A small budget, badly spent

The MIT Media Lab put 54 people in an EEG study earlier this year and had them write essays — some with ChatGPT, some with Google, some with nothing. After four months, the ChatGPT group had the lowest brain engagement, the weakest ownership of their own text, and trouble quoting work they had supposedly written. The authors call it cognitive debt. The paper isn't peer-reviewed yet, the design has obvious holes, and I still find the result hard to ignore.

I think it's also pointing at the wrong thing.

Cognitive tools come in two kinds. Some of them teach you. The abacus, Arabic numerals, a good notebook. After enough use, you can do the work in your head, or at least better than you could before you met the tool. Others amplify you while you have them and leave you no better, sometimes worse, when you don't. GPS is the canonical example. A calculator for arithmetic you used to do by hand is another. An LLM, used to write your essay for you, is plainly a third. The atrophy worry is real for the second category. It is not the whole story for AI. The same tool can sit in either column. The question is what you ask it to do.

Working memory is small. There's a hard limit on how many distinct things a brain can hold active at once, and it is a low single-digit number. Psychology has been refining the constant for seventy years without disturbing the shape of the claim. That capacity is the budget out of which both retention and synthesis are paid. Most of what passes for thinking in a working week is retention. Names, agreements, what someone said three weeks ago, who reports to whom this month. The brain is a poor database, slow, lossy, biased, and a remarkable synthesiser. The two share a budget, and retention always wins, because retention has deadlines and synthesis doesn't.

When the retention budget runs out, and it always does, synthesis is what gets quietly skipped. The recurring concern across three different conversations, the contradiction between what someone says and what their calendar shows, the person whose meetings have slowly drifted away from the decisions they're supposed to be in, none of that gets noticed. Not because anyone is dim. Because the working memory was spent remembering who said what.

A second brain, used right, is not an assistant. It is a memory. Markdown files in Obsidian, structured fields, links between people, meetings and decisions. Claude is the part that turns yesterday's conversations into structure overnight, so that today's question has somewhere to land. I bring the question, the system holds the receipts. Note what the system doesn't do. It doesn't decide. It doesn't write the email I should be writing. It doesn't summarise away the texture of a conversation. The output is plain markdown I can read line by line. The skills that run on top are narrow, one file, one folder, one job each. If I disappeared the LLM tomorrow, the notes are still there in a format I can grep, and I am better at navigating my own relationships and decisions than I was before I built the thing.

That's the test. After a year of using it, do I think better when the tool is off? For my setup, yes. For the MIT cohort, clearly no. Different tools. Same word.

The moment that converted me had no cleverness in it. I'd been carrying a low-grade sense that one part of the business was slow to decide and another part kept asking the same questions twice. I'd assumed it was process. After a few weeks of meeting transcripts being folded into per-person notes, the shape of the problem changed under me. A handful of people were sitting on top of communication paths nobody had drawn deliberately, and the right inputs were not reaching them. Decisions queued behind them, not because they were blockers by design, but because the graph had quietly grown that way. Nobody had drawn the structure. The data drew it for me. What changed wasn't my IQ. It was my working memory budget. Retention had moved into a file. Synthesis was what was left. The insight didn't require cleverness; it required not having to remember the inputs.

The pattern in the surveys of knowledge workers using AI is consistent and worth taking seriously: the more you trust the model, the less critical thinking you do; the more you trust yourself, the more of it you do, at a higher cognitive cost. The risk vector isn't AI in general. It is the moment you stop being the one asking and become the one consuming. The discipline now is different from the one most knowledge workers have practised. It is not remembering. It is choosing what to ask. The questions I bring to a queryable second brain are themselves the work. A vague question gets a vague answer. A sharp one — where does this person sit in the communication graph, and who isn't in their feed? — gets a useful one. That is the part of thinking the next ten years will actually reward: asking the right thing of a system that can answer almost anything.

I don't worry that AI will make me think less. I worry that with retention finally outsourced, we will find out who had the good questions all along. That is a more interesting fear, and probably the right one.