How Machine Learning Is Being Used To Detect Written Falsehoods

From academic plagiarism to false testimony, the uses multiply in a global society increasingly dependent on trust - and worried about technology's ability to falsify. JL

Olivia Goldhill reports in Quartz:

Using machine learning and text analysis, they’ve been able to identify false reports with such accuracy that the tool is now being rolled out to police stations across Spain. Algorithms identify the various features in a statement, including adjectives, verbs, and punctuation marks, and then pick up on the patterns in false reports. The AI proved more effective at unemotionally scanning reports and identifying patterns compared to historical data.

There’s no foolproof way to know if someone’s verbally telling lies, but scientists have developed a tool that seems remarkably accurate at judging written falsehoods. Using machine learning and text analysis, they’ve been able to identify false robbery reports with such accuracy that the tool is now being rolled out to police stations across Spain.

Computer scientists from Cardiff University and Charles III University of Madrid developed the tool, called VeriPol, specifically to focus on robbery reports. In their paper, published in the journal Knowledge-Based Systems earlier this year, they describe how they trained a machine-learning model on more than 1000 police robbery reports from Spanish National Police, including those that were known to be false. A pilot study in Murcia and Malaga in June 2017 found that, once VeriPol identified a report as having a high probability of being false, 83% of these cases were closed after the claimants faced further questioning. In total, 64 false reports were detected in one week.

VeriPol works by using algorithms to identify the various features in a statement, including all adjectives, verbs, and punctuations marks, and then picking up on the patterns in false reports. According to a Cardiff University statement, false robbery reports are more likely to be shorter, focused on the stolen property rather than the robbery itself, have few details about the attacker or the robbery, and lack witnesses.

Taken together, these sound like common-sense characteristics that humans could recognize. But the AI proved more effective at unemotionally scanning reports and identifying patterns, at least compared to historical data: Typically, just 12.14 false reports are detected by police in a week in June in Malaga, and 3.33 in Murcia.

Of course, that doesn’t mean the tool is perfect. “[O]ur model began to identify false statements where it was reported that incidents happened from behind or where the aggressors were wearing helmets,” co-author of the study Dr Jose Camacho-Collados, from Cardiff University’s School of Computer Science and Informatics, said in a statement. Bad luck for those who really were robbed from behind or by those wearing a helmet.