Rather than fact-checking every piece of fake news, researchers are teaching strategies that recur across different pieces of fake news. When used as language models, humans are better at distinguishing truth from falsehood on unfamiliar topics.
Faced with a generative AI capable of writing untruths as persuasive as those of humans, case-by-case fact-checking no longer follows, and training courses that teach to doubt have the disadvantage of also making people doubt the truth. Researchers propose another defense: targeting not the content of each statement, but the persuasion strategies that come back from one to the other.
They mapped this "manual" of disinformation from 214,646 verified claims in 51 languages. It covers social influence, information manipulation, and appeals to emotion, organized around a core of fear, conspiracy, and fabrication. Briefed with this manual — and a reminder that persuasion does not mean lying — a language model better distinguishes between true and machine-generated false claims about events after its training.
A real but targeted effect
In a pre-recorded experiment, the same briefing improved the participants' discernment of the statements they said they did not know, without undermining their confidence in the real ones. On all the statements, however, the gain was not statistically significant. Its advantage: it did not need to wait for a specific verification to be published.
Read the full article on the original source
Go to the source →