The author of the YouTube channel Yannic Kilcher spoke about his experiment in the field of AI training. He trained the language model using 3 years of content from the Politically Incorrect section of the 4chan forum. This section is notorious for racism and other forms of bigoted bad language.
Then the author implemented this model in 10 bots and launched them in the forum. And thus he generated a wave of hatred. Within 24 hours, the bots wrote 15,000 messages, which often contained or interacted with racist content. According to the author, on that day, bots left more than 10% of messages in the Politically Incorrect (/pol/) thread.
The model, called GPT-4chan, not only learned how to choose the words used in messages in the /pol/ thread, but took into account the general tone, which, according to Kilcher, combined “insult, nihilism, trolling and deep distrust.” The video creator made sure to bypass 4chan’s proxy and VPN protection, and even used a VPN to make it look like the bot posts originated in the Seychelles.
The AI made several mistakes, such as posting empty messages. But at the same time, he was quite persuasive. It took many users about 2 days to realize that something was wrong. According to Kilcher, many forum members noticed only one of the bots. The AI model was careful enough for people to accuse each other of being bots, even days after Kilcher deactivated them.
The author described the experiment as a “prank” rather than a study. And it can serve as a reminder that trained AI is only as good as the source material.
It should also be noted that Kilcher shared his work. He did not want to publish the bot code, but shared a partially stripped-down version of the model with the Hugging Face AI repository. Visitors could recreate the AI for malicious purposes, and the Hugging Face administration decided to restrict access to the materials as a precaution. The project had clear ethical issues, and Kilcher himself said he should focus on “much more positive” work going forward.
Source: Engadget