Leggi in italiano

Current lie detection systems, such as the polygraph, are unreliable for use in a legal setting. Could AI replace them? Credit: Getty Images.

A research group from the IMT School of Advanced Studies Lucca, and the University of Padua, has developed an artificial intelligence algorithm that can, to a certain extent, identify lies in written texts. The algorithm, based on a large language model, has achieved an accuracy level of 80% in distinguishing truthful stories from false ones, beating the average performance of human intuition. The results are published in Scientific Reports1.

The human ability to recognize lies is limited, in controlled tests achieving around 50% accuracy. Techniques such as the polygraph, often fail, and many agencies do not recommend to use the in the legal field. “But IT models are already used in specific sectors, for example to identify false reviews online,” says Giuseppe Sartori, professor of forensic neuropsychology at the University of Padua, and co-author of the work.

The authors started with a language model called FLAN-T5, similar to GPT, and trained it with a database of true and false narratives compiled by asking hundreds of participants to respond both honestly and falsely to questions on personal opinions, autobiographical memories and future intentions.

The results showed an average accuracy in detecting lies of 80%, with a better performance in unmasking false opinions. The authors concede that, because it was only tested in a laboratory setting with fabricated texts, the reliability of the algorithm is still limited. “We are still far from practical use in the legal field,” Sartori says “but we are confident that we will be able to get closer to this in the future by expanding studies and increasing the amount of data used”.