What AI assistants get wrong about people, and why it is hard to correct

Ask an assistant about a company and a wrong answer costs a sale. Ask it about a person and a wrong answer is a different kind of problem, because there is no support desk, no correction form, and no way to know how many people have already been told.

I do this professionally for a small number of clients, and the pattern of what goes wrong is consistent enough to be worth writing down.

What kinds of error actually occur?

Four, roughly in order of how often I see them.

Staleness. A person's role, company or location is described as it was several years ago, because that version of them was written about more than the current one. This is the most common by a wide margin and the least dramatic.

Conflation. Two people with the same or similar name are merged into one, and the resulting description is a person who does not exist. Common surnames make this near-inevitable.

Omission. A person who is genuinely notable in their field returns almost nothing, because the record of them lives in places that are not easily read: printed press, closed platforms, other people's coverage rather than their own.

Invention. Plausible detail that has no source. Rarer than the discourse suggests, but the hardest to argue with, because there is nothing to correct.

• Stale: an old role or an old company · fixable at the source

  The most common one, and the most tractable. Publish the current version somewhere authoritative.

• Omitted: notable, and largely absent · fixable at the source

  Also tractable. There is nothing to correct, only something to create.

• Conflated: merged with a similar name · harder, and slower

  Needs disambiguating signals across several sources before it settles, and it settles slowly.

• Invented: a plausible detail with no source · harder, and slower

  Nothing to point at, so the only lever is making the accurate record more prominent than the gap.

Why can you not simply ask for a correction?

Because there is nothing that holds the sentence.

A search result points at a page, and a page has an owner. An assistant's answer is assembled at the moment you ask, from a model trained on a corpus plus whatever it retrieves. There is no record with your name on it sitting in a database waiting to be edited, so there is no edit to request.

What you can change is the material the answer is assembled from, which is slower, indirect, and the only lever that exists.

What actually moves it?

Making the accurate version of the record more available, more consistent and more authoritative than the inaccurate one.

In practice that means a small number of unglamorous things. A page the person controls that states the current facts plainly. Consistency across the places that already rank, so the sources agree rather than compete. Entries in the structured, machine-readable references these systems lean on heavily, where those are appropriate and the person genuinely qualifies. And correcting the third-party records that are carrying the stale version, which is usually where the actual error lives.

None of that is a trick. It is record-keeping, done deliberately, in public.

How long does it take?

Longer than people want, and the honest answer is that it varies enough that I would distrust anyone who gives you a date.

Changes to your own pages can be reflected quickly. Changes that depend on a model's training data may not be reflected for a very long time, or at all in the version currently deployed. What improves faster is anything the system retrieves live rather than recalls, which is a growing share of how these answers get built and the reason the work is worth doing at all.

Who should care about this?

Fewer people than the anxiety around it suggests.

If your name returns nothing much and nobody is looking you up, this is not your problem and I would not sell it to you. It becomes real when someone is making a decision about you in a room you are not in: a funder, a counterparty, a journalist, a board, a family office deciding whether to take a meeting.

That is a narrow set of situations. In those situations it matters considerably, because the paragraph is read as neutral fact rather than as something anybody wrote.

What would I do first?

Look yourself up, in a private window and in two assistants, and write down what is wrong in the four categories above. Most people discover something stale rather than something scandalous, and stale is the version you can fix in an afternoon.

The commercial version of this problem, for businesses rather than people, is in what AI assistants actually read on your website. Where it concerns a family name rather than a company, that is work I do quietly.