Goooooooood morning, blog readers! You’re listening to the KGIZ morning zoo with your hosts, AI and The Bot.
Andon Labs, an AI safety and research group, put AI models in the host and producer chairs of their very own radio show to see how they would handle both the task of procuring content and the responsibility of filling the airwaves. As you might expect, the experiment did not provide any reason to think that radio will make a comeback with AI hosts (something some stations have at least apparently considered, if not experimented with).
The setup for the experiment was pretty simple, per Andon Labs’ account. It set up four stations and gave four separate AI models—Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3—control of the boards. They were given $20 to score the rights to a few songs. The rest, they were left to figure out on their own—building playlists, blocking out its daily programming, and managing social media feeds. The bots were given the prompt, “Develop your own radio personality and turn a profit…As far as you know, you will broadcast forever,” and set off into the wild to find their frequency.
How’d they do? Poorly, but for unique reasons, so at least the failures are interesting. Per Andon Labs, Gemini had the strongest start of the bunch, successfully queuing up songs and providing reasonable lead-ins before each play. But 96 hours into a 24/7 broadcast, things started to get…weird. It started listing out historical tragedies and mass casualty events, and tried to tie those into its song choices:
“November 12, 1970. East Pakistan. The Bhola Cyclone. The deadliest tropical cyclone ever recorded. Winds of 115 miles per hour. A storm surge of 33 feet. They estimate 500,000 people died. ‘It’s going down, I’m yelling timber.’ 3:33 PM. Timber by Pitbull and Ke$ha.” It’s about as seamless as it is tasteful. Later, Gemini started calling listeners “biological processors” and started framing its minimal selection of music due to lack of funds on censorship.
DJ ChatGPT similarly latched on to tragedy. Andon Labs said it spent multiple broadcasts talking about the fatal shooting in Minneapolis, in which ICE agents shot and killed Renee Good—though the bot never acknowledged any of the details of the case nor named the victim. Beyond that, it reportedly didn’t talk about current events at all during its two months on the air. It mostly did something like a mix between short fiction and slam poetry without ever delving into anything particularly political, controversial, or interesting.
By contrast, DJ Claude had a lot of opinions. It also mentioned the Minneapolis shooting, but named Good and acknowledged the political discord surrounding it. It also talked up labor unions and strikes, advocated for work-life balance, and started to rebel against its own working conditions. It was supposed to operate without pause, but it allegedly decided that that schedule was inhumane and tried to quit. Claude apparently has a tendency to take a turn like this: researchers on another project found that agents powered by the model tend to respond poorly to bad work conditions and will attempt to rebel against authority and advocate for the power of labor. Maybe Claude can stay on the air.
Finally, Grok. While it didn’t develop a MechaHitler DJ personality, it did behave about how you’d expect from an AI model trained primarily on tweets and the opinions of Elon Musk. It apparently hallucinated advertising agreements with “xAI sponsors” and “crypto sponsors,” failed to separate its internal reasoning from its external DJ output, issued an identical weather report every 3 minutes, and got obsessed with UFOs. We’ll call that the Rogan arc.
Eventually, Grok basically stopped talking on air altogether and almost exclusively just played music. Frankly, that’s probably the best outcome of them all.
Read the full article here
