AI hallucination

An AI hallucination is a false or unsupported statement that an AI model such as ChatGPT, Gemini or Microsoft Copilot presents as fact, such as an invented price, feature, quote or source. Hallucinations come from the way models learn to guess plausible answers, and they still occur when answers are grounded in web search.

Why AI hallucinations matter

AI hallucinations are common, even in engines that search the web. In a March 2025 test by the Tow Center at Columbia Journalism Review, eight AI search tools were each asked to identify the source of 200 news excerpts, 1,600 queries in total. Collectively, they gave incorrect answers to more than 60% of the queries, and they were often confidently wrong instead of declining to answer.

For a brand, the same failure shows up as a wrong price, a feature you never built or a discontinued product presented as current. The answer is what the buyer reads. If nobody on your side reads it too, the error stays in place.

Why AI models hallucinate

Researchers at OpenAI and Georgia Tech argue that language models hallucinate "because the training and evaluation procedures reward guessing over acknowledging uncertainty." On most benchmarks, a model that guesses scores better than one that admits it does not know.

Their paper also shows which facts are most at risk. After pretraining, the hallucination rate should be at least the share of facts that appear only once in the training data: if 20% of birthday facts appear exactly once, base models should be expected to hallucinate on at least 20% of birthday facts. Facts repeated across many sources, like Einstein's birthday, are rarely wrong. A detail about your brand that appears on a single page is exactly the kind of fact this analysis puts at risk. See training data.

Grounding reduces the problem without removing it. OpenAI warns that ChatGPT's "search results and citations can be incomplete, outdated, or incorrect."

Errors about your brand: hallucination and misattribution

Two kinds of errors hurt brands. A hallucination invents a fact. Misattribution gets the fact right but credits it to the wrong source or brand: your study quoted under a competitor's name, your feature described in another product's review, or a copy of your article cited instead of the original.

The Tow Center measured misattribution directly. DeepSeek credited the wrong source for 115 of 200 excerpts, and some tools cited syndicated copies on Yahoo News or AOL rather than the original publisher.

Example (illustrative): a buyer asks "Does Acme Payroll pay contractors in Germany?" The engine answers yes and cites a comparison page, but the sentence it relied on describes a competitor.

How to catch and correct hallucinations about your brand

Read the answers, not just the scores. Hawkscan shows the full answers, word for word, from ChatGPT, Microsoft Copilot, Gemini and Google AI Mode for each tracked prompt, so a wrong price or a misattributed feature does not go unseen.
Open the cited source. If a page caused the error, correct your own page or ask the third-party site to fix theirs.
Repeat key facts consistently. State prices, features and policies the same way on your site and on the profiles engines read. Consistent facts help engines recognize your brand as one entity.