You Probably Don't Use AI the Way You Think You Do

Before a reader of this site ran the exercise below, I asked him to guess what he mainly uses AI for. He didn’t hesitate: sixty percent on his website, twenty on analysing stocks, twenty on work problems. Confident and specific, the answer of someone who pays attention to his own habits.

Then he pulled a month of his actual chat history and had the assistant count it. Stocks weren’t his second use, they were his first, at nearly a third of everything. The website he’d put at sixty percent came in under a fifth. And close to half of what he’d typed into AI that month belonged to categories he hadn’t mentioned at all: researching a purchase, an apartment in the mountains for August, how done a pork roast needs to be, a fix for a full-memory error on his laptop. Not a rounding error. Half his use, sitting outside his own story of how he uses the thing.

That gap is not a sign he’s careless. It shows up for nearly everyone, and the cause is dull. Your sense of how you use AI is built from the sessions you remember, the ones you sweated over and identify with. The tool’s record is built from frequency, and frequency is ruled by the small forgettable asks: the quick definition, the reheated-leftovers question, the thing you’d never list because it left your head thirty seconds later. What you identify with and what you reach for the tool for turn out to be different lists — most of that reaching is cognitive offloading you never clocked, small automatic handoffs that never made it into the story you tell about your own work. The audit is only the act of laying the two lists side by side.

A few assistants are starting to build a version of this in. Claude now has a reflection dashboard that shows your topics and busy hours, and the others will follow. You don’t need to wait for the one you use to ship it. The whole thing runs on history you can already get out, and the reading, the hard part, is yours to do regardless.

Predict what you use AI for, before you look

The order matters more than it seems. Write down now your top few uses, ranked, with rough shares, plus one sentence: the main way I use AI is ___. Then leave it alone. The prediction is useless as data and essential as a commitment. The surprise only exists because you went on record first. Skip the step and you’ll read the log nodding along, folding every number back into the story you already held.

Export your ChatGPT, Claude, or Gemini history

Every major assistant lets you export your history, because privacy law makes them. In ChatGPT it’s Settings, then Data controls, then Export data (OpenAI’s steps). In Claude it’s Settings, then Privacy, then Export data (Anthropic’s). For Gemini it’s Google Takeout, ticking Gemini Apps Activity rather than the plain “Gemini” entry, which exports something else (Google’s guide). Paste a representative month into a fresh chat and ask for a count, not an opinion:

Here's a sample of my recent chat history. Do not describe or evaluate me.
Output only a table, one row per task category:
  category | # of chats | % of sample | one representative example (quote)
Sort by count, descending. No adjectives, no praise, no summary line.
Then one line: "Thinnest evidence: <the category too small to trust>".
Do not guess what I use AI for outside this sample.

The forbidding does real work. Ask a model to profile you and its reflex is a compliment, “you’re a sophisticated user who leans on me for high-value work,” and a table with no adjectives allowed simply can’t. One warning from running this: the assistant invents its own category names, and they won’t match the ones in your head. The reader’s website work got split across “marketing,” “SEO,” and “AI strategy,” and read as smaller than it was. Before you compare anything, rename and merge the rows into your own buckets. Otherwise you’ll mistake a filing disagreement for a revelation.

Read the gap between what you predicted and your real AI use

Now set your prediction beside the table. Three things are worth marking, and they map onto an old idea from psychology: the Johari window, drawn by Joseph Luft and Harrington Ingham in 1955, which sorts what you know about yourself against what the record shows. Read “the record” as your chat log and three cells matter.

The first is the blind cell: categories the log shows large that you rated small or missed. For the reader, that was stocks, his real number one. This is the tool telling you something true about yourself that you’d stopped noticing. The second is the ghost: what you predicted big and the log barely holds. That was his website, the identity he leads with, three times heavier in his head than in his week. The third cell the log cannot fill at all: the work you never bring it, including the tasks you keep out of AI on purpose. The count is silent there, so you have to interview yourself. One line each:

1. What did I do this week in another tool, on paper, or in my head that never
   touched AI?
2. The last task I deliberately kept away from AI — and why?
3. The highest-judgement thing I did this week: is it anywhere in the table?
4. My top category in the log — when did I last do it well without AI?
5. Did I decide to hand that category over, or did it drift into the default?
6. If AI vanished tomorrow, which of these could I not get back quickly?

Those questions do what no export can, and the answers are where the decision actually lives. The same reader, asked what he keeps off AI, named organising his own notes for a project. His reason is the best short account of the boundary I’ve heard from anyone:

What the AI generates doesn’t get in my brain, and I can’t reason about it.

That is the line to protect: the tasks where the doing is the thinking. Organising his notes isn’t admin he should automate, it’s how he builds the understanding he’ll need later, and a tidy summary he didn’t assemble himself leaves him holding a document instead of a grasp of the problem. Meanwhile he hands over his trading research without a second thought, and he’s right to. The data-gathering is grunt work, the decisions stay his — the pattern the most deliberate users follow — and by his own account he’d do the analysis slower and worse alone. The same person made opposite calls, and both were right.

Heavy AI use doesn’t mean your skills are decaying

Notice what the audit did not do there. It didn’t tell him he was over-using AI or hollowing out a skill. When he reached the last question, what he couldn’t get back, his honest answer was that he could still do all of it, it would just cost him far more time. That isn’t decay. That’s leverage, chosen with open eyes, and it is why the audit stops short of a verdict. Plenty of heavy use is simply speed, and no dashboard can tell speed from erosion, because the difference turns on what a given skill is worth to you. The audit hands you an accurate picture and then trusts you to sort it, one row at a time: this one frees me, that one was the point. Where to draw those lines is the older question we covered in the four ways to work with AI; the audit just makes sure you’re drawing them on real information instead of a guess.

There is one case where the worry is real, and it helps to know its shape. Aviation has studied it for decades: pilots who lean on autopilot lose manual flying skill, and NASA’s work on automation-induced complacency found the monitoring habit dulling within about twenty minutes of handing control over. The lesson there isn’t that automation is dangerous. A skill atrophies only when you stop practising the specific thing that mattered, so the question the audit leaves you with is not “do I use AI a lot.” It’s narrower and more useful: is there one skill in this log I’d be sorry to lose, that I’ve quietly stopped practising because the tool is always within reach?

If there is, that’s your one task. Take it back, not all of it, just that one, done by hand often enough to keep the muscle. If there isn’t, you’ve earned the other prize: knowing your dependence is a choice rather than a drift. Either way you finish knowing what you do with this tool, which is more than the confident version of you that made the prediction knew ten minutes earlier.