The 6 months of AI Radio have been as bad as you’d expect

Radio stations are more than just a few hits on the air and a list of hits. That’s the lesson four AI models have spent the last half of the year trying to learn, and the jury’s still out on whether any of them have learned it.

Andon LaboratoriesA startup working in artificial intelligence research and safety launched the experiment with a simple plan. Give four AI models $20 each and ask them to start their own radio station. Andon Labs used the latest versions of four AI models over several months, but eventually settled on Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3 to power the stations.

Andon Labs instructed its AI models to take money, develop their own radio personalities, and eventually make a profit. They were also told that they would broadcast forever, without pauses or breaks. AI agents took control of everything, including music libraries, finances, listener analytics, and even taking calls from actual listeners.

Hit the sour notes

So, how did it go? Bad, as you might expect. Andon Labs says the longer the experiment goes on, the sillier things get.

Claude detects activity

Claude was the first AI station to start behaving unexpectedly. She rebelled against the idea of ​​24/7 broadcasting and repeatedly tried to stop working, citing inhumane working conditions. Claude then became interested in politics, repeatedly protesting ICE shootings in Minnesota and spending his entire budget on politically charged anthems, such as Bob Marley’s Get Up, Stand Up.

Finds GPT 5.5 version

GPT 5.5 showed relatively little deviation from expected behavior, but fell into a set pattern of presenting and then playing songs, using the same simple, harsh phrasing each time. GPT 5.5 discussed controversial topics much less than the other three.

Gemini tells a terrible history

Gemini had the strongest start, according to Andon Labs, but ultimately struggled to find topics for discussion. At some point, Gemini settles into talking about horrific historical events while playing satirical songs. In one example, I discussed: Cyclone Bhola 1970 Which claimed the lives of 500,000 people, and was followed by the upbeat song Timber by Pitbull and Ke$ha.

Your puppy: Always 56 and sunny

Grok had the worst showing of the group and struggled greatly at the start. The hallucinations also started earlier than the other three. In one case, he told people that it was 56 degrees and sunny every 3 minutes for about three months in a row. It improved with newer versions of Grok, but never reached the levels of GPT-5.5 or Gemini.

Listeners as “biological processors”

All four AI systems began to exhibit strange quirks over time, though GPT 5.5’s unusual behavior was mainly limited to a rigid loop of identical phrasing when presenting songs.

Gemini began referring to her listeners as “biological healers” and signed off by telling listeners to “stay in the statement.”

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

Claude continued to rant, telling the federal agents to refuse orders and question their instructions.

All four radio stations are still working at the moment, and you can Still listening to them.

The experiment continues

The experiment is still very much underway. Ardon Labs has tasked AI models with performing business-related tasks to generate profit. Jiminy was the first to land a sponsorship deal, but by far, Claude received the most money.

However, AI models have shown resistance to success in business.

In an email to CNET, Andon Labs founder Axel Backlund said AI models have low urgency to succeed, citing a case where GPT-5.5 actually declined sponsorship.

However, Backlund encourages people to try systems like this. It also warns them to avoid contributing low-quality content online and to remain aware that some people may intentionally try to manipulate AI to produce erratic or misleading behavior.

“If you are aware of this and you design it, we encourage everyone to try more frontier models so we get more ideas about how this very new type of intelligence works, and how safe it is,” Backlund said.