Characterising Deception in AI: A Survey

Abstract

In 2000, it was predicted that artificially intelligent agents would inevitably become deceptive. Today, in a world seemingly awash with fake news and in which we hand over control of our home environments to faux-human smart devices, it is timely to review the types of deception that have actually emerged. By reference to examples from diverse branches of AI, we classify research on deception into five novel categories, which we describe according to the human characteristic with which they most closely align: imitating, obfuscating, tricking, calculating and reframing. We offer this as a way for those within AI to recognise connections across the discipline and to suggest how AI-instigated..