Artificial intelligence in lie detection: Why do cognitive theories matter?

Abstract

In the field of psychological research on deception detection, the rise of machine learning algorithms, or artificial intelligence (AI), has sparked discussions about its potential benefits and risks. Most researchers argue strongly for the inclusion of good theories in the design, training, and application phases of AI. In this letter we ask an important follow-up question: what makes a good theory? And why do they matter in detecting deception? To this end, we argue that mechanism-driven, and cognitively-informed, theories are the ones AI researchers need to be looking for. This is particularly important in deception detection, where false positives and negatives can result in irreversible legal consequences. Crucially, mechanism-driven theories allow us to know 1) what features are we extracting (e.g., a particular facial expression in time, higher peak in P300, etc.)? And importantly, 2) how are these features related (subordinate or superordinate) to the entity we are inferring (e.g., memory recognition, anxiety, or deception)? Answers to these questions can help forensic experts anticipate where the majority of our AI's mistakes are (i.e., false negatives or false positives), and allow nonexperts such as policymakers to adjust decision-making criterion to compensate for such errors via legal or other means if needed be (e.g., a more liberal criterion for detecting deception during the investigation phase, but later switches to conservative criterion in court). These logical inferences all start from mechanistically and cognitively-informative theories.