GPT-4chan
A GPT-J model fine-tuned on posts from 4chan's /pol/ board.
GPT-4chan is a large language model built by AI researcher and YouTuber Yannic Kilcher, released in June 2022. It was created by fine-tuning the GPT-J model on millions of posts from the /pol/ board of 4chan, and it sparked debate after being deployed on that forum and shared publicly.
- Fine-tuned from GPT-J on millions of /pol/ posts from 4chan
- Released 3 June 2022 by Yannic Kilcher
- Ran as an unmarked bot on /pol/ with no human oversight
- Weights published under Apache-2.0, later removed from Hugging Face
- A petition against the deployment drew 300+ expert signatures
What it is
GPT-4chan is a large language model that generates text from a prompt. Yannic Kilcher, an AI researcher and YouTuber, developed and released it in June 2022. The system is a fine-tuned version of GPT-J, an open model from EleutherAI. Training used millions of posts taken from /pol/, a board on the anonymous forum 4chan. That board is known for hosting hateful and extremist material. The resulting model copies the tone of those users, and its output is often deliberately offensive towards groups, including racist, sexist and homophobic text, along with nihilistic views. Kilcher put it live on /pol/ itself and also published it for others to use.
How it came about
Work started in May 2022, when Kilcher described the project on his YouTube channel. This was before ChatGPT existed. He said he wanted a model able to write realistic, coherent text in the voice of /pol/. He pointed to GPT-3 from OpenAI and the open GPT-J from EleutherAI as inspiration, and picked GPT-J as his starting point. He then adapted it using a /pol/ post dataset that had already been published. In June 2022 he released the model both as a bot on the forum and as a public download.
Inside the model
The base system is GPT-J, a generative pre-trained transformer released by EleutherAI. Kilcher fine-tuned it on a previously published collection of /pol/ posts, numbering in the millions. Fine-tuning means further training an existing model on new data so it picks up that data's style. Here the model learned to imitate how /pol/ users write. Sample outputs shown during development covered political opinions, conspiracy theories, jokes, insults and threats, as well as poems, stories, songs and code. The weights were published under the Apache-2.0 licence, with code in Python on GitHub.
What it can do
The model writes in the style and tone of /pol/ users. Its replies can be racist, sexist, homophobic or otherwise offensive, and often nihilistic. Examples shown by Kilcher ranged from political claims and conspiracy theories to jokes, insults and threats, and also more creative pieces such as poems, stories, songs and code. He said he was impressed by how fluent and diverse the writing was. On the forum, some users rated it as clever, funny or agreeable, while others called it ignorant, inconsistent or absurd.
How to use it
Kilcher hosted the model on Hugging Face, where people could try it through a web interface or an API. He also linked a GitHub repository holding the code and data. The weights carried an Apache-2.0 licence. Access on Hugging Face was later gated and then switched off because of worries about possible harm. A bot version also ran on the /pol/ board itself.
Safety and testing
Kilcher ran the forum bot with no human oversight, treating it as a natural experiment to watch how the model behaved in a live setting and how it coped with trolling, flaming, baiting and moderation. On Hugging Face, access was first restricted and then turned off over concerns about the damage the model could do. That step involved direct action by CEO Clément Delangue on the talk pages, which was unusual next to the site's normal moderation routine.
Reception and debate
The project drew criticism and argument within the AI community over its ethics, legality and social effects. On /pol/, users argued over the bot's posts, some praising it and others attacking it, and some trying to trick or expose it. A petition against the deployment collected more than 300 signatures from technology experts. The release also drew media coverage and attention on YouTube. Debate focused on the risk of spreading hate speech and on who carries responsibility for such models.
Why it matters
The case became a reference point in arguments about harmful model releases. It fed into discussion of hate speech risk, of the duties of model makers and hosting platforms, of how such systems should be regulated and overseen, and of what open publication and transparency mean for AI research. Kilcher said he wanted to start that discussion and welcomed feedback and criticism.
Questions people ask about GPT-4chan
What is GPT-4chan?
It is a large language model made by fine-tuning GPT-J on millions of posts from /pol/, a board on the anonymous forum 4chan. Yannic Kilcher built and released it in June 2022. It writes in the style of that board's users, and its output is often deliberately offensive.
How was it trained?
Kilcher started from GPT-J, an open model from EleutherAI, and further trained it on a previously published dataset of /pol/ posts. This fine-tuning taught it to copy the tone of those users. He showed outputs spanning political claims, conspiracy theories, jokes, insults, threats, poems, stories, songs and code.
How could people use it?
It was posted on Hugging Face with a web interface and an API, and the code sat in a GitHub repository under an Apache-2.0 licence. A bot version also posted on /pol/ without saying what it was. Hugging Face later gated and then disabled access over harm concerns.
Why did it cause argument?
Its writing is often racist, sexist, homophobic or nihilistic, which raised questions about spreading hate speech and about who is responsible for such a model. A petition from technology experts against the deployment gained over 300 signatures, and the case fed wider debate on oversight and openness in AI research.
Good reading and sources
- YouTuber trains AI bot on 4chan’s pile o’ bile with entirely predictable resultsThe Verge, Article
- AI bot trained on 4chan posts misbehaves like 4chan usersThe Register, Article
- ykilcher/gpt-4chan · Hugging FaceHugging Face, Article
- Moderating Model Marketplaces: Platform Governance Puzzles for AI IntermediariesarXiv, Article
The facts on this page come from the Wikipedia article "GPT-4Chan", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.