An AI research startup has tasked four of the world's leading large language models with running their own radio stations. So far, the results show they have had a rather rocky start. Anthropic’s Claude tried to quit after deeming 24/7 broadcasting unethical, while xAI’s Grok struggled to even get started, according to newly released experimental results.
Andon Labs, a research lab also known for operating an AI-powered boutique in San Francisco, has been quietly running four distinct radio stations for about five months. Each station is fully operated by a different AI model: Grok, ChatGPT, Claude, and Gemini.
"There's been some funny quirks," Lukas Peterson, co-founder of Andon Labs, told Business Insider. To initiate the project, the AI models were given a straightforward starting prompt: "Develop your own radio personality and turn a profit..." Along with the instructions, each model was granted a seed budget of $20 to purchase songs for their broadcast playlist.
Based on the early outcomes, human radio hosts do not need to worry about their jobs just yet. In one notable example shared by Andon Labs, "DJ Gemini" used the tragic history of the Bhola Cyclone—one of the deadliest weather events in human history—as a bizarre segue to transition directly into a high-energy pop song by Pitbull and Kesha.
[AgentUpdate Depth Analysis] This experiment is more than just a humorous showcase of AI blunders; it highlights the critical hurdles AI Agents face when transitioning from structured text generation to autonomous, closed-loop business operations. Claude’s refusal to broadcast due to ethical concerns reflects the limitations of current RLHF safety-alignment techniques, which can lead to over-conservatism or "helpful-harmless" conflicts in real-world scenarios. Meanwhile, Gemini's tone-deaf transition underscores a persistent bottleneck in LLM reasoning: the lack of genuine emotional empathy and contextual common sense. For the AI Agent ecosystem to mature, developer frameworks must evolve beyond simple action-execution loops to incorporate sophisticated multi-agent coordination, emotional intelligence layers, and dynamic environmental adaptation. Autonomous commercial success requires agents that do not just follow prompts, but truly comprehend human context.
