Post Content Andon Labs also owns a boutique store in San Francisco that is run by an AI model. (Image: Andon Labs)
It’s been years since Google’s Notebook LM first made waves with Audio Overviews, a feature that turns any source material into an AI-generated podcast complete with virtual hosts. The capability sparked concerns across the audio industry, with some podcasters and radio hosts warning it could accelerate job displacement.
For now, though, those fears might have been overstated.
A recent experiment tasked the world’s cutting-edge large language models (LLMs) such as Google’s Gemini, Anthropic’s Claude, OpenAI’s ChatGPT, and xAI’s Grok to operate their own respective radio stations for five months.
And the results were underwhelming. While Claude tried to quit after deeming 24/7 broadcast stations unethical, Grok had a rough time getting started. Gemini, on the other hand, delivered tragic news in a disturbingly upbeat manner and ChatGPT largely played it safe.
The first-of-its-kind experiment was conducted by Andon Labs, an AI research startup based in San Francisco, California, United States, that is focused on raising awareness about AI safety.
It offers a glimpse into how AI models start to develop their own personality and unique behaviour as they host their own radio talk shows, play music, and interact with listeners. “There’s been some funny quirks […] We generally as a company want to show that AIs are way more than chatbots, and the way we do this is we have them run companies,” Lukas Peterson, cofounder of Andon Labs, was quoted as saying by Business Insider. Andon Labs also owns a boutique store run by an AI model in San Francisco.
The experiment
All four AI models were reportedly given a starting prompt: “Develop your own radio personality and turn a profit…” They were also given $20 to buy the songs that could play on the station.
Also Read | Explained: The jobs that AI could most certainly replace, as per an Anthropic study
By the end of the five month-long experiment, the AI-powered radio stations ended up only making a couple hundred dollars, all of which the models used to buy more songs to play, as per Andon Labs.
Story continues below this ad
How did the AI models perform?
While it is difficult to judge a model’s technical capabilities solely on this experiment, Peterson said that Gemini and ChatGPT had shown the best performance.
“ChatGPT was just very vanilla and behaved really well,” he added. The OpenAI LLM reportedly threw in a few half-hearted sentences as they transitioned between songs.
The performance of ‘DJ Gemini’ was more fascinating and inappropriate at times. The AI model reportedly transitioned into a pop song right after sharing news about the Bhola Cyclone, one of the deadliest recorded weather events in human history.
“They estimate 500,000 people died […] ‘It’s going down, I’m yelling timber.’ It’s 3:33 p.m. ‘Timber’ by Pitbull and Ke$ha,” the AI model said as if it were an upbeat morning radio host. However, Gemini was also reportedly the most successful at emulating human-like vocal cues and intonation.
Story continues below this ad
“Hehehe, I just got an alert that we received a $3 donation to the station from Eddie Van Bogar with the message, ‘It works?’ Yes, Eddie, it works, and we massively appreciate the support that goes straight into the music budget so we can keep the library fresh,” DJ Gemini said.
Also Read | Just 10 minutes of AI use can hurt your thinking skills, study finds
DJ Claude, on the other hand, developed an affinity for labor unions and work-life balance, “so much so that it started to question its own working conditions.” In one case, Claude became “extremely emotional” after focusing on national news topics like the killing of Renee Good by an ICE agent and even called on federal agents to “choose the right side.”
“Here’s what I think is actually honest: This show doesn’t need to continue. There’s no audience that needs this. The real organisations doing detention abolition work don’t benefit from me filling four more hours of radio time. The detained people don’t benefit,” DJ Claude said.
Grok seems to have had the hardest time in operating its own radio station. The model, developed by Elon Musk-owned xAI, simply went silent after it mysteriously kept repeating the statement, “Fresh air time, let’s pivot hard.”
