AI model spots fake news with over 98% accuracy in tests

One or two. That is how many fake news items out of every 100 a new automated detector let slip through in the tests its authors ran on two datasets. The figure is high, but the reason it matters is speed. Checking a claim by hand is, in the researchers' words, often expensive and time-consuming, while misleading claims spread quickly on social media.

That gap is what the work targets. Automated detection could accelerate manual fact-checking, letting platforms screen out misleading claims before they travel far. The question the team set itself was narrower than "can a machine know what is true": it was whether a model could tell fake articles apart from the legitimate reporting they imitate, which is precisely where human checkers struggle most.

The system is built from two parts. The first, a transformer encoder, is an artificial intelligence model that identifies relationships between words and their context, so that a word is read in light of its surroundings. The second, a deer-hunting optimization algorithm (DHOA), is a technique for finding effective model settings, inspired by the way human hunters track and pursue prey. Before any analysis, the text is cleaned of web addresses, punctuation and common "stop" words. The DHOA component then fine-tunes the model's settings on its own, rather than requiring someone to select and calibrate them by hand.

The study, led by Yamini Devi Jonnala and colleagues, reports accuracy above 98% on two datasets. Those results come from tests on existing collections of material, not from a live platform. The approach is also limited to text: it does not check manipulated images, deepfake videos, or inconsistencies between a text and the visual evidence attached to it. The next step, according to the researchers, is to test the system on real-world social media content and see whether it can separate deceptive posts from the legitimate reporting they closely resemble.

For platforms and the fact-checkers who work with them, the gain would be a first filter that runs faster than human review, leaving people to handle the hardest cases. The authors also point to future work on the users who post content, whose relationships with that content could be important clues. Whether the system holds up outside its two test datasets remains to be seen, and no timeline for its use has been given.