When an AI gets you wrong, the problem is upstream

The assistant is not lying about you. It is accurately repeating something inaccurate that it found, and it has no reason to prefer your own account.

By Aaron Stalberg · 20 September 2026

I have a professional interest in this, so read the rest with that in mind. I have also watched it happen to myself, which is the only reason I think about it as much as I do.

Ask an assistant what it knows about a specific person and you will usually get something broadly accurate assembled from three or four sources it has decided to trust. If one of those sources is wrong, the answer is wrong, and it will state the error with the same confidence as everything else. Nothing in the output tells you which part came from where.

Assistants do not rank, they resolve

A search engine returns a list and leaves the judging to you. An assistant has to commit to an answer, and committing means deciding which sources to believe. That decision tends to favour things that are specific, dated, plainly written, and repeated across more than one place.

Notice what is not on that list. Being the most authoritative source is not the same as being the one an assistant prefers. A short trade-press article from seven years ago that states something clearly and never gets updated can outrank a formal record that is more accurate but harder to parse.

Why the primary source often loses

This is the part that surprises people. The authoritative record is frequently not the thing being quoted:

• It may be gone. Records get deleted, withdrawn, or quietly expired.
• It may be hard to read. A register entry or a decision document is written for legal precision, not to answer a question.
• It may be undated in a way that makes it useless for recency checks.
• It may be behind a wall, or unindexed, or simply never crawled.
• It may not say the thing plainly. A record can establish a fact without ever stating it in a sentence a model can lift.

Meanwhile the secondary source is clean, dated, quotable, and indexed. From the model's point of view it is the better citation. It is doing what it was built to do.

So correcting the AI does almost nothing

You can report an error to an assistant, and you should if the output is genuinely false. But it is usually a weak move, because the model's next answer will be assembled the same way from the same sources. Correct the output and you have fixed one sentence. Correct the source and you have fixed every future answer that quotes it.

This is why, when something inaccurate is circulating about you, the work has a specific order:

1. Find out what the assistant is actually citing. Most of them link their sources. That list is the real to-do list, and it is usually shorter than you expect.
2. Correct the sources you can reach. Publishers respond to specific, evidenced correction requests. They do not respond to complaints about tone.
3. Put a primary, dated, plain-language account on a domain you control. This is the step most people skip, and it is the one that changes what gets quoted. A model needs something citable; if your own site does not state your position in a liftable sentence, it has nothing to use.
4. Then be consistent everywhere. Same facts, same wording, across your site, your profiles, and any directory that lists you. Contradiction makes a model hedge, or drop you.

The uncomfortable part

There is a version of this that turns into reputation management, and I want to be careful about it. The aim is not to bury true things. If a source is accurate, it should stay, and no amount of publishing fixes a fact you dislike.

What is worth doing is narrower than people assume. It is making sure the accurate version exists, is dated, is plainly written, and is on a page a model can actually read. In my experience most people who are misrepresented online have never done that one thing. They have complained about the output instead of publishing the correction.

An assistant cannot quote an accurate source that does not exist. That gap is the whole problem, and it is the only part of it you fully control.