Cognitive debt: when organizations replace thinking with AI

Last spring I reviewed a risk assessment that a colleague had generated with ChatGPT. The format was perfect. Headings, numbered risks, mitigation actions. It looked like something a senior analyst would produce.

I asked him about risk number four. He could not explain it. He had not written it. He had not thought about it. He had prompted, skimmed, and submitted.

That moment stuck with me. Not because the tool was wrong (it was not, actually) but because the person responsible for the assessment no longer understood his own document.

What the MIT study actually found

MIT researchers recently put a name on this. They call it cognitive debt. In a study with 54 students writing essays, 83% of those who used ChatGPT could not recall a single sentence from their own text four minutes later. Their brains showed roughly half the neural connectivity of students who wrote without AI. The tool did the work and the human watched.

The researchers also found something more useful. When students who had first written on their own were given ChatGPT in a later session, their brain activity increased and their essays scored highest. The order matters: write first, augment second. They called the method “draft-then-augment.”

I think most people in my feed will read this as a study about students. I read it as a study about organizations.

Why the document is never the value

I have spent twenty years in operational environments: CNC machines at Volvo CE, 24-meter trucks on Swedish highways, a SaaS platform used by drivers at Scania, SCA and SSAB, digitalization projects at Riksdagen and Stockholms stad, business development at Volvo Group. In every one of those settings, the value was never in the document. It was in the thinking behind it.

The production planner who understood why the schedule broke on Thursdays. The driver who knew which loading dock had the broken ramp. The project manager who remembered that the last integration attempt failed because of a union agreement nobody had read. That kind of knowledge does not come from a prompt. It comes from years of paying attention.

What I see now is organizations adopting AI at the output layer without protecting the thinking layer. The risk assessment gets generated, the strategy memo gets generated, the project plan gets generated. For a while everything looks fine: documents are better formatted than before, they arrive faster, and they contain reasonable content.

Something is missing, though. The person who wrote the risk assessment cannot explain risk number four. The person who wrote the strategy memo has not thought through the second-order effects. The project plan looks solid until someone asks why we chose this sequencing and not the other.

How cognitive debt compounds in an organization

This is cognitive debt at the organizational level, and it behaves like financial debt in one important way: it accumulates quietly in good times and shows up all at once in bad times. It does not appear in a quarterly review, because nothing is technically wrong with the deliverables. The audit passes. The compliance report is filed. The board pack looks professional.

The cost shows up when something goes wrong and the person responsible reaches for the document and realizes they do not understand it. A regulator asks a follow-up question that is not in the report. A customer escalates a risk that was listed but never internalized. A senior leader changes role and the person inheriting the file has nothing but the file. The institutional memory that should have been built during the writing process was never built, because the writing was outsourced to a tool that does not stay in the organization.

The pattern compounds. New hires learn to prompt rather than to reason, because that is what is rewarded. Senior people stop catching errors because the format looks senior. Decisions get made on documents nobody owns. After a year or two, the organization has the same headcount, the same outputs, and noticeably less capability to handle anything the AI has not seen before.

How to use AI without buying the debt

I am not arguing against AI. I use Claude Code every day, I have roughly 900 logged AI conversations, and I have built thirteen projects with these tools. The question is not whether to use AI. It is how to use it without giving up the thinking.

A few things I do in my own work, which translate reasonably well to teams:

Draft first, then augment. This is the MIT finding applied directly. For anything that requires judgment (a risk assessment, a strategy memo, a project plan), I write the structure and the reasoning before I let the model touch it. The model is a second reader, not a first writer. When I skip this step, the output is faster and worse, because I lose the part where my brain has to commit to a position.

Explain before submitting. A simple test before sending anything AI-assisted: can I explain every paragraph without looking at the screen? If not, I rewrite the paragraph in my own words. This is what my colleague failed at when I asked about risk number four. If you cannot defend a sentence, it is not your sentence.

Use AI for what it is good at. Reformatting, summarizing source material I have already read, drafting boilerplate, checking for inconsistencies, generating variants for a paragraph I am stuck on. These uses do not produce cognitive debt because the thinking has already happened.

Protect the thinking layer at team level. If you lead a team, the relevant question is not “how do we adopt AI faster”. It is “how do we make sure people still understand what they ship”. That can mean reviews where the author has to defend the document verbally, or a rule that AI-generated sections are marked as such until someone has rewritten them.

The test

If you lead a team that uses AI for operational decisions, ask one question. Can the person who produced the document explain it without opening the file?

If the answer is no, the document is not the problem. The process is. The MIT study confirms what I suspected: the order matters. Think first, then augment. The organizations that skip the thinking step are building a competence gap that will not be visible until it is expensive.