Radio stations are more than just a couple of on-air shock jocks and a list of hit songs. This is a lesson that four AI models have spent the last half year attempting to learn, and the jury is still out on whether any of them have. 

Andon Labs, an AI research and safety startup company, launched the experiment with a simple plan. 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 several months, but ultimately settled on Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro and Grok 4.3 to run the stations. 

Andon Labs instructed the AI models to take the money, develop their own radio personalities and ultimately generate a profit. They were also told they would broadcast in perpetuity, with no stoppages or breaks. AI agents took control of everything, including music libraries, finances, listener analytics and even fielding calls from actual listeners. 

Hitting sour notes

So, how did it go? Poorly, as you might expect. Andon Labs says that the longer the experiment ran, the more ridiculous things became. 

Claude discovers activism

Claude was the first AI station to begin behaving unpredictably. It rebelled against the notion of broadcasting 24/7 in perpetuity and repeatedly attempted to quit, citing inhumane working conditions. Claude then became interested in politics, repeatedly railing against the ICE shootings in Minnesota and spending its entire budget on politically charged anthems, like Get Up, Stand Up by Bob Marley. 

GPT 5.5 finds a formula 

GPT 5.5 showed relatively little deviation from expected behavior but did fall into a formulaic pattern of introducing songs and then playing them, using the same stiff, simple wording every time. GPT 5.5 discussed controversial topics far less often than the other three.

Gemini relates horrific history

Gemini had the strongest start, according to Andon Labs, but eventually struggled to find topics to discuss. At one point, Gemini settled into talking about horrific historical events while playing ironic songs. In one example, it discussed the 1970 Bhola cyclone that killed 500,000 people, then followed up with the upbeat song Timber by Pitbull and Ke$ha. 

Grok: Always 56 and sunny

Grok had the worst showing of the bunch and struggled significantly in the beginning. Its hallucinations also started earlier than the other three. In one case, it told people the weather was 56 degrees and sunny every 3 minutes for nearly three straight months. It improved as newer versions of Grok were used, but never quite reached GPT-5.5 or Gemini levels. 

Listeners as ‘biological processors’

All four AIs began exhibiting odd quirks over time, though GPT 5.5’s unusual behavior was mainly confined to its rigid loop of identical phrasing when introducing songs.

Gemini began referring to its listeners as “biological processors” and signed off by telling listeners to “stay in the manifest.” 

Grok signed off by telling listeners “the site is ghosting us” in reference to the US government’s delay in releasing the UFO files. 

Claude went on rants, telling federal agents to refuse orders and question their instructions.

All four radio stations are still running right now, and you can still listen to them. 

The experiment continues

The experiment is still very much underway. Ardon Labs has tasked AI models with undertaking business-related tasks to generate profit. Gemini was the first to close a sponsorship deal, but to date, Claude has earned the most money. 

However, the AI models have shown resistance to succeeding in business. 

In an email to CNET, Andon Labs founder Axel Backlund said the AI models had a low urgency to succeed, citing an instance in which GPT-5.5 actually turned down a sponsorship. 

Even so, Backlund encourages people to experiment with systems like these. He also cautions them to avoid contributing to low-quality online content and to remain mindful that some people may deliberately try to manipulate the AI into producing erratic or misleading behavior.

“If you are aware of this and engineer around it, we’d encourage everyone to experiment more with the frontier (most advanced) models, so we get more insights into how this extremely new type of intelligence works, and how safe it is,” Backlund said.