What is an AI hallucination?

An AI hallucination is a statement by an AI system that sounds convincing but is factually wrong or not backed by any source. The language model invents facts, figures, quotes, or sources and phrases them with the same confidence as correct information. AI hallucinations happen because a language model calculates the most likely continuation of a text and does not check whether its statement is true.

Term profile at a glance

The term is a metaphor borrowed from psychology. A person who hallucinates perceives something that does not exist. A language model perceives nothing; it produces text that looks like a fact. That is why some experts use the word confabulation, which in medicine describes the unconscious filling of memory gaps.

Why do language models invent facts?

Language models invent facts because they are trained on probability and have no built-in stop mechanism for a knowledge gap. A large language model predicts word by word what is most likely to come next. An invented year can be just as likely for the model as the correct one, as long as it fits the sentence.

AI hallucinations have 4 main causes:

1. Probability replaces knowledge: The model stores patterns from its training data. It has no verified fact database, so the most plausible answer wins.
2. Rare facts stay blurry: A birthday, a company location, or a revenue figure that appears only once in the training material cannot be reconstructed reliably from patterns. This is exactly where the model most often inserts an invented detail.
3. Evaluation rewards guessing: In September 2025, OpenAI showed in the study “Why Language Models Hallucinate” that common benchmarks count only correct answers. A guess has a chance of scoring, while “I don’t know” earns zero points. Models therefore learn to guess.
4. Errors carry forward: In long answers, every new sentence builds on the previous ones. A detail invented early pulls further invented details along with it that fit logically.

The figures from the same OpenAI study show how strongly evaluation matters. On the SimpleQA test, the older model o4-mini answered 24 percent of the questions correctly and 75 percent incorrectly, because it almost never abstained. The newer gpt-5-thinking-mini abstained on 52 percent of the questions, got 22 percent right, and was wrong on only 26 percent. The accuracy is almost the same, while the error rate drops to a third.

What types of AI hallucinations are there?

There are 4 types of AI hallucinations: invented facts, invented sources, contradictions of the source material, and confused entities. Research broadly distinguishes between errors against the world (factuality) and errors against a supplied source (faithfulness).

For businesses, the first and the fourth type are the most dangerous. An invented source is usually caught during review. A wrong price or a confused entity, on the other hand, looks credible and often goes unnoticed until a customer asks about it.

What is semantic drift and why does the plausible win?

Semantic drift is the gradual shift in meaning that occurs when a language model summarizes, rephrases, or processes content over several steps. Each individual step picks a phrasing that sounds good and stays close to the original. Taken together, the statement moves away from the source text until it claims something else.

The model prefers the plausible over the correct because that is exactly what it is optimized for. For a language model, plausible means: The phrasing appears frequently in similar texts. Correct means: The statement matches the world. The model can only measure the first from its training data. 3 patterns come up again and again:

• Negations disappear: “Does not issue court-proof expert reports” becomes “issues court-proof expert reports” when condensed, because the affirmative form is more common in everyday language.
• The unusual is smoothed into the usual: An unusual figure, such as an 11-month contract term, slips to the industry-standard 12, because that number dominates the training data.
• Contradictions get resolved: If the model finds two different values on two pages, it picks the more common one or blends both into a third.

The third point is why fact consistency carries so much weight for AI visibility. If a website states 15 years of experience on the homepage, 12 in its company profile, and 10 in a business directory, the model decides on its own which figure applies. If all sources state the same value, there is no gap for the model to fill with something plausible. That is why taismo applies one rule to every website: One answer per fact, sitewide.

How do you protect your brand from false AI statements?

You protect your brand from false AI statements by giving the model a clear, consistent, and easily retrievable answer to every important question. A decision from Canada shows how serious the issue is: In February 2024, Air Canada had to pay a customer a refund that the chatbot on its own website had promised based on an invented rule for bereavement fares. The company was held liable for what its AI said.

5 measures reduce the risk the most:

1. Standardize your facts: Prices, time frames, headcount, locations, and founding year read the same on every page and in every profile. Every discrepancy is an invitation to guess.
2. Make the entity unambiguous: Structured data connects the company, people, and profiles via sameAs into a single node in the Knowledge Graph. If there are namesakes, a disambiguatingDescription separates the right person.
3. Build citable answer pages: A grounding page answers the core questions about your business in clear sentences with figures and dates. Systems using retrieval-augmented generation pull exactly these passages at answer time.
4. Align third-party sources: Business directories, review platforms, and press articles feed into the picture a model has of you as brand mentions. Outdated descriptions there can outweigh your own website.
5. Measure AI answers regularly: Ask ChatGPT, Gemini, Perplexity, and Google AI Mode the same questions about your brand and compare the answers. A structured measurement of AI visibility follows exactly this pattern.

A GEO audit narrows down where false statements about your business come from. We check which sources the models cite, where your own information contradicts itself, and which profiles are still spreading an outdated version. One thing to keep in mind: AI systems only pick up corrected information after the next retrieval or training run. A fixed source therefore takes effect with a delay.

Frequently asked questions about AI hallucinations

Can AI hallucinations be avoided completely?
AI hallucinations can be reduced, but by current knowledge they cannot be ruled out entirely. OpenAI also states that the accuracy of a language model will never reach 100 percent, because some questions simply cannot be answered from the available information.

Does ChatGPT hallucinate less with web search?
With web search enabled, ChatGPT usually hallucinates less on factual questions, because it bases its answer on retrieved pages. The risk then shifts to the quality of those pages: If they contradict each other or are outdated, the model adopts the error.

How can you spot an AI hallucination?
The most reliable way to spot an AI hallucination is to check specific details such as figures, names, quotes, and sources against the original source. Red flags are source references without a link, very round numbers, and details that do not appear in any second source.