Prediction Machines – Comparative Analysis of Theory of Mind Abilities in Machine Learning Models

Abstract

With the advent of Machine Learning models in the current global markets, questions pertaining to artificial mental states and Artificial General Intelligence (AGI) come to the forefront. The current study seeks to find support for possible mental states by subjecting three different Machine Learning models – GPT 3.5 (Generative Pre-Trained Transformer), Bard, and GPT 3 – Ada to psychological tests concerning Theory of Mind (ToM). To discern another’s mental state one must be capable of distinguishing oneself as an entity separate from the other. The three models were presented with the Strange Story Task and the Theory of Mind Scale. The idea of artificial ToM by observation and prediction has been posited. As hypothesized, GPT 3.5 significantly outperformed GPT 3 (Ada) and failed to outperform Bard. Further implications and limitations have been discussed.