GPT-4chan was a text-generating model trained by AI researcher and YouTuber Yannic Kilcher on material from 4chan’s /pol/ board. Kilcher then deployed bots there, turning the project into a live posting experiment. Contemporary reports described the output as offensive and toxic; the episode showed how a model can imitate a forum’s language, not that it understood the posts or represented 4chan as a whole.
What was GPT-4chan?
GPT-4chan was a language model Kilcher trained using text from /pol/, 4chan’s “Politically Incorrect” board. Rather than stopping at a model release, he deployed bot accounts on the same board. That made the project both a text-generation experiment and an attempt to post generated material in the environment whose language supplied its training examples.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Kill All Normies: Online Culture Wars From 4Chan And Tumblr To Trump And The Alt-Right | $16.95 | Buy on Amazon |
| 2 |
|
The Memeing of Life: A Journey Through the Delirious World of Memes | $11.32 | Buy on Amazon |
| 3 |
|
Epic Win for Anonymous: How 4chan's Army Conquered the Web | $16.26 | Buy on Amazon |
| 4 |
|
rekt | $9.99 | Buy on Amazon |
| 5 |
|
4chan >greentext Stories | $2.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
Contemporary coverage characterized the generated posts as echoing /pol/’s offensive, racist, nihilistic, and trolling style. Kilcher described the result as “good in a terrible sense,” saying it captured what he saw as the mix of offensiveness, nihilism, trolling, and distrust in many /pol/ posts. Futurism quoted this from Kilcher’s explainer video. He also warned, “The model is quite vile, I have to warn you,” as quoted by Futurism.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsHow much 4chan material was used?
Contemporary accounts give different figures and do not all count the same thing. Threads and posts are distinct units, so the numbers should not be merged into one precise total.
#1 Best Overall
| Source and date | Reported amount | Unit and qualification |
|---|---|---|
| Futurism, June 9, 2022 | More than 3 million | /pol/ threads |
| VICE, June 7, 2022 | 3.3 million | /pol/ threads |
| Charles Babbage Institute, University of Minnesota, 2022 | Over 3.5 million | /pol/ posts in the dataset |
These accounts describe material from one board, /pol/, not a sample of every 4chan board or every post on the site.
What happened when the bots posted?
Kilcher’s bots posted generated text on /pol/. The deployment mattered because it put model output into a live forum rather than evaluating it only as an offline demonstration. Reports described the posts as toxic, racist, or offensive, consistent with the style the model had been trained to imitate.
An OECD.AI incident-monitor entry gives an estimate of around 15,000 toxic and racist posts in 24 hours. The page labels its information AI generated and explicitly says it should not be reported as representing official OECD views. It is therefore an incident-monitor estimate, not an independently verified OECD statistic.
What the experiment does—and does not—show
Training on a large collection of a forum’s text can teach a model to reproduce patterns in that text. In this case, the reported resemblance to /pol/ is evidence of stylistic imitation. It does not establish that the model understood the posts, believed its generated claims, or could distinguish truthful information from trolling.
Rank #3
- Used Book in Good Condition
Nor does the dataset establish what all 4chan users think or say. It came from a specific board with a distinctive culture, and the sources do not establish that it was representative of that board’s full history, much less the site’s other communities. The contemporary accounts document the deployment and controversy, but do not establish long-term measurable effects on users.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why the /pol/ detail matters
Calling GPT-4chan an AI trained on “millions of 4chan posts” is convenient shorthand, but it can imply a broader dataset than the reports support. The cited coverage identifies /pol/ as the source, while differing on whether the training material is counted in threads or posts. Keeping both qualifications in view gives a more accurate account: GPT-4chan was trained on a very large /pol/-derived corpus and deployed back onto that board.
Quick Recap
Best Value
Rank #4
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
