AI hallucinations: definition, causes, and how to reduce them
Article published on November 7, 2026, based on OpenAI's research publication on the causes of hallucinations (September 2025), court decisions and press coverage of the cited cases, and our training practice on Copilot, ChatGPT, Claude, Gemini and Mistral.
An AI hallucination is a false, fabricated, or unrelated response that a language model presents with the same certainty as a correct one: a non-existent legal precedent, a made-up figure, a quote attributed to the wrong person. This isn't a bug that the next version will completely fix, but a consequence of how these models work: they produce the most plausible text, not the truest. Models in 2026 hallucinate far less than those in 2023, but they still do, and the remaining errors are harder to spot because they're buried among otherwise excellent responses. This article explains what a hallucination is, where it comes from, what it has already cost, and ten techniques that truly reduce the risk.
The essentials in five lines
- A hallucination is a false or unverifiable claim produced by generative AI and presented as a fact.
- She comes of the functioning of LLMs, which predict the most likely continuation of a text, and of their training, which rewards a confident answer rather than an "I don't know".
- The most risky cases These include legal references, sources, precise figures, quotes, and recent or rare facts.
- To reduce itThe model is given the reference documents, asked to quote the exact passage, and allowed to say that he does not know.
- Human verification remains mandatory for everything that is distributed: the tool provides the information, the person who sends it is responsible.
AI hallucination: definition
We speak of hallucination when a generative AI system produces content presented as accurate when it is false, fabricated, or unfounded in the information it possesses. The term, borrowed from psychology, is debated: the model doesn't "see" anything; it calculates. Some researchers prefer "confabulation," which better describes the phenomenon: filling a gap with a plausible invention, without any intention to deceive. However, "hallucination" has become the accepted term, in both French and English.
Two features distinguish a hallucination from a simple mistake. First, it is plausible The name of the court decision resembles a real name, the figure is of the correct order of magnitude, and the cited author sometimes actually exists. Furthermore, it is Nothing in her tone signals doubt. It is this combination that makes her dangerous in a business setting.
The five types of hallucinations
| Type | What is happening | Example |
|---|---|---|
| Invented fact | The model asserts a fact that does not exist. | A company creation date, a manager, a false revenue figure |
| Source manufactured | He cites a reference that doesn't exist, or that doesn't say what he attributes to it. | A legal precedent, a scientific article, or an invented standard |
| Infidelity to the document | He summarizes or analyzes a provided document, adding what is not already there. | A meeting report that mentions a decision that was never made |
| Calculation or logic error | He "calculates" in his head and makes a mistake, or follows a flawed line of reasoning. | An incorrect column sum, a miscalculated percentage |
| False capacity | He claims to have done something he did not do | "I checked on the website," even though no search had been initiated. |
The third type is the most underestimated. It's often believed that giving a document to the model eliminates the risk. While this does significantly reduce it, a summary can still slip in a missing nuance, merge two points, or attribute a sentence to the wrong speaker. The longer the document and the more vague the request, the greater the risk.
Why AI is hallucinating
The first reason lies in the very principle of an LLM. A large language model is trained to predict the most probable next word. It has no fact database to consult; it has learned regularities from billions of texts. When a question concerns a fact well-represented in its data, the most probable continuation is often correct. When it concerns a rare, recent, or absent fact, it still produces the most probable continuation, which resembles a correct answer. We detail this mechanism in LLM: definition and operation.
The second reason is more surprising. In September 2025, researchers from OpenAI and Georgia Tech published "Why Language Models Hallucinate." Their thesis: the methods used to train and evaluate models reward guesswork rather than admitting uncertainty. Most tests score a correct or incorrect answer, without awarding points for "I don't know." Like a student facing a multiple-choice question with no penalty for incorrect answers, the model has an incentive to take a chance. The authors also link the hallucinations to events that appear only once in the training data: the model doesn't have enough examples to learn them, but enough to invent something similar. They propose that assessments stop penalizing inaction.
Next come reasons related to usage:
- The date of acquaintance : without web search, the model ignores what happened after its data was stopped, and may fill in with outdated information.
- A question that suggests the answer : “What case law confirms that…” prompts the model to find one, even if there is none.
- The lack of context : without your internal documents, the model invents your processes based on what companies "generally" do.
- Conversations that are too long : beyond a certain volume, the model mixes elements from different moments of the exchange.
- Complacency : models tend to go in the direction of the user, which can lead them to confirm a false premise.
