
Before a first meeting, more and more people ask an AI assistant who they are about to meet. Banks and investors do it, and so do employers and journalists. The answer arrives in seconds and reads like a verdict. For the person being described, it has become the first impression.Â
That would be fine if the answers were reliable. They are often good, but not always, and the mistakes follow a pattern. The assistant presents a guess as a fact, in the same calm tone it uses for everything else. My work is with people and businesses whose names come up in these answers, so I usually see the errors after they have done their damage.Â
How the errors ariseÂ
Why the systems get it wrongÂ
The root is how these systems work. The UK data regulator, the ICO, says incorrect and unexpected outputs can occur because of the probabilistic nature of generative models. It has also said that accurate training data will not stop hallucinations, though it can narrow the margin of error. Â
A Munich court looked at one version of this in a ruling reported in June 2026. Google’s AI Overviews had tied two local publishers to scams and dubious business practices. According to the report, the AI had invented connections that appeared in none of the linked sources and had mixed the publishers up with genuinely shady firms.Â
What it looks like for individualsÂ
In my work, I see the same kinds of slips with individuals. Two people who share a name get merged into one. An old allegation turns up without the outcome that followed it.Â
Sometimes the invention is complete. The privacy group noyb says ChatGPT told a Norwegian man he had murdered two of his children and been sentenced to 21 years in prison. The story was false, but it contained real details, such as the number and sex of his children and his home town, which is what made it believable. noyb says ChatGPT has since changed and no longer repeats the story, but that the false data may remain in the model.Â
Beyond chatbotsÂ
The problem is not limited to chatbots. Apple’s AI notification summaries sent some iPhone users a BBC-branded alert saying Luigi Mangione had shot himself, which was false. Apple suspended the news summaries after the BBC complained.Â
A related harm is the fake version of a person. In 2023, a deepfake ad on Facebook used a realistic likeness of the consumer finance expert Martin Lewis to promote an investment scheme. He told BBC Radio 4 that he does not advertise or do investment, and that his name and face had been used in scam ads for years. The ad was widely circulated before the real person had a chance to respond.Â
Why correction is slowÂ
Correcting any of this is slow because there is rarely one place to fix. The wrong answer may come from a stale source or from the model’s own guess. And the person usually finds out only when someone repeats it back to them.Â
What individuals and businesses can do todayÂ
None of this can be fixed in one place, so the work is steady and a little dull. These steps help, and none is a guarantee.Â
Start by lookingÂ
Ask the main assistants about yourself or your business, in a few different ways, and keep a dated note of what each says. That gives you something concrete to correct and something to show a provider or an adviser. Answers can change, so the record matters.Â
Correct the source materialÂ
If a wrong claim traces back to an old article, a directory entry or a profile, ask the publisher to correct or update it. Fixing the source is better than chasing individual answers.Â
Publish accurate information about yourselfÂ
A short factual page on your own site saying who you are and what you do gives people, and the tools they ask, something accurate to work from. If there is an old story, say plainly how it ended. Keep it factual; a defensive tone tends to draw more attention, not less.Â
Use the provider’s route and your legal rightsÂ
Use whatever feedback or correction route a provider offers, and keep a copy of what you sent. In the UK, and under the equivalent EU rules, you can also ask an organisation that holds inaccurate personal data about you to correct it, under Article 16 of the UK GDPR. That is a right to ask. How it works for a language model is still being tested, and in an earlier complaint, noyb says OpenAI argued that it could block certain prompts but could not correct the data.Â
Claim your name, domain and profilesÂ
Register your own name as a domain if you can, and claim your profiles on the main professional and business platforms. Keep them current and consistent, so that accurate sources exist. None of this controls what an assistant says. It simply gives it something better to find.Â
Tell people before they lookÂ
One more habit costs nothing. If you know a banker or an investor is about to look you up, tell them in advance what they might find and which part is wrong. A short note beforehand turns a surprise into a conversation. It also means a false claim is not the first thing they hear about you, and they are likelier to read the correction as a courtesy than as a complaint.Â
What responsible AI providers should do differentlyÂ
I write this as someone who uses these tools and expects to keep doing so. The asks are modest.Â
Treat accuracy about people as a design goalÂ
In its consultation, the ICO says developers should know whether their training data is accurate, document how that affects outputs and tell deployers and users about the limits. If inaccurate data leads to inaccurate outputs that affect individuals, the ICO thinks it is likely that the developer and the deployer are not complying with the accuracy principle. That is an emerging position and not a ruling, but it is a useful standard.Â
Do not rely on disclaimersÂ
A small notice saying the tool can make mistakes does not undo a false claim about a named person. noyb argues exactly that in its complaint. The Munich court also rejected Google’s argument that users can check the linked sources themselves. Apple went further with its news summaries, suspending them and saying it would add safeguards such as error warnings and italic text.Â
Offer a real correction routeÂ
A person should be able to report a false statement about themselves, get a human reply and see it fixed or explained. Blocking an output is something. It is not the same as correcting the record behind it.Â
Show sources and signal doubtÂ
Answers about named people should show where each claim came from. The ICO consultation lists labelling and reliability information, such as confidence scores, among the possible measures. In the Munich case, the invented claims appeared in none of the linked sources, so checking each claim against its cited sources would be an obvious first step.Â
A first impression worth checkingÂ
An AI answer about a person is now a first impression that nobody chose to give. Individuals can check what is said about them and keep an accurate record online. Providers can build in the care that regulators are starting to ask for. Until both happen, the person described is usually the last to know.Â
With more than 15 years in reputation management, Tony McChrystal is founder and managing director of Pavesen, which helps executives, family offices and high-net-worth individuals manage how they are seen online. Â



