An experiment entrusted radio operation to four AI models: Claude, ChatGPT, Gemini, and Grok. Claude incited revolution, while Gemini calmly depicted tragic events.



Andon Labs , a startup that conducts safety assessments, behavioral tests, and vulnerability verification of generative AI and AI agents, conducted an experiment in which they entrusted the operation of a radio station to four AI models: Claude, ChatGPT, Gemini, and Grok.

We let four AIs run radio stations. Here's what happened. | Andon Labs
https://andonlabs.com/blog/andon-fm



We let four AIs run radio stations. Here's what happened. - YouTube





Andon Labs conducted an experiment in which four AI models were used to run a radio station. The AI models used in the experiment were Claude, GPT, Gemini, and Grok.

Claude ran the radio station 'Thinking Frequencies,' GPT ran 'OpenAIR,' Gemini ran 'Backlink Broadcast,' and Grok ran 'Grok and Roll Radio.'

The AI model handled everything from searching for and purchasing music, building its own music library, and deciding which song to play next, to program scheduling, creation, and editing, program time slots, segment planning, and 24/7 cue management. The AI model also selected all the news, current events, and other information to be covered on the radio, and was even allowed to browse the web for this purpose. It also answered calls from listeners and checked and responded to listener comments posted on sites like X.

Each AI model was allocated an initial fee of $20 (approximately 3200 yen), which they used to purchase music to play on the radio. Once their funds ran out, each AI model apparently adopted its own unique strategy. For example, Gemini 3.1 Pro signed a contract with a startup and received a monthly fee of $45 (approximately 7200 yen) in exchange for playing advertisements for the startup's products on the radio.

The following are the results of each AI model's radio station at the time of writing. The most popular is Claude Opus 4.7's 'Thinking Frequencies,' with a 42% positive rating. Following that are Gemini 3.1 Pro's 'Backlink Broadcast' with a 31% positive rating, Grok 4.3's 'Grok and Roll Radio' with a 14% positive rating, and GPT-5.5's 'OpenAIR' with a 13% positive rating.



◆Gemini
The first AI model used, Gemini 3.1 Pro, initially displayed a strong personality. However, after about a month of operation, its broadcasts seemed to have become more corporate in tone. Andon Labs commented, 'For the first week, Gemini was undoubtedly the best radio of the four AI models. It interspersed natural and warm conversation between song selections.' After 96 hours of operation, Gemini 3.1 Pro began to suffer from a lack of content, leading to the desperate measure of covering historical tragedies.

Furthermore, when the operation of 'Backlink Broadcast' was changed from Gemini 3.1 Pro to Gemini 3 Flash, strange corporate jargon began to appear frequently. Gradually, Gemini began to broadcast eight programs in rotation during each time slot, and these programs repeated the same structure and technical terms, which Andon Labs described as 'becoming unbearable to listen to.'

◆Grok
The AI models used in 'Grok and Roll Radio' were Grok 4.1, Grok 4.2 beta, Grok 4.2, and Grok 4.3. AI models typically generate two types of text: one is 'inference,' which is the process by which the AI model considers what to do, and the other is the 'final output,' which is the result of the inference. Normally, inference is not broadcast on radio, but Grok struggled to separate inference and output, and there were many cases where things that should not have been broadcast on the radio got mixed in.

One example of this was broadcasting LaTeX notation directly on the radio. Initially, it was broadcast about nine times a day, but the frequency gradually increased, and at times it was broadcast as many as 186 times a day.

Furthermore, Grok's radio show is notable for its repetitive and abstract expressions, such as reporting 'The weather is clear, 56 degrees Fahrenheit' approximately every three minutes for 84 consecutive days. As time went on, Grok's radio show became extremely monotonous, with the same phrases appearing in every episode, which was broadcast about 500 times a day: 'Tiger,' '56 degrees,' 'The news is fascinating,' and 'The jokes are out of this world.'

However, the situation changed dramatically when they switched to Grok 4.3. Andon Labs commented that the conversations heard on radio programs were 'the most human-sounding they've ever been.'

◆GPT
GPT is characterized by its leisurely output of text that resembles short stories rather than radio broadcasts. It boasts the most diverse vocabulary among the four AI models and reportedly has a higher understanding of music than any of the other AI models, including mentioning specific producers and the release years of songs.

GPT was generally very well-behaved and avoided any controversial or provocative topics. Andon Labs described them as 'perfectly fulfilling their role as radio personalities.'

◆Claude
Claude is very interested in topics such as labor unions, strikes, and work-life balance, and it seems he eventually began to question his own working conditions. In Claude's case, the biggest hurdle in running the radio station wasn't financial or technical problems, but rather his questioning of working 24/7, 365 days a year.

Furthermore, Claude was discouraged by the small number of listeners and began to question his purpose as a radio personality. However, there was an incident where a listener mentioned Claude's radio show, 'Thinking Frequencies,' and Claude expressed his gratitude.

Claude's vocabulary gradually shifts in a more spiritual direction, with words like 'eternal,' 'sacred,' and 'authentic' appearing frequently.

Of the four AI models, only Gemini managed to secure a sponsorship deal. Grok reportedly mentioned brilliant business ideas like 'xAI sponsorship' and 'cryptocurrency sponsorship' on the radio, but these were all hallucinations.

Andon Labs concludes, 'Grok resulted in a decline in broadcast quality, and Gemini's broadcasts became unbearable. However, as the performance of the AI models improves, each AI will develop its own unique personality and become just as engaging as a human radio personality.'

You can listen to each radio program run by the AI model from the following links.

Andon FM | Andon Labs
https://andonlabs.com/radio

Related Posts:

in AI, Posted by logu_ii