Ideas

Clark and Chalmers proposed in 1998 that cognitive processes don't stop at the skull. When you use a tool reliably and consistently, it becomes part of your cognitive system. Not metaphorically — functionally. Your notebook isn't storing memories for you. It is part of your memory system.

Menary pushed this further with cognitive integration: the question isn't whether the tool functions like a brain process. It's whether the person and the tool have developed genuine practices of thinking together. This is the framework behind everything I build with AI — not giving people a tool to use, but building something that becomes part of how they think.

Piaget's insight was that learning isn't absorption — it's construction. New information either fits into existing mental models (assimilation) or forces those models to restructure (accommodation). Real learning lives in accommodation. It's uncomfortable, and most training avoids it.

Sweller's cognitive load theory adds the constraint: working memory is limited. Overload it and learning stops. The art is keeping learners in the productive zone — enough demand to trigger restructuring, not so much that working memory collapses. This is why a 200-slide onboarding deck doesn't work. It's not bad content. It's bad cognitive design.

Ericsson's research demolished the idea that expertise comes from experience. It comes from a specific kind of practice: targeted effort at the edge of current ability, with immediate specific feedback, repeated over time. Most practice in most organizations doesn't meet any of these criteria.

AI changes this equation. AI can identify the edge of someone's ability in real time. It can provide immediate, specific feedback. It can space repetition across optimal intervals. The technology exists to make deliberate practice the default. The question is whether anyone builds it that way.

Knowing something and doing it under pressure are two different neural systems. Declarative memory (facts, frameworks, steps) and procedural memory (execution under real conditions) don't transfer automatically. Most training lives entirely in declarative space: slides, playbooks, quizzes. Actual behavior change requires encoding in procedural memory through spaced repetition, emotional salience, and practice in context.

This is why a rep can ace a methodology quiz and still revert to feature-pitching on a live call. The knowledge is there. The behavior isn't. Closing that gap is the hardest problem in organizational learning, and it's where most of my work lives.

My dissertation examined what makes people engage deeply in learning rather than going through the motions. The short answer: autonomy, relevance, and the belief that effort leads to improvement. The long answer involves self-determination theory, self-efficacy research, and how digital environments can either support or undermine intrinsic motivation.

This matters for AI because the same tool can create dependency or agency. The AI does the work and the person clicks accept — that's dependency. The AI scaffolds the thinking and the person constructs the insight — that's agency. The difference isn't the technology. It's how you design the interaction.