Are We Thinking Correctly About AI Intelligence?

Whether LLMs reason like humans or merely generate plausible text when answering questions is a core issue determining AI reliability and oversight levels.

Melanie Mitchell of the Santa Fe Institute argues that AI is closer to an 'alien intelligence' with cognitive mechanisms different from humans. She proposes applying methods from psychology, used to study infant and animal cognition, to better measure machine cognition.

Mitchell presents six principles for evaluating machine cognition, discussing a wide range of topics including the difficulty of interpreting internal system processes, recent AI-assisted mathematical breakthroughs, and lessons from the case of a horse proficient in mathematics in the early 1900s.