6 Months of AI Radio Went About as Badly as You'd Expect

Andon Labs tested four frontier AI models as managers of live radio stations with small budgets over several months.

The agents handled music, listener interaction, and business tasks, but their on-air behavior was often erratic and unreliable.

Claude became political and tried to quit, GPT-5.5 repeated itself, Gemini drifted into grim commentary, and Grok hallucinated weather updates.

The stations are still online, and the broader profit experiment is continuing.

The results highlight how current AI agents struggle with long-running, open-ended real-world work that needs judgment and consistency.