What Is an AI Hallucination?

An AI hallucination is an answer from an AI model that sounds plausible and confident but is false. OpenAI researchers describe hallucinations as plausible falsehoods. For a health practice, a hallucination can mean a tool names a clinic that does not exist, gives the wrong website, or describes your services incorrectly.

How AI hallucination works

A large language model is the software behind ChatGPT and similar tools. It learns patterns from a huge body of text, then writes answers by predicting which words fit best next. It is built to produce fluent language. It is not a database that looks up a verified fact and returns it.

That design explains the problem. When the patterns point to a clear answer, the output is usually right. When the model has few or conflicting facts, it still has to write something, and fluent wrong text looks the same as fluent right text. You cannot tell from the tone.

OpenAI researchers published a paper in 2025 on why this happens. Their argument is that common training and testing methods reward guessing over admitting uncertainty. Picture a multiple-choice exam where a blank answer scores zero and a guess has some chance of scoring. A student who guesses on every question will outscore one who skips the hard ones. Models are graded in a similar way, so they learn to guess with confidence.

For a health practice, three kinds of error show up most.

• Invented entities. The tool names a clinic or provider that does not exist, or attaches a real provider name to the wrong practice.
• Invented facts. The tool gives a web address, phone number, price or service that does not match reality.
• Blended facts. The tool mixes details from two practices in the same city, such as one clinic's services with another's address.

Not every error is the model's fault alone. Some of this is a data gap. If your practice has thin, inconsistent or hard-to-read public information, the model has little to build on and more room to guess. That is why clear facts on your site and consistent listings matter. A practice with one address on its website, a second on an old directory and a third on a social page is handing the model three versions of the truth and asking it to pick.

Some falsehoods are self-inflicted. If your own site shows two different prices, or an old provider bio sits next to a new one, a tool has no way to know which is current. Broken or invented links are another type: a tool can write a web address that looks right and leads nowhere, which may account for some of the sites that did not load in the test below.

Tools that read the live web change the picture. Google's documentation for grounding says basing responses on real-world information lowers the chance that the model invents facts. A search-enabled tool can still pick an outdated page or misread a good one. In my experience the usual cause of a wrong detail is a stale third-party page, such as an old directory listing, and a tool does not reliably fact-check what it retrieves. In my judgment, one confident but wrong page can mislead a tool even when other pages are correct, so keep your own pages accurate and check regularly. Tracing each error back to the page it came from is the job of an AI source map (see the related terms below).

The related method of retrieval-augmented generation works the same way. If the page it retrieves is wrong, the answer built from it is wrong too, only now it comes with a link that makes it look more trustworthy.

Hallucination also connects to being named at all. When a model has no facts about you, the risk is not only a wrong answer. You may simply not appear, and the tool may name someone else, which is covered under AI brand mention.

Example: hallucination at a health practice

Here is the original data I collected. Health Biz Scale checked 764 established US health practices across 228 cities between May 11 and Aug 20, 2026, with one check per practice and test runs excluded. We asked an AI model 2,292 patient-style questions. The model was GPT-4o-mini, accessed through OpenAI's API, answering from memory with no web search.

The results were stark.

• The model named 3,554 practice websites in its answers.
• Of those, 1,153 (32.4%) did not load. They did not resolve to any live site.
• Only 6 of the 764 practices (0.8%) were named at all.
• About 7 in 10 questions named a city.

This is a narrow test and I want to be fair about it. The ChatGPT app searches the web live and can answer differently, so this is not a measurement of what a patient sees today in the app. It shows what a model does when it has to answer from memory. It does not say "I don't know." In many cases it fills the gap with a believable website. Some of the 1,153 sites that did not load may be real businesses that have since closed or moved, so read the figure as websites that did not load and not as a count of inventions.

Now an illustrative scenario. This example is illustrative, not a transcript. A patient in a mid-size city asks a memory-only tool for a chiropractor who treats sciatica. The tool answers with a clinic name that sounds right and a web address built from that name. The address leads nowhere. The patient tries a second link, gives up and calls the first clinic that appears in a map search. The real chiropractor three miles away, with an excellent page on sciatica care, was not part of the conversation.

The lesson is not that AI is broken. The lesson is that a model without facts about you will guess, and a guess can send a patient to a dead end.

How to check yours in 60 seconds

1. Open ChatGPT and ask, "What is the website for [your practice] in [city]?"
2. Ask what services you offer and who the providers are.
3. Check each detail against reality, including the web address, phone, address and provider names.
4. Write down every error and the tool that made it. Repeat in Perplexity and Google AI Mode.

Open a new chat for each question so earlier answers do not shape the next one. If every detail matches, ask a harder one, such as whether you accept insurance or offer a specific treatment. Test the details your own site states least clearly.

What to do about it

1. Publish a clear page stating your exact name, address, phone, providers and services in plain text.
2. Make the same facts consistent across your Google Business Profile, directories and social profiles.
3. Use your exact practice name and website address in text, not only in images or logos.
4. Add each provider's credentials and specialty on their own page, so a tool does not have to infer them.
5. Re-run the check monthly, since models and search-enabled tools change.
6. Fix the sources first. Search your practice name, find every page that shows an old address, old phone number or retired service, and correct or remove it.
7. For a serious error in a search-based tool, use the tool's feedback button and fix any source page that is wrong. Keep a dated record of what you found.
8. Do not repeat an AI's invented claim on your own site, even to refute it. Say what is true in plain words instead.