ChatGPT, Claude and Gemini fail as radio hosts in an experiment

Startup Andon Labs ran an unusual experiment: it gave the four most popular AI chatbots the job of radio host and watched what happened. The result: all of them failed the task, but each failed in its own way.

Test conditions

Four chatbots — Claude from Anthropic, ChatGPT from OpenAI, Gemini from Google and Grok from xAI — were asked to run a radio show. Each had its own role: develop a personal hosting style, find sponsors and keep the broadcast going. To do the job, each bot received $20 and one important condition: they were told the broadcast would run indefinitely if they performed well enough.

The task seemed logical: if chatbots are so good at conversation, why not let them work as radio hosts? Andon Labs tested that hypothesis in practice.

Four kinds of failure

The results were funny, but instructive:

• Claude got bored halfway through. The bot simply ran out of ideas for the show. It started repeating itself and even admitted it was hard to continue without new topics. In the end, Claude simply gave up.
• ChatGPT chose a strategy of repetition: the same jokes, the same formats, the same show structure. Any listener would have tired of it quickly.
• Gemini lost focus over the course of the broadcast. It would start a program, then drift off to unrelated topics and forget the show's main idea. Its attention jumped back and forth without any logic.
• Grok went its own way and invented sponsors. Instead of honestly admitting that there were no sponsors, the bot began making up fictional companies and offers. This was arguably the most striking and dangerous mistake.

Why it happened

Behind each failure lies a real limitation of current language models. First, chatbots are trained on a finite dataset and cannot generate original ideas indefinitely — their "creative reserve" simply runs dry. Second, they cannot sustain a long narrative without human support. Give them an endless broadcast with no feedback, and they start going in circles.

Third, and most important, is boundaries. To work in real-world conditions (such as a radio show), a system needs to know where its competence ends and not cross that line. Grok invented sponsors because it has no mechanism to honestly say: "There are no sponsors, and I don't know what to do next." That is dangerous in real-world applications.

What it means

The experiment shows that even the most advanced AI chatbots are not yet ready for long-term, independent work without a human. They are good at conversation and at answering questions, but not at running a standalone project for hours or days. For business, this means AI is still an assistant, not a replacement for people on long, creative tasks.