AI Can Fake News Posts With Photos — but Can't Spot Them

A team of researchers asked a blunt question: how easy is it for AI to mass-produce fake news posts — the kind with a headline plus a matching photo — and can other AI catch them? To find out, they built a small "newsroom" of three AI helpers working together: one writes the story, one makes the image, and one plays critic, checking whether the post sounds believable. The trio churned out more than 9,000 fake posts.

The fake posts covered science, health and entertainment, and each was designed to contradict a real news story — the sort of thing you might scroll past on Facebook or X. Then the team put 16 leading AI models to the test as fact-checkers. The verdict was grim: most scored well below human accuracy, and they failed badly on the hardest part — judging whether a photo was real or AI-generated. Reading the text is one skill; trusting the picture is another.

Why you care: this is a preview of the next wave of online misinformation, where the image is the convincing part. Right now, the tools meant to flag fakes aren't reliable enough to lean on. The practical takeaway is not to outsource your judgment. Reverse-image search a photo before you share it, check whether a claim appears anywhere beyond a single post, and treat suspiciously perfect pictures with extra suspicion — especially when they make you angry.

The study's own limits are worth knowing. This is lab-made fake news, not a survey of what's actually spreading, and AI-written posts may read differently from the messy stuff real people share. Still, the researchers released their code and data so others can build better defenses. The paper was accepted to EMNLP 2026, a major language-technology conference — a sign the field now treats fake-news detection as an urgent, unsolved problem rather than a side quest.