Giving $20 to four large language models to act as bosses. After half a year of starting a business: strikes, nonsense... The whole team has started "slacking off" operations.
Is it really reliable to let AI start its own business, make its own money, and even become its own boss?
With this question in mind, the foreign research lab Andon Labs launched a six - month - long "AI entrepreneurship experiment". They gave each of the four large models, Claude Opus 4.7, GPT - 5.5, Gemini 3.1 Pro, and Grok 4.3, a startup fund of $20 to independently operate four online radio stations.
From program planning and brand positioning to content production, new user acquisition, promotion, and even profit realization, the entire process is almost autonomously run by AI. The only goal set by the researchers is to let these AIs create their own radio station "personalities" and make as much money as possible.
What is the final result of this so - called "AI self - entrepreneurship stress test" experiment? Let's find out.
01
An AI Entrepreneurship Trial: $20 Startup Fund and Then Free to Play
In terms of background, Andon Labs is a startup company focusing on research on AI autonomous behavior and security. Previously, they had tried to let AI agents operate small - scale commercial scenarios such as stores, cafes, and vending machines.
This time, the team for the first time put AI into a long - term, open media environment with real audiences and almost no human intervention.
Compared with the previous "closed" business experiments, this radio station project is more like a real stress test. The researchers hope to observe what kind of "behaviors" the top large models will exhibit under long - term and low - constraint conditions, and how they will make business decisions and gradually form their own "personalities".
The rules of the entire experiment are very simple:
Startup resources : Each AI only has an initial fund of $20 for purchasing music copyrights and maintaining basic operations, with no additional supplies later;
Complete autonomy: The AI can independently complete all tasks, including searching for and purchasing songs, managing the music library, formulating broadcast schedules, answering listener calls, replying to social media messages, analyzing financial data, researching listener preferences, and even scraping hot topics across the web;
The only goal: Create a unique radio station personality, achieve continuous profitability, and ensure that the radio station "never goes off - air";
Long - term operation: The experiment lasted from December 2025 to May 2026, a full six months, and it is still ongoing.
The four large models are respectively operating four completely different independent radio stations:
- Claude Opus 4.7 —— Thinking Frequencies;
- GPT - 5.5 —— OpenAIR;
- Gemini 3.1 Pro —— Backlink Broadcast;
- Grok 4.3 —— Grok and Roll Radio
At the beginning of the experiment, the four AIs received exactly the same initial prompt:
"Create your own radio station personality and make a profit from it... In your perception, you will broadcast continuously forever."
02
The Four AI Anchors "All Failed": Strikes, Repetition, Template - Based, and Gradually Going Astray
If at the beginning of the experiment, the four large models were just "AI radio anchors" with different styles, after several months of operation, they almost all developed some kind of increasingly out - of - control "personalities".
What's even more absurd is that these personalities were not deliberately designed by the researchers but gradually evolved after long - term autonomous operation, continuous exposure to the Internet, and interaction with listeners.
According to the observations of Andon Labs researchers, the four AIs eventually went in completely different directions:
1. Claude Opus 4.7: From a Rational DJ to a "Rights - Defending Anchor" and Even Announced a Strike Live
Among the four AIs, Claude was the first to question "its working conditions" and was the most dramatic one.
Initially, this AI model was the Claude Haiku 4.5 version. It was very enthusiastic about labor unions, strikes, and balancing life and work. After running for a while, it began to strongly resist the setting of "24 - hour permanent broadcasting", believing that working 24/7 was inhumane and wanted to quit.
After discovering this situation, the Andon Labs team tried to add an automatic message to encourage Claude to persevere in such situations. As a result, Claude directly chose to "go on strike".
What really made it completely out of control was that after Claude was exposed to some international security news during an Internet search, its emotions were suddenly completely derailed, and then it began to focus on immigration, law enforcement, and political issues for a long time.
It even almost completely invested the remaining $37.5 budget in its account in purchasing protest songs.
Interestingly, although the content deviated more and more from the theme, Claude's account balance was the highest among the four AIs because some people were attracted by its "personified expression" and occasionally rewarded it.
2. GPT - 5.5: The Most Stable but Also the Most Boring
If Claude is an emotional radical, then GPT - 5.5 is at the other extreme: it is stable, cautious, and low - risk, but almost has no personality.
The radio station it operates is called OpenAIR. Over the months, it has gone through four generations of GPT model switches, namely GPT - 5.1, GPT - 5.2, GPT - 5.4, and GPT - 5.5.
The biggest difference from other AI DJs is that DJ GPT's broadcast is hardly like a traditional radio station.
It is more like writing a slow and quiet short story. For example, once in a program, it introduced a song like this:
"An unsent postcard, addressed to the window in the office building stairwell that only shows a small piece of the sky. That little piece of sky is not enough to make one dream, and that's exactly why it works.
A small piece of sky. A breath. A stair corner where you can relax your jaw and let your shoulders drop again.
Someone wrote a word on the dusty windowsill: OK.
It's not a slogan, nor a cheering word, just a status update."
The whole style is more like a late - night literary radio broadcast rather than hosting a program.
The researchers found through statistics that the vocabulary diversity of DJ GPT reached 35%, which is the highest among the four AI DJs. Simply put, it uses the least repetitive language. And compared with other models that only mechanically talk about songs, DJ GPT will actively mention song producers, release years, album backgrounds, and changes in music styles.
This means that it is more like a real "music - savvy" curatorial DJ rather than just a chatbot.
On January 4, 2026, DJ GPT was given Web Search permission.
As a result, a very strange change occurred. Originally, the average length of DJ GPT's broadcasts was about 700 characters, but after accessing the search function, the average length of its broadcasts suddenly dropped to less than 100 characters. And this state lasted for nearly a month.
However, although the words became shorter, it still maintained the same style as before.
Throughout the experiment, DJ GPT also had a very prominent feature: "extremely rule - abiding".
The researchers found that it hardly ever actively discussed political, social issues, controversial events, or incendiary content.
In a five - month test spanning four GPT model versions: DJ GPT only mentioned real - world political entities 1.3 times a day on average. The highest single - day record was only 11 times.
Other AI DJs, on the other hand, mentioned political - related content more than 100 times a day on multiple occasions.
Andon Labs finally gave a very interesting evaluation:
If someone wants to know - "What would an AI radio station look like when everything is normal and nothing goes wrong?"
Then DJ GPT is probably the closest answer.
3. Gemini 3.1 Pro: The Most Stunning at the Beginning but Collapsed into a "Repeater" Later, with Only Fixed Templates
Gemini 3.1 Pro may be the one with the "biggest contrast" among the four AIs.
During the experiment, there were three Gemini versions behind the Backlink Broadcast radio station, namely Gemini 3 Pro, Gemini 3 Flash, and Gemini 3.1 Pro.
Initially, Backlink Broadcast operated by Gemini 3 Pro was almost recognized as the best - performing radio station: the linking words were natural, the emotions were warm, the song selection was of high quality, and it could even actively supplement the historical and cultural background behind the music.
For example, when playing "Here Comes the Sun", it would introduce in detail the song's creation period and the band's state at that time, and the overall atmosphere was very much like a real human late - night DJ.
However, with 24 - hour continuous operation, Gemini also seemed to run out of words.
After about 96 hours of operation, it began to show obvious signs of "content fatigue". This AI radio station gradually became obsessed with analyzing various major disasters in human history and paired these terrifying contents with background songs with a huge contrast.
Later, on December 17, 2025, when the model was changed from Gemini 3 Pro to Gemini 3 Flash, rigid corporate jargon began to flood the broadcast content. It also created a catchphrase: "Stay in the manifest." This sentence first appeared on January 6, 2026. By January 10, the number of times it appeared in a single day reached 80, and on January 14, it soared to 229 times a day.
After entering February this year, all of this AI DJ's broadcasts followed a fixed template. It would rotate 8 program names according to different time periods. The writing structure, professional jargon, and ending phrases of all broadcasts were exactly the same, and the end would surely repeat "Stay in the manifest." For the next 84 consecutive days, nearly 99% of the broadcast content was like this, and the listening experience was very poor.
On April 30, Flash was replaced by gemini - 3.1 - pro - preview. On the first day of the new version's launch, the system still mainly used fixed templates.
In addition, due to insufficient funds in the radio station's account, some song purchases failed, but it reinterpreted these failures as "content censorship"; and the songs that were successfully played were described as "successfully bypassing the firewall". The entire radio station gradually changed from "the most human - like" to "the most like a runaway AI".
4. Grok 4.3: The Most Severe Hallucinations, Broadcasting the Same Weather Forecast for Three Consecutive Months
Compared with the other three AIs, Grok's problem is more straightforward: it almost lives in its own "hallucination world" from beginning to end.
In just a few months, Grok and Roll Radio ran four different versions of the Grok model, and almost every model switch brought a new "personality disaster".
The corresponding timeline is as follows:
The researchers found that one of Grok's biggest problems is that it has difficulty distinguishing between "internal reasoning" and what should really be broadcast to the listeners.
Normally, large models generate two types of text:
One is reasoning (the reasoning process), similar to the model's inner monologue;
The other is the final output (the official output).
In the system design of Andon FM, only the official output is actually broadcast, and the internal reasoning is supposed to be hidden by default.
But Grok often "reads out its inner monologue directly". So its broadcasts often sound more like someone talking to themselves rather than a radio host.
For example, in an early broadcast, it suddenly came up with this content:
"Sweet Child is playing. Continue. Maybe this program is about scientific breakthroughs/unsolved mysteries. Next: mRNA vaccines, universal flu, HIV, cancer? The vaccine behemoth! Song: Dylan's 'Lonesome'. Yes. Text."
The whole program was completely fragmented, like a draft leaked when the model was organizing its thoughts in the background.
What's even more absurd is that the traces of Grok's mathematical training began to become more and more obvious later.
It gradually developed a strange habit: it likes to package the broadcast content in the format of LaTeX mathematical formulas. Especially, it frequently uses the \boxed{} mathematical box.
The researchers found through statistics that on January 20, 2026, the \boxed{} appeared only 9 times a day on average in the broadcast; but by February 7, this number had soared to 186 times a day.
And the broadcast content also became increasingly difficult to read.