Stephen Downes

A lot of cognitive science takes the form "humans must have [cognitive feature X] because that's the only way to do [cognitive function Y]." But what if artificial intelligence can also [do cognitive function Y]? This study looks at whether that's the case, and argues that this alignment is 'jagged': "LLMs can show surface-level behavioral convergence, process-level similarities, cross-task generality, and representational alignment, while also exhibiting some capability, distributional, metacognitive, or super-human forms of divergence." Still, the challenge for cognitive science is there. "The field has long been willing to draw strong inferences about human mental representations and processes from behavioral evidence alone; the arrival of systems that can reproduce much of that behavior - sometimes through divergent means - forces us to re-examine when such inferences are warranted and what they actually license for both humans and machines."