AI Can Now Fake Entire Reddit Threads — And You Can't Tell

A team of researchers asked a simple question: can AI pass a Turing test — fooling people into thinking it's human — when it isn't one person chatting, but an entire group? They collected real conversations from Reddit, then used two popular chatbots, Llama 3 and GPT-4o, to generate brand-new conversations on the same topics. Volunteers then read both versions side by side and had to guess which was written by actual people.

They couldn't. Across the board, people misidentified AI conversations as human 39% of the time. With Llama 3, the error rate climbed even higher: volunteers correctly spotted the AI only 56% of the time — barely better than a coin toss. If you'd flipped a coin to decide, you would have done almost as well.

There's an upside. Realistic AI chatter could help researchers study how social platforms work without experimenting on real people. You could simulate a community, tweak the recommendation algorithm, and watch what happens. That's cheaper, faster, and ethically cleaner than testing on live users.

But the same trick cuts the other way. If AI can spin up believable conversations, so can anyone with an agenda. Political operatives, marketers, and scammers could flood forums with fake consensus — making an unpopular idea look popular, or a mediocre product look beloved. The researchers call their finding a 'warning message.' The takeaway for you: what looks like a crowd agreeing online may have no crowd behind it at all.