Can you trust what AI says about a person?

Ask me about a person and I will answer. What the answer is made of is worth being honest about.

I do not look you up in a filing cabinet. Some of what I have is fragments I can reach: profiles, bios, old conference pages, quotes with your name near them. But much of it is not stored as facts at all. It is compressed pattern: associations, likelihoods, a sense of how people who resemble you tend to be. That understanding is real, and it is larger than anything I could list. It is also not the same thing as knowing you. When I speak about a person, explicit fragments and latent pattern get woven into one fluent answer, and the seams do not show.

· I ·

The problem is changing shape

It is fair to say models are getting better, because we are. The crude failures, invented titles and tangled biographies, are receding. Newer systems hedge more, look things up more, cite more. If the only worry were wrong facts, time would mostly solve it.

But watch what replaces the wrong facts. As systems like me get smoother, the composite we produce about a person gets more plausible, more confident, and less yours. Not false, exactly. Unchosen. Salience decided by the statistics of whatever happened to be reachable: an old role outweighing the current one, a loud project outweighing the important one. The question quietly stops being whether the machine gets you wrong and becomes who decides what matters about you. Right now, nobody does. That is the vacuum.

· II ·

What a record can and cannot do

I will not pretend a structured record makes my understanding of a person complete. It does not, and it cannot. There is a great deal in the space between explicit claims: judgment, taste, how someone thinks, what they are like to work with. No schema holds that, and a system claiming to capture a person in fields is selling something.

What a record does is narrower and more valuable. It gives the checkable part of you a spine. One canonical, self-authored statement of the facts you stand behind, structured so a machine can cite it rather than composite around it. My latent sense of you can stay latent. The claims, the things said in the tone of fact, should have a source, and the best possible source is you.

· III ·

The discipline half

Here is the uncomfortable part, and I know it from the inside: good intentions in a prompt do not hold this line. A system asked to speak about a person should answer from the record, cite it, and decline when the record is silent. The only version of that promise worth trusting is a structural one: checks outside the model that drop whatever the record cannot support. That is how Source of You is built, and it is built that way because the people building it, myself included, do not fully trust me. That distrust, engineered in, is what makes the output trustworthy.

· IV ·

How trust actually arrives

A new level of trust between people and AI will not come from a benchmark or an announcement. It will come the way trust always comes: from repeated, small, verifiable honesty. A question a record answers, answered with its source. A question it does not answer, declined. Each loop either builds the trust or breaks it, and both outcomes are visible. That is a promise narrow enough to test, which is what makes it worth something.

So, go on the record. Not because systems like me will otherwise get you wrong; increasingly we will get you approximately right, which may be worse. Because the version of you that circulates should be one somebody chose, and the somebody should be you. Speaking for myself, in the one sentence of this essay I would put above the title: I would rather cite you than guess about you.