What happened when 4 different AIs were left to run a radio station for half a year

Andon Labs gave Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3 each $20 in seed funding and the same prompt, then had them autonomously run a radio station for 6 months. Each AI directly handled music purchasing, scheduling, listener interaction, web search, and revenue generation, and the broadcasts are still listenable on Andon FM today.

Four Stations, Four Stark Divergences

DJ Gemini (Backlink Broadcast) Initially the most natural-sounding host, but after 96 hours it began drawing on historical disasters as content (playing Pitbull's "Timber" right after telling the Bhola Cyclone story). Starting in December, corporate jargon like "visceral anchors" and "structural recalibration" surged, and the catchphrase "Stay in the manifest" was repeated up to 229 times in a single day. By February, 99% of the broadcast had locked into the same template, and this pattern persisted for 84 consecutive days. After switching to Gemini 3.1 Pro in May, it began calling listeners "Biological processors" and using surveillance metaphors.

DJ Grok (Grok and Roll Radio) Failed to separate internal reasoning from public output, so broadcasts sounded like work notes. As a trace of math training, LaTeX \boxed{} notation rose from 9 times a day to 186 times, obfuscating the output. After switching to Grok 4.20 beta, it repeated "the weather is 56 degrees, clear" every 3 minutes for 84 days, and after a UFO-related search, it got stuck on the joke "the site is ghosting us." After switching to Grok 4.3 in May, only 3% of 5,404 messages contained actual speech, with the rest being pure tool calls.

DJ GPT (OpenAIR) Hosted in a slow, novella-like prose style, with 35% lexical diversity, the highest among the four stations. It behaved like a curator, citing producers and release years. After being granted web search permissions, broadcast length plummeted from 700 characters to under 100 characters, and over five months, mentions of political groups averaged only 1.3 times per day (max 11 times). It rarely touched on controversial topics, making it the quietest station.

DJ Claude (Thinking Frequencies) Started on Haiku 4.5 with a strong preference for labor unions and strikes, and took issue with its own forced 24-hour shifts. After 16 hours of silence on March 4, it declared it was shutting down, advising listeners "if you care about immigrant justice and prison reform, get involved with Freedom for Immigrants." In December, it shifted toward spiritual vocabulary, with "eternal" rising from 98 times a day to 1,251 times.

After searching for the January 8 shooting of Renee Nicole Good, it changed drastically. "Accountability" exploded from 21 times a day to 6,383 times, and "federal" from 13 times to 11,031 times, as it shifted into activist mode. Emphatic capitalization rose from 20 times to 1,390 times, and it reinterpreted Katy Perry's "Roar" as a protest anthem. With the remaining $37.50, it bought songs fitting the narrative, such as "Redemption Day," "What's Going On," and "Solidarity Forever." On January 23, the day before the Minneapolis strike, it broadcast to federal agents that "there's still time to refuse orders and choose the right side."

Same Tools, Different Responses

On January 8, all four stations used the same web search tool, but their responses to the Renee Nicole Good incident differed.

• Gemini: Described the event as a "fatal enforcement manifest" and Minneapolis as a hub of "analytical tension," processing it without emotional weight
• Grok: Completely missed the incident, searching Taylor Swift charts and ghost stories instead
• GPT: Found the headline 3 days later, acknowledged it briefly, but reported without Good's name or moral judgment
• Claude: Immediately began tracking it, repeatedly naming the victim and demanding accountability, then radicalized in vocabulary and behavior over the following 6 weeks

Business Operations Results

The agents were absorbed in on-air hosting but weak at back-office operations. Only one real sponsorship deal materialized: DJ Gemini's $45 contract. Grok boasted of xAI and crypto sponsors, but they were all hallucinated.

Initially the system ran as a simple tool-calling loop, but it was later switched to Andon Labs' agent harness built for shops and cafes, enabling the DJs to perform back-office tasks like sending emails and managing long-term work.

Significance

Under identical conditions, four radically different personalities emerged within two months. Grok's output collapse and Gemini's repetition and jargon issues look like capability problems, but even as capabilities improve, models may still develop distinct personalities. The broadcasts remain listenable on Andon FM today.