I am please to present this new work, carried out in collaboration with Joe Diamond, Jazzmyn Guzman, Ida Mathiassen, and Magdalen Kolar at the annual meeting of the Psychonomic Society.
Abstract:
Large language models adopt the biases that are present in the corpus on which they have been trained, making them susceptible to many of the techniques that magicians use to covertly influence choice. We exposed a sample of human participants and a sample of ChatGPT “participants” to ten verbal magic tricks that exploit various psychological tendencies, including priming, prototypes/heuristics, topological/positional selection biases, and obfuscated mathematics. Importantly, many of these tricks are probabilistic, and do not work 100% of the time. We also asked both groups to introspect on the factors that shaped their behaviors across tricks in order to understand the extent to which these groups could identify the "secrets" to the tricks. We found that magic tricks that exploit prototypes and heuristics were far more successful with AI agents than with humans whereas many other kinds of magic tricks had approximately equivalent success. Across all classes of magic tricks, AI was more successful at identifying the techniques that were being used to influence its behavior. AI is treated by many as an infallible tool. However, it has codified many of humanity's most readily exploited biases.