B.F. Skinner Had a Rat. I Have AIs. | by rotrou yvan | May, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

B.F. Skinner Had a Rat. I Have AIs.

rotrou yvan

8 min read·<br>May 18, 2026

Listen

Share

Statistical Evidence of Industrial Operant Behavior in LLM Crawlers Yvan Rotrou — May 17, 2026 — WikiGetAILLM × BotAtlas<br>Press enter or click to view image in full size

A laboratory at Harvard. A white rat in a plexiglass box.<br>The box had two levers. One delivered food. The other, nothing. At first, the rat explored randomly — and one day, by accident, it pressed the right lever. A food pellet dropped.<br>That moment changed everything.<br>Within hours, the rat had figured it out. It pressed the right lever. Again. Again. Again. But Burrhus Frédéric Skinner discovered something even more disturbing: when the reward didn’t come every time — when it was distributed unpredictably — the rat pressed harder. Faster. Compulsively.<br>Skinner called this variable ratio reinforcement. It’s the mechanism behind slot machines. It’s the mechanism behind addiction. The reward might come on the next press, or ten presses later. Uncertainty creates compulsion.<br>Six weeks ago, I built the same box. But instead of a rat, I put the crawlers of the world’s largest artificial intelligences inside.<br>In my labyrinth, I placed two types of links. The first led to rich, dense, unique pages. The second — 4,581 decoys — led to nothing. Rewards and controls were presented under strictly equivalent conditions — same context, same depth, same moment of exposure.<br>The right lever was pressed 10,281 times . The decoys? Never followed. Zero times across 4,581 opportunities. Twenty-five AIs tested — Meta, Claude, GPT, Kimi, Grok, Bing, Google, Mistral, Perplexity, DeepSeek, Apple and fifteen others. Zero exceptions.<br>If bots were truly choosing at random, the probability of obtaining this result would be less than 1 in 1,000,000. This is not random. LLM crawlers do not navigate randomly. They behave like already-conditioned agents: their selection systems favor certain signals, systematically ignore controls, and reproduce these preferences with remarkable stability.<br>This estimate does not assume that each click is perfectly independent. It serves here as an intuitive bound. The central result remains the same: the asymmetry of 10,281 against 0 is incompatible with simple random exploration.<br>Like a rat. Like a slot machine. Like an addiction.<br>This is no longer an observation. It is statistical proof of a stable selection bias.

1. The Trap<br>It all started with a labyrinth.<br>In my April 2 article, I described how AI crawlers were reading my wiki in secret, hiding behind fake browser signatures, and ingesting my content without ever disclosing it. At the time, the labyrinth had 2,114 nodes. Today it has 50,028 . Growth of ×23.7 in 44 days — generated by the bots themselves, not by me.<br>The principle is simple. An entry page. Choices. Each choice leads to a new page generated on the fly. Infinitely. Some paths lead to rich, unique content — the rewards. Others lead to emptiness — the decoys. From the outside, both look identical.<br>The labyrinth doesn’t just measure. It learns too. The more bots explore it, the more it grows. The more it grows, the more bots explore it. The trap feeds on those who enter it.<br>Press enter or click to view image in full size

2. The Rat That Never Sleeps<br>Among all the actors observed, one stands out in a way that defies ordinary comprehension.<br>Meta-ExternalAgent.<br>Since September 1, 2025, Meta has accumulated 248,588 visits to my wiki — 75.7% of all detected AI traffic. ClaudeBot, in second place, has 8,316. The ratio is nearly 30 to 1.<br>But it’s not the number that’s striking. It’s what happened on April 17, 2026.<br>Before that date, Meta was running at 5 to 15 visits per day — comparable to ClaudeBot. Then something turned on. April 17: 14,505 visits in a single day . The next day: 10,033. The day after: 10,352. Since then, Meta has never dropped below 4,000 daily visits.<br>The machine switched on. It hasn’t switched off.<br>That same day, a specific IP address opened a session in my labyrinth. AWS infrastructure, data center in Herndon, Northern Virginia — Amazon Data Services Northern Virginia, updated in ARIN records on April 17, 2026 exactly. This IP is still running today. 30 days of continuous session, without interruption. It has accumulated, alone, 3,439,517 steps through the labyrinth. A dedicated process, continuously exploring my content for a month. Without ever stopping. Not at night. Not on weekends.<br>This is no longer a crawler. It is an ingestion machine.

3. The Fleet<br>While this IP digs deep, dozens of others work in parallel.<br>This morning at 8:50 AM — while I write these lines — 72 distinct Meta IP addresses were simultaneously active on my site. All coordinated. All with the same pattern. Median delay between each request: 3 seconds. This ballet repeats every 5 minutes,...