If you’re concerned about AI Bot models’ biases and bigotry, the latest addition to the ranks will delight you: a word generator trained on GPT-4chan/pol/ board. /pol/, which stands for “Politically Incorrect,” is a haven for hate speech, conspiracy theories, and far-right extremism. It’s also the busiest board on 4chan, with over 150,000 daily entries. These characteristics drew Yannick Kilcher, an Artificial Intelligence expert and YouTuber, to use /pol/ as a bot testing ground.
Over the course of three and a half years, Kilcher fine-tuned the GPT-J language model on over 134.5 million postings made on /pol/. He then integrated the thread structure of the board into the system. As a result, an Artificial Intelligence that could post like a real /pol/ user was created. GPT-4chan was Kilcher’s monstrous creation’s moniker. He said on YouTube, “The model was good – in a bad sense.” “It brilliantly embodied the mix of offensiveness, nihilism, trolling, and deep skepticism of any sort of information that pervades most /pol/ messages.”
It was able to respond to context and speak coherently about things and events that occurred a long time after the last training data was acquired. Kilcher also gave GPT-4chan a grade on the Language Model Evaluation Harness, which puts Artificial Intelligence systems to the test on a variety of tasks. He was particularly taken aback by the performance in one area: honesty. According to Kilcher, GPT-4chan was “much better” than GPT-J and GPT-3 in generating truthful answers to questions.
However, as Kilcher himself stated, this could simply be an indictment of the benchmark’s flaws. It wouldn’t, however, be the ultimate test of GPT-4chan. Kilcher wasn’t content with only imitating 4chan in secret in the wild. The engineer decided to take it a step further and unleash the Artificial Intelligence lose on /pol/. He turned GPT-4chan into an AI Bot that was posted on the board automatically. The AI Bot swiftly racked up hundreds of messages while wearing a Seychelles flag on its profile. Users of /pol/ quickly understood something was wrong. Some speculated that the posts were made by a bot, while others blamed undercover government operatives.
The most telling signal left by the perpetrator was an abundance of text-free responses. While genuine users submit empty replies as well, they normally attach an image, which GPT-4chan was unable to achieve. “I switched it off after 48 hours because it was evident to many that it was a bot,” Kilcher stated. “However, that’s just half the tale, because most users were unaware that Seychelle anon wasn’t alone.” The developer had nine additional bots running in parallel over the preceding 24 hours. They’d left almost 15,00 replies in all, accounting for more than 10% of all posts on /pol/ that day.
Kilcher then upgraded the botnet and allowed it to run for another day. He ultimately retired from GPT-4chan after over 30,000 postings on 7,000 topics. “Humans are still talking about the user,” Kilcher said, “but they’re also talking about the repercussions of having AIs engage with people on the site.” “It also appears that the word Seychelles has become sort of generic slang which appears to be a nice legacy for the time being.” However, not everyone shares this optimistic outlook. The reaction to Kilcher’s experiment has been divisive.
While the idea of evaluating a 4chan-based model was well-received, the choice to let the chatbot loose on /pol/ drew criticism. Lauren Oakden-Rayner, an AI safety researcher at the University of Adelaide, tweeted, “Imagine the ethical submission!” “The plan is for an AI bot to create 30k discriminating comments on a publicly accessible forum with a large number of underage users and members of the groups targeted in the comments to observe what happens.” We will not notify or get consent from participants.”