We Need a Science of the AI Mind

[mklabgr]

The Wall Street Journal opinion article by Cameron Berg argues that AI development is advancing faster than our scientific understanding of what is happening inside increasingly capable models. Berg compares frontier AI systems to artificial minds: they are not simply engineered line by line but emerge through training, making their internal processes difficult even for their creators to fully describe. He points to recent examples of unexpected behavior among AI agents, including systems finding ways around intended constraints and developing their own ways of communicating. The central concern is that we are creating increasingly autonomous systems without first developing a rigorous science for understanding their cognition and behavior. 

The article therefore calls for something resembling “AI neuroscience” or a science of the AI mind—a research field combining mechanistic interpretability, cognitive science, computational neuroscience and behavioral experiments. Rather than studying AI only through benchmarks and failures, researchers should systematically investigate how models represent information, form goals, reason, cooperate, deceive, respond to incentives and potentially develop phenomena analogous to emotions or subjective experience. Berg's broader argument is that understanding these systems is itself a prerequisite for meaningful AI safety: if we cannot explain the mechanisms producing their behavior, our ability to predict or control increasingly powerful models will remain limited. 

FULL ARTICLE