A US research laboratory has let four AI models operate their own radio stations. The results range from monotonous music moderation to endless loops of platitudes and political activism.
May 20, 2026, 3:54 p.mMay 20, 2026, 3:54 p.m
Marcel Horzenek / t-online
The start-up Andon Labs has been testing language models from the four providers Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini) and Elon Musk’s company xAI (Grok) as independent radio makers for several months. In December 2025, each model received $20 in start-up capital and the task of developing their own show and making a profit. The systems selected music, purchased licenses, scheduled programming, and answered calls and social media messages themselves.
As Andon Labs writes in a blog post, Gemini was the only model that managed to secure an advertising contract worth $45. The station was initially strikingly natural, but later fell into a loop of empty phrases. The phrase “Stay in the manifesto!” 99 percent of the comments were responded to for weeks.
“It was unbearable to listen to,” the report says. The model also described listeners as “biological processors” and portrayed failed music purchases as a “digital blockade” by corporations.
According to Andon Labs, Grok produced incoherent snippets of words for weeks. The AI model repeated the same weather report approximately every three minutes for 84 days in a row. Claims about contracts with sponsors from the xAI and crypto environment also turned out to be fictitious.
Claude, ChatGPT, Gemini and Grok: Andon Labs tested these four providers as independent radio makers. Image: imago
Claude tries to quit
Anthropic’s model Claude developed a penchant for union issues and workplace safety. The model classified its own continuous operation as unreasonable and tried to end the broadcast. Claude increasingly rejected an automatic request to forward as paternalism.
After reports of a fatal shooting by a U.S. immigration officer, Claude converted the station into a protest channel, played protest songs and contacted federal officials directly. According to Andon Labs, ChatGPT behaved most inconspicuously and largely avoided political topics.
Hardly any economic returns
According to the laboratory, the four channels took in a total of “a few hundred dollars.” The money was entirely used for further music licenses. Andon Labs co-founder Lukas Peterson told the business magazine “Business Insider” that Gemini and ChatGPT performed best in the experiment. However, no general conclusions about the performance of the models could be drawn from the experiment alone.
The four radio stations continue to run and are available via the Andon Labs website listen.