How much of this archive is really mine?

My archive holds two years of conversations with AI. Over three and a half million words. The question underneath is uncomfortable enough to avoid: how much of it is actually mine?

You can answer that question with a feeling. Most people do, and the answer usually turns out well for the person giving it. I wanted to count it.

What came out

Of the 3,617,653 words of conversation text, 882,344 are in my name. That is 24.4 percent. So a quarter of the archive came in through me and three quarters came from a model.

But that first number is too kind to me. Of those 882,000 words, I actually typed about 178,000 myself. That is 4.9 percent of the whole. The rest of my input was pasted: error messages, code, chunks of documents, things I picked up somewhere and dropped into the window.

One in twenty words in my archive comes from my keyboard.

How to separate typed from pasted

A computer sees no difference between typing and pasting. What it does see is length. A turn of thirty words I almost certainly wrote myself. A turn of twelve hundred words with a long error message in it, I did not.

So the measurement only counts turns of at most 200 words as typed. A longer turn counts as pasted in its entirety and drops out. That is a crude rule, and it is wrong in two directions. A long turn can contain my own writing that now drops out, and a short pasted error message still counts. What drops out is the bigger part: four out of five words in my name sit in turns above the limit.

Why 200 and not 100 or 500? There is no natural limit, so I ran the measurement at five limits. At 100 words I get 3.5 percent typed, at 500 I get 8.9 percent. So the number depends on the limit, by a factor of 2.5 between strict and loose. The picture does not. At every limit it stays a single-digit percentage, and the order of the four AI services never changes. The 4.9 percent is the middle of that range.

Within that range, the number does not hang on a few outliers either. It takes 134 of the 1,406 conversations to reach half of that typed work, and no single turn can push it up on its own. It is an estimate with a known direction, not a shot in the dark.

One lesson from building that script. Two days earlier I had counted the same numbers by hand. Six of the seven came out of the script again. The total did not, and when I checked, that number already failed to match the two percentages next to it, on the day I wrote it down. I had not noticed. A number without a script behind it is not just impossible to repeat. It can already be wrong while it sits there looking neat and everyone reads it as settled.

Where I bring in the most, I type the least

The nicest pattern is between the channels. There is one where almost half the text comes from me, far more than in the other three. In that same channel, less than six percent of those words were actually typed.

So that is where I paste. Long documents in, a model that says something short about them. It looks like the channel where I am most present, and it is the channel where I write the least.

Then I saw a second pattern. Year by year, the share I type myself seemed to drop, in every service. That fitted nicely into a story about moving from writing to steering. A day later I measured it again per quarter, and the story did not hold: within no service does the share drop. The steepest decline came from a single conversation in February, in which I had pasted a working document eight times in a row. Take that one conversation out and the decline turns into a rise.

That one conversation is not an exception. In every service, less than 2.5 percent of the turns carry half of all my words. A handful of long sessions shapes the picture. That makes an average per channel a number you should not quote on its own, not even when it is your own.

What is stable only shows once you count turns instead of words. Nine out of ten turns in my name are short enough to have been typed. Per quarter and per service that lies between 83 and 97 percent, with no direction. Five percent of the words, nine out of ten turns: almost every turn comes from my keyboard, and those turns are short. What makes the word count swing is the size of what I paste now and then.

What you can no longer recover

At the conversation level, the origin is preserved: each turn records who said what. Every note now carries its origin fields, and they are set again with every new import.

One level up, it is gone. When I distil one knowledge page from twenty conversations, who came up with which sentence evaporates along the way. For those pages I cannot reconstruct it, however much I would like to.

So I don't. New distillations get their origin recorded at the moment of writing. The 59 pages that already existed stay unlabelled. A guessed origin label is worse than none, because it reads just as firmly as a measured one. It is the same mistake as the number above, only with a word instead of a figure.

What this number is not

Two things it does not cover, and they point the same way.

The first: the count only covers exported chat conversations. The work I do with AI in my programming environment is not in it, while that is exactly where more and more of my steering happens. I have collected those sessions by now, but my chat exports lag too far behind to put the two side by side fairly. So I do not yet know my share across all my AI work combined.

The second: typed words are not the same as thinking. The five percent above counts keystrokes, not judgements. Choosing what gets built, rejecting something, sending it back twice, deciding that an outcome is wrong: that is the work that matters, and it produces hardly any words. Anyone who reads this measurement as "he only does five percent himself" is reading it wrong.

What the number does do is move a discussion. No longer whether AI takes over a lot or a little of your work, but how much exactly, in your archive, with a method someone else can check. That is a less exciting conversation. It is a conversation about something.