Let’s talk about the person you were six years ago.
Not the flattering version. The real one. The one with opinions you’d now have to explain, the certainty you’d quietly like to walk back, the version of your taste and your work that a fair judge would call a draft.
Now put a hiring panel in a boardroom, and on the screen behind them a two-page summary of that person, written by a machine, assembled from everything you ever said in public. Somebody reads it out. It sits in the room before you do.
And here’s what nobody says in that moment. The summary isn’t wrong. Every line is sourced. That’s the problem.
There used to be a version of this where you were the primary narrator of your own past. Before the record was total, a man who changed careers at forty-five told his own story at the interview, in his own words, with his own editing. He got to say what it meant. He got to leave things out. Not because he was lying — because that’s what a person is. A moving target with a memory that curates itself.
That wasn’t a loophole. That was how a human made it from one decade into the next.
And we’ve swapped it for something that never forgets and never reinterprets.
The mechanism is boring, which is why it’s winning.
Consistency is cheap to verify. Contradiction is expensive to explain. When you’re building a system that has to score a human being — a job, a loan, insurance, access — the most convenient signal in the world is: does this person match their own file. It needs no judgment, no context, no interview, no human who has to stand behind the call. The machine can compute it while you’re still parking the car.
So consistency becomes a virtue by default. Not because anyone decided it should be. Because it’s the only thing the system can measure cleanly.
And once consistency has a price, changing your mind acquires a cost.
That should bother you more than the surveillance does. We’ve spent a decade arguing about what the machine is allowed to know. We’ve spent almost no time on the harder question — what it’s allowed to hold you to.
Because the archive isn’t a memory. A memory is something you carry, and you get to reinterpret it. A file is something that happens to you. And when the file gets pulled into a summary by a model with no interest in why you were wrong, you aren’t being remembered. You’re being enforced.
Look at where this actually bites.
A founder. Three years ago, before the pivot, she published pricing and a promise about what the product would never do. The market moved. The business moved. The promise is wrong now, and she’s on a stage saying so out loud. Somebody in the audience has the old page open. The machine has the old page open forever, and it has no reason to surface the correction next to the original. Half a million words of context, and the sentence that gets quoted is the one she’d take back.
A kid at sixteen, performing an identity for an audience of eleven people. Legally a child. Developmentally a larva. At twenty-six he’s applying to teach. Everything is fine except the record, which has no concept of age.
A doctor who said something cautious and slightly wrong on a podcast in 2019, in an area where the guidelines have since reversed. His actual position has moved twice. The search result hasn’t moved at all.
None of those people are dishonest. Every one of them did the one thing that separates a living mind from a database. They updated.
And here’s the uncomfortable part, where the machine’s logic is completely right and completely wrong in the same breath. It’s correct that a person who says two different things on two different days is a less stable dataset. It’s just that being a stable dataset was never the point of being a person.
The most valuable people I know had a position, held it for years, then killed it in public when the evidence turned. That’s not inconsistency, it’s the whole function of judgment. A conviction isn’t a promise. It’s a current best guess you’re obliged to be willing to abandon.
So what does the machine do with someone who changes? It flags risk. Not out of malice — structurally. A model trained on outcomes learns that volatility is a signal, and a signal is a signal. Nobody has to want you punished. The system only has to prefer the flat line.
Which is how you end up with a generation doing the arithmetic and deciding to stay exactly the same. You can already watch it in how people write in public. Everyone is careful. Everything is hedged years ahead of time. Nobody says the half-formed thing anymore, because the half-formed thing is the only thing that can be quoted back at you, and a finished thought never gets recorded. We built an incentive structure that punishes thinking out loud and rewards saying nothing, and then we called it safety.
Operator lesson, if you build anything that touches a human record.
Design for revision on purpose. If your product holds a reputation, a history, a profile, or a score, ask the question most teams never ask: can this system represent a person changing? Can a claim be superseded without the old one staying co-authoritative? Can a customer correct the summary you generated about them, and does that correction count going forward — or is the summary the truth and the human the disruption?
An appeal that just runs the same model a second time is a re-roll with a form attached.
And timestamp everything. A record of a person should carry its dates, its sources, and the plain fact that it’s a snapshot of a Tuesday. Not a verdict with no date on it.
The bigger thing sits underneath all of this, and it isn’t about AI.
Sovereignty has a time dimension, and we keep ignoring it. We argue about who holds your data. We should be arguing about who holds your past, and whether you have the authority to overrule it.
The right to change your mind isn’t a nice-to-have. It’s upstream of almost everything else we care about — the right to disappear, the right to be unclassifiable, the right to be forgiven. All of it is the same request in different clothes: don’t let a copy of me that I no longer agree with outvote me permanently.
You’re going to be wrong about something this year. You’ll hold it confidently, and then you’ll watch it come apart, and you’ll have to walk it back in front of people who remember the old version.
Do it anyway.
The alternative is a life spent maintaining an earlier edition of yourself for an audience that isn’t watching. There’s nothing sovereign in that. There’s just a person slowly becoming their own file.
You’ve got this.
⚠️ File-mutation verifier: 1 file(s) were NOT modified this turn despite any wording above that may suggest otherwise. Run git status or read_file to confirm.
• /tmp/noop_probe_ignore.txt — [patch] Failed to read file: /private/tmp/noop_probe_ignore.txt