Melanie Mitchell Sets 6 Principles for Assessing AI Cognition

1 articles · Updated · Quanta Magazine · Aug 20

Mitchell argued that today’s AI lacks adequate tests of cognition, saying large language models may produce convincing reasoning-like text without reasoning in human ways.
Her six principles call for guarding against anthropomorphic bias, using control experiments, varying benchmarks to test robustness, probing models’ internals, separating performance from competence, and studying failures rather than hiding negative results.
A Clever Hans-style warning runs through her framework: AI can appear to solve a task while exploiting hidden cues, as in benchmark designs where models answered diagram questions even without seeing the diagrams.
Mitchell said AI should be treated as an “alien intelligence” and studied with methods borrowed from developmental and comparative psychology, while mechanistic interpretability could become a neuroscience-like tool for understanding how models actually work.