Fake news detection systems now use AI to label training data instead of hiring humans to do it

A new method uses large language models to cheaply label examples of fake news across different domains, with human reviewers checking the AI's work. This means fake news detection systems can be trained on much larger datasets without the cost of manual labeling, making cross-domain detection (where a system trained on political fake news can spot health misinformation) actually feasible instead of theoretically interesting.

Fake news detection has always been trapped by the labeling problem: you need thousands of examples marked as real or fake to train a system, but hiring people to read and label all those examples is expensive and slow. This paper shows that AI can do the labeling work, checked by humans, at a fraction of the cost. What becomes possible is systems that work across domains instead of being trained narrowly on one type of fake news — a political misinformation detector that also catches medical hoaxes, not a tool that's useless the moment the target switches.