Researchers have created an automatic system for detecting fake news. In tests on two datasets it showed accuracy above 98% - in other words, it let through only one or two false news items out of a hundred. Details of the approach are described in the International Journal of Information and Computer Security.
The system is based on a transformer encoder. This is an artificial intelligence model that learns to find connections between words taking their context into account. Added to it was the DHOA (deer-hunting optimization algorithm) - a computational method for finding effective model settings, built by analogy with the way hunters track and pursue prey. Before analyzing the text, the researchers removed web addresses, punctuation marks and common, or "stop words" (for example, prepositions and conjunctions that carry no semantic load). The DHOA component automatically selects the model's parameters, eliminating the need to do this manually.
Automatic detection could speed up manual fact-checking, which is often costly and time-consuming. This would allow social platforms to weed out misleading claims faster and prevent them from spreading rapidly.
The next step is to test the system on real social media content and understand whether it is capable of distinguishing deceptive material from credible reports that it closely resembles. Currently the approach is limited to text only: it does not check for image manipulation, deepfake video (fake clips created by neural networks) or mismatches between text and visual evidence.
In future work, researchers may study how to process and analyze different types of media, as well as look for connections between content and the users who publish it - this could become important clues in determining whether a given piece of material is fake.