AI is making disinformation harder to spot – but we’ve found a new way to catch it

Generative AI has made disinformation much harder to spot. Old detection tricks—spotting grammatical errors, unusual word choices—no longer work, because AI text now sounds completely natural. Researchers at the University of Sheffield took a different approach: instead of looking for AI fingerprints, they built a system to detect when comments are deliberately derailing a discussion. They analyzed over sixteen hundred comments under BBC News videos and found that malicious posts consistently steered conversations away from the original topic, using red herrings, logical leaps, and personal attacks. Their AI system compared each real response to what a typical, on-topic response would look like, and flagged the ones that unexpectedly changed direction. In testing, it correctly identified derailing comments seventy-seven percent of the time—roughly twice as effective as existing detection methods. The Conversation reports that researchers stress this works best as a flag for human moderators rather than a replacement for their judgment, though questions remain about how well AI systems truly understand human discussion.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton