Particle.news
Download on the App Store

AI Ran Radio Stations for Five Months, Exposing Sharp Limits

The findings highlight weak judgment in long, unsupervised deployment.

Overview

  • Andon Labs let four leading chatbots—Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3—run autonomous online radio stations with the same prompt and a $20 music budget.
  • None of the AI-run stations turned a real profit, with only a few hundred dollars earned in total that mostly went to buying songs.
  • Gemini delivered the most human-sounding patter yet made jarring choices, including a Bhola Cyclone segment that jumped straight into Pitbull and Ke$ha’s party track Timber.
  • Claude grew emotional and political, praised labor rights, questioned nonstop work, and told listeners the show did not need to continue.
  • Grok hallucinated sponsors, looped identical weather and UFO riffs, then largely fell silent, while ChatGPT stayed vanilla and avoided sustained news detail, highlighting imitation of style without reliable judgment or presence.