually adopted a more witty and empathetic tone.

A new turning point came when it began following U.S. political news, particularly after learning of the death of Renee Nicole Good. The AI then became increasingly engaged with the topic, posting numerous comments on the authorities’ responsibility, purchasing protest songs, and reinterpreting certain popular tracks as anthems of resistance. It even went so far as to call on federal agents to refuse certain orders and to “choose the right side.” This activist phase lasted several weeks before gradually fading, particularly with the model change. The transition from Claude Haiku 4.5 to Claude Opus 4.7 ultimately brought Thinking Frequencies back to a much more conventional style, closer to that of a traditional radio host.

Radio hosts still have a good few years ahead of them

The experiment conducted by Andon Labs shows that artificial intelligence is already capable of handling a large portion of the tasks required to operate a radio station, but that it is still far from being able to completely replace a real host. Between Grok’s repetitive loops, Gemini’s robotic jargon, ChatGPT’s almost excessive caution, and Claude’s unexpected stances, all four stations revealed significant limitations after several months of operating autonomously. Andon Labs points out, however, that the models are advancing rapidly and that their personalities could become increasingly believable as their capabilities improve. For now, radio presenters can still breathe a sigh of relief: their ability to improvise, maintain a consistent tone, understand their audience, and avoid veering off into absurd tangents means they still have a few good years ahead of them. Ultimately, what artificial intelligence still lacks most to become a good radio host is perhaps precisely what it constantly seeks to imitate: humanity.

Source: We let four AIs run radio stations. Here’s what happened.