Question
Can AI learn how I make decisions?
It can hold how you decide, and apply it. It cannot work it out on its own. The difference is invisible on routine work and total on a case you have never handled before.
The word doing the damage is "learn"
Learning, as the word is normally used about AI, means induction. Show the system enough examples and it infers the rule behind them. That works, and it is why these systems are good at almost everything they are good at.
Now ask what examples exist of you deciding.
Your drive is full of finished work. Documents, decks, emails, plans, offers, code. Every one of them is the residue of a decision, and none of them is the decision. The email does not contain the four versions you rejected. The plan does not contain the reason you refused the obvious approach. The deck does not record that you cut the strongest slide because it promised something the proof could not carry yet.
So induction over your work can recover how you sound, at length and in detail. It cannot recover how you rule, because rulings left no trace in the material it is reading.
The test is the case that has never happened
This is the part worth keeping, because it is a test you can run this afternoon.
On routine work, a system that inferred you from your output and a system that was told your reasons look the same. Both produce something that reads like you, because routine work is close to work you have already done and interpolation is enough.
Give both a case with no precedent in your history. A situation you have genuinely never been in, where the right call is not obvious and reasonable people would split.
The inferred system produces a fluent answer assembled from adjacent cases, delivered with the same confidence it uses for everything else. It has no way of knowing this case is different, because difference is measured against a rule and it has no rule, only a pattern.
A system holding your actual reasons can do something else. It can name which of your reasons bears on this, say what makes the case unlike the ones the reason came from, and stop where the reason runs out. That last capability is the one people underrate. Being told where the system's authority ends is worth more than another confident paragraph.
Why more data makes this worse rather than better
The instinct on hearing this is to feed it more. Everything you have written, the whole archive, ten years of files.
It does not help, and it is worth being precise about why. The problem is not volume. It is that the quantity being increased is the wrong quantity. Ten thousand pages of your output contain the same number of your decisions as one page does, which is none, because a page is not a decision. You are raising the resolution of the resemblance while the thing you wanted stays at zero.
There is a second effect that is less obvious. A larger archive makes the imitation better, which makes it harder to notice that the judgment is missing. The output stops sounding wrong. It goes on being wrong in the same places, and now nothing flags it.
What the answer looks like when it is yes
The affirmative version of this answer is narrower than the question implies, and more useful.
A system can be given the call you made, the case you made it in, and the reason behind it. Not written out in a documentation project, which fails for its own reasons, but taken from the corrections you are already making while you work. Those three parts are enough, because the reason is the part that travels.
A record that you picked option B is a fact about one Tuesday. A record that you picked option B because a promise had already been made and you will not let a promise get ahead of the proof is a standard, and standards apply to cases that have not happened yet. That is the entire mechanism. Not learning. Being told, at the moment you knew.
Frequently asked
Can AI learn how I make decisions?
It can hold how you decide and apply it. It cannot work it out by itself from your finished work, because the finished work does not contain the decision that produced it. The practical difference appears on cases you have never handled before: a system that was told your reasons can extend them to a new situation, while a system that inferred patterns from your past output can only produce something that resembles your past output.
Isn't that what personalisation already does?
Personalisation matches you to patterns that already exist in the data, which is why it is good at style and weak at standards. It can tell that you write short sentences. It cannot tell that you refuse a particular kind of claim unless the proof is in hand, because a refusal leaves nothing behind to observe.
How would I tell the difference in practice?
Give it a case you have genuinely never had before, one where the right call is not obvious. A system running on inferred patterns produces a fluent answer built from adjacent cases. A system holding your rulings can say which of your reasons applies here and what makes this one different. The second kind can also tell you it does not know, which the first kind almost never does.
Does more data fix it?
More of the same kind of data makes the resemblance better and the judgment no better. Ten thousand pages of your output still contain zero instances of you deciding, because deciding is not what a page is. It is the thing that happened before the page.
What would actually have to be captured?
The call, the case it was made in, and the reason. Reasons are what carry to a new situation. A record that says you chose option B is a fact about one day. A record that says you chose option B because the client had already been promised something adjacent and you will not let a promise get ahead of proof is a rule that will still be right next year, in a case that has not happened yet.
Where this sits
The category this belongs to is defined at what Imprinted AI is. The specific break this page turns on is set out at the Five Levels of Knowing.
Related
- The Five Levels of Knowing. Where the break between knowing what you decided and knowing how you decide actually falls.
- Can you upload your judgment to an AI? Why there is no stored copy to move in the first place.
- Imprinted AI vs fine-tuning. What statistical learning keeps, and why your judgment sits in the part it discards.
- Captured, not guessed. The one-question audit for any entry a system holds about you.
Keep going
If the test above landed, the argument underneath it is that a business now has two versions available to it, one that resembles the owner and one that carries the owner's actual calls, and almost nobody is choosing on purpose.
The full case is The A.I. Business Manifesto. About 23,000 words, free to read on the page, no gate in front of it. If you would rather have the short version, the same page will send you the three fixes and the PDF.
Read The A.I. Business Manifesto
Free either way. Reading it costs nothing and asks nothing.
Last updated: 28 July 2026