What is cognitive load?

Cognitive load is the amount of working memory a task demands. Working memory is small, it is easily exceeded, and once it is exceeded performance and learning both collapse. The theory is John Sweller's, it dates from the 1980s. It is the reason a well-designed AI interface and a badly designed one can produce opposite outcomes from the same underlying model.

The three types, and why the distinction decides everything

Intrinsic load is the task itself. You cannot remove it without changing what is being done.

Extraneous load is waste: a confusing interface, badly organised information, having to hold something in your head that could have been on the screen. Removing it is a straight gain, always.

Germane load is the effortful processing that constructs understanding. Removing it feels exactly like removing extraneous load, and does the opposite of helping.

That last sentence is the whole practical problem. From the inside, having the difficulty taken away feels the same whether the difficulty was waste or whether it was the thing doing the work.

Why this matters for AI

Every AI tool reduces cognitive load. The question is always which kind.

Summarising a badly formatted document removes extraneous load and is unambiguously good. Generating the argument you were about to construct removes germane load, and the work still gets done, so nothing signals the loss.

The clearest demonstration is Bastani and colleagues. Students with an unrestricted GPT-4 tutor and students with a guardrailed one used the same underlying model. Grades rose in both cases while the tool was available. When it was withdrawn, the unrestricted group scored 17 per cent lower than students who never had access, while the guardrailed version largely removed the harm. Same model, opposite outcome, decided by which load the interface removed.

This is the mechanism underneath desirable difficulty. Germane load and desirable difficulty describe the same thing from two research traditions: the effort that produces the learning.

The three-way split is contested

The three-way split is contested within the field. Some researchers argue germane load is not a separate type but a description of how intrinsic load is processed, and measuring the three separately in practice is genuinely difficult. The distinction is useful and it is not settled science.

Applying it to professional knowledge work is also an extension rather than a finding. Cognitive load theory was developed for instructional design with novice learners, and whether the same categories behave the same way for an experienced practitioner using an AI tool has not been tested.

Which load is the tool removing?

Cognitive load gives the design question its sharpest form: which load is this tool removing? Ask it about a specific tool, a specific task and a specific person. It is answerable. Ask it about AI in general and it is not.

It also explains a pattern this research keeps returning to. Organisations automate to reduce load, load reduction improves output, output is what gets measured, and the germane load, the effortful processing that built the capability, disappears with everything else. Nobody chose it. It is not visible in any number anyone tracks. That is capability debt described in cognitive terms.

Related SuperSkills research

On the learning-science sibling, desirable difficulty. On the offloading decision, cognitive offloading. On learning with AI, how humans learn with AI. On the organisational version, capability debt. See should AI attend my meetings.

Key sources

• Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
• Bjork, R. A. and Bjork, E. L. Desirable difficulties in theory and practice.
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9).

About this definition

Cognitive load theory is John Sweller's and is not a SuperSkills coinage. Its internal disputes are noted above rather than smoothed over. Rahim Hirji is the author of SuperSkills (Kogan Page, 2026). Reviewed quarterly.