Predicting Criminal Behavior

Although it has not been officially published, a recent study on the use of facial recognition to distinguish criminals from law-abiding citizens is fueling controversy over appropriate applications for machine learning. The report, which appeared in an online, open-sourced journal, claims that the software was able to correctly identify criminals from photos with 90 percent accuracy by analyzing three facial features: upper lip curvature, inner eye corner distance and nose-to-mouth angle.

The artificial intelligence system was first trained with 1,670 pictures of Chinese men, half of whom were convicted criminals. It was then presented with an additional 186 photos and asked to identify criminals and non-criminals. Based on this, the researchers determined that criminals have a greater degree of dissimilarity in facial features than non-criminals. In addition (supposedly), on average, the nose-to-mouth angle is 19.6 percent smaller, lip curvature is 23.4 percent larger and inner eye corner distance is 5.6 percent shorter in criminals.

As with previous attempts to judge a person’s character from appearance alone (e.g., physiognomy in ancient Greece and craniometry in 19th-century Britain), the results of this study are dubious, and the danger of using such technologies for “predictive policing” may far outweigh the benefits. More research is needed before tools like this are widely deployed.