Three expensive cases
| Business | What happened | Result |
|---|---|---|
| Mata vs. Avianca, New York, 2023 | Lawyers file a brief citing six decisions invented by ChatGPT, which had assured them that they existed. | Judge Castel fined them $5,000 on June 22, 2023. The case became a global benchmark on the subject. |
| Moffatt v. Air Canada, Canada, 2024 | The Air Canada website assistant directs a customer to a refund policy that does not exist. | The court finds the company responsible for the information provided by its assistant and orders it to compensate the client. |
| Deloitte, Australia, 2025 | A 237-page report submitted to the Ministry of Employment, paid for at 440,000 Australian dollars, contains references to non-existent work and a fabricated quote from a judge. | The firm refunds part of the contract and publishes a corrected version mentioning the use of Azure OpenAI. |
The phenomenon is far from anecdotal. As of early June 2026, the database maintained by researcher Damien Charlotin listed over a thousand American cases in which court documents contained erroneous content generated by AI, approximately 40% of which originated from lawyers. The common thread in these cases is not the tool used, but the lack of verification prior to publication.
Where the risk is highest
| High risk | Moderate risk | Low risk |
|---|---|---|
| Legal references, standards, and articles of law cited from memory | Summary of a long document provided | Rewrite, correct, change the tone of a text you have written |
| Sources, bibliographies, citations | Synthesis of several contradictory documents | Translate a provided text |
| Precise figures, statistics, prices | Data analysis without calculation tools | Propose a plan, ideas, titles |
| Recent or little-known facts: an SME, a person, a product | Minutes of a meeting based on a transcript | Structuring notes into a report, with proofreading |
The rule that emerges is simple: the less background information you provide, the higher the risk. When AI reworks content you give it, it rarely makes a mistake. When it has to produce a fact, a source, or a figure itself, it can make a mistake, and that's where verification must be systematic.
Ten techniques to reduce hallucinations
No single technique eliminates the risk, but their combination greatly reduces it. They work in all assistants: Copilot, ChatGPT, Claude, Gemini, and Mistral's Vibe.
- Provide the reference documents. Attach the contract, procedure, report, or use a project, connector, or agent linked to your sources. The model then responds based on your text rather than its memory: this is the principle of RAG, detailed in RAG: definition and limits.
- Restrict explicitly. "Only answer using the attached documents. If the information is not there, write: not found in the documents."
- Allow for doubt. "If you're not sure, say so and indicate your level of confidence." This counterbalances the model's tendency to respond at all costs.
- Demand that the passage be quoted. "For each statement, quote the sentence from the document that justifies it, along with the page number." A quotation can be verified in ten seconds; a statement cannot.
- Enable web search for recent newsand open the suggested links: a link that exists does not always say what the model attributes to it.
- To calculate using a tool. For a table, request data analysis (the model writes and executes code) rather than a "mental" calculation.
- Ask neutral questions. “Is there case law on…?” rather than “What case law confirms that…”.
- Cut out. One specific question per request, and a new conversation per topic, rather than a long exchange that mixes everything up.
- Choose the right mode. Methods of reasoning and in-depth research involve more verification before answering; they are slower but more reliable on factual matters. See The right AI model for each task.
- Proofread a second time. Ask the model, in a new conversation, to point out the statements in the text that are not supported by the sources. Then verify the points you have raised yourself.
A prompt type that combines several of these techniques:
"Using only the three attached documents, list the service provider's obligations. For each obligation, quote the exact sentence in quotation marks and indicate the document and page number. If an obligation you expect is not included, write it down. Do not add any information external to the documents."
These reflexes aren't acquired by reading a list; they're developed through practice on your own files. That's the purpose of our training. Mastering Generative AIwhere each participant learns to use the tool on their own documents and then check the results. Our guide Prompt engineering provides the complete method for writing instructions.
How to check an AI response in five minutes
Verification doesn't need to be lengthy if it's targeted. Here's a simple checklist to apply before distributing any AI-generated content.
| Element | Question to ask oneself | How to check |
|---|---|---|
| Each number | Where is he from ? | Find the primary source; otherwise, delete it. |
| Each reference | Does she exist, and does she say that? | Open the link, search for the decision on an official basis, read the passage |
| Each quote | Was it spoken or written by that person? | Find the original text |
| Each proper noun | Does the person, the company, the product exist? | One search is enough |
| A document summary | Is each point included in the document? | Verify three statements about the text using a sampling method. |
| Calculations | Were they made by a tool? | Recalculate the total in the spreadsheet |
What the company needs to organize
Hallucinations are not treated solely at the individual user level. Three decisions fall under the purview of the organization:
- A written verification ruleAccording to the AI charter, all content produced using AI is verified by the person who publishes it, who remains responsible for it. The Air Canada case serves as a reminder that the company is accountable for the information generated by its tools.
- Control levels depending on the stakes A simple review for an internal email, verification of each source for a client document, double validation for a legal opinion or a decision affecting individuals. For systems that make or guide a decision, the AI Act mandates human oversight in high-risk cases; see AI Act: What your company needs to do.
- Tools linked to reliable sources An assistant connected to your up-to-date documents is far less likely to hallucinate than one that responds from memory. It's often the best investment to prevent hallucinations.
One last point, often overlooked: overconfidence. A study we discussed shