Our laptop says the witness is unreliable

LAWYERS and judges use skill  and instinct to sense who might be lying in court. Soon they may be able to rely on a computer, too. An AI system trained on false statements is highly accurate at spotting deceptive language in written or spoken testimony. It can also be used to weed out fake online reviews of books, hotels and restaurants. The system is the work of computational linguists Massimo Poesio at the University of Essex in Colchester, UK, and Tommaso Fornaciari at the Center for Mind/ Brain sciences in Trento, Italy.  It is based on a technique called stylometry, which counts how often certain words appear in a passage. The method is often applied  to determine who wrote a piece  of text, but software can employ  it to pick out deception instead.  The strategy is to seek out the overuse of linguistic hedges such as “to the best of my knowledge”, or overzealous expressions such as “I swear to god”. “But all previous studies had used deceptive texts created in the lab,” Poesio says. “What has been missing was a system that could work on real-world lies.” So he and Fornaciari trained  a machine learning system by feeding it Italian courtroom depositions and statements  by defendants known to have committed perjury. The researchers say it is now nearly 75 per cent accurate at indicating whether a defendant or witness is being deceptive. “We can achieve an accuracy that is way above chance,” says Poesio. Sam de Silva, chairman of the London-based Law Society’s technology and law committee, “ What was missing was a system that could work on real-world lies rather than texts created in the lab” says that such a system could act as an aid to a counsel’s “gut feel” about a witness. “There’d be no bar on using it, but it depends how reliable it is and how much it will cost,” he says. The researchers are now feeding the system online book reviews known to be fake, in order to expose authors who have written fawning reviews of their own works. “We can now assess the likelihood a review is deceptive, again with an accuracy way above chance,” says Poesio, who will present the results at a conference on computational linguistics in Gothenburg, Sweden, next month.  Paul Marks  ■