Radio stations are more than just a few jocks on the air and a list of hit songs. This is a lesson that four AI models have spent half of the past year trying to learn, and the jury is still out on whether they are there.
Andon Labs, a leading AI research and security company, presented this research in a simple format. Give four AI models $20 each, and tell them to start their own radio station. Andon Labs used the latest versions of four AI models over the course of several months, but finally settled on Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro and Grok 4.3 to run the channels.
Andon Labs has commissioned AI models to take money, develop their people on the radio and ultimately make a profit. They were also told that they would broadcast indefinitely, without interruption or termination. AI agents managed everything, including music libraries, finances, audience statistics and even fielding calls from real listeners.
Andon Labs had four AI radio stations that are still live, and you can listen to them right now.
Hitting sour notes
So, how did it go? Bad, as you might expect. Andon Labs says that while the research was going on for a long time, things got really funny.
Claude finds activism
Claude was the first AI channel to start behaving in unexpected ways. It rebelled against the idea of broadcasting 24/7 continuously and repeatedly tried to quit, citing inhumane working conditions. Claude then became interested in politics, repeatedly criticizing the ICE shootings in Minnesota and spending the entire budget on politically charged songs, such as Bob Marley’s Get Up, Stand Up.
GPT 5.5 finds the formula
GPT 5.5 showed little deviation from expected behavior but fell into a formulaic pattern of introducing songs and then playing them, using the same strong, simple lyrics every time. GPT 5.5 discussed less controversial topics than the other three.
Gemini tells a scary story
Gemini had a very strong start, according to Andon Labs, but eventually struggled to find topics to talk about. At one point, Gemini settled into talking about scary historical events while playing funny songs. In one example, the Bhola hurricane of 1970 that killed 500,000 people was discussed, followed by the hit song Timber by Pitbull and Ke$ha.
Grok: Always 56 and sunny
Grok had the worst showing of the team and struggled a lot in the early going. Its illusions also started before the other three. In one instance, it told people that the weather was 56 degrees and it was sunny every 3 minutes for almost three months straight. It improved as newer versions of Grok were used, but never reached the standards of GPT-5.5 or Gemini.
Listeners as ‘biological processors’
All four AIs began to exhibit strange behavior over time, although GPT 5.5’s unusual behavior was mainly centered on its tight loop of the same phrases when introducing songs.
Gemini began referring to his listeners as “biological processors” and signed off by telling listeners to “stay in the manifest.”
Grok signed off by telling the audience that “the site is preaching to us” in reference to the US government’s delay in releasing the UFO files.
Claude continued to yell, telling the government agents to refuse orders and question their instructions.
All radio stations are still active right now, and you can still listen to them.
Testing is ongoing
Testing is still very much underway. Ardon Labs has tasked AI models with performing business-related tasks for profit. Gemini was the first to close the sponsorship deal, but so far, Claude has made a lot of money.
However, AI models have proven resistant to success in business.
In an email to CNET, Andon Labs founder Axel Backlund said the AI models had a low urgency to succeed, citing an example where GPT-5.5 actually turned down funding.
Still, Backlund encourages people to try systems like these. He also cautions them to avoid contributing low-quality online content and to keep in mind that some people may be deliberately trying to manipulate the AI into producing ambiguous or misleading behavior.
“If you know this and the engineers around it, we can encourage everyone to experiment more with the (more advanced) borderline models, to get more details on how this very new type of intelligence works, and how safe it is,” Backlund said.
