A general rule of thumb online is not to engage with bots. But what if the bot you're chatting or sparring with on the web was built using your own words? Would you even notice you were chatting with an artificial intelligence?
As Vice reports, this scenario played out in a rather disturbing fashion recently when a spate of messages appeared on the popular 4chan board Politically Incorrect, or /pol/. In over 24 hours, a mysterious poster dropped more than 15,000 messages.
As YouTube Yannic Kilcher described it in a recent video, "people started to notice there was something strange about this user. Some people loved him, as they agreed with his opinions. Other people hated him, as he appeared to be everywhere.”
Turns out, this user was Kilcher himself. He created a bot that ingested 3.5 years worth of posts from /pol/ and unleashed it via nine bots. It wasn't pretty. The forum is widely known for hate speech, and offensive material quickly proliferated. (As The GuardianThe Guardian reports, the shooter from last month’s massacre in Buffalo posted a 180-page manifesto containing ample plagiarized content from the board.)
Kilcher created the bot in just a few weeks, and his newly minted army of posters scanned the dataset every five minutes, generated unique comments, and published them under the same anonymous account. Given that AI is only as smart as the content it's trained on, /pol/ was quickly flooded with even more racist, misogynistic, and antisemitic content than usual.
"There were a base of users recognizing the bots for being bots, [but] there were still plenty of other users who didn't," Kilcher says, even after he admitted to being behind it.
While 4Chan typically requires users to complete a CAPTCHA to prove their humanity, Kilcher bypassed it by getting 4Chan Pass, a $20/month premium subscription that lets members to skip the CAPTCHA. 4Chan Pass also allows posting from a proxy server, which is usually not permitted. Kilcher’s bot ran on one based in the Seychelles, a tiny island nation off the coast of Eastern Africa.
After 24 hours, Kilcher took the bot down and outed himself on the platform as the creator. He then tweaked it and put it up again for a final 24 hours. Across the two days, the bot created over 30,000 posts, Kilcher says.
He also posted the model to AI community Hugging Face; the site later blocked access to it. That prompted concern from AI experts who feared it would be used by people with more nefarious intent than a YouTube prankster. Kilcher seems unbothered:
The move may remind some of Microsoft's short-lived AI chatbot Tay, which got real racist, real quick after observing the behavior of people it was chatting with online. The difference there is that people quickly noticed Tay had gone off the deep end; on /pol/ it was the volume of content that caught people's attention versus what was actually being said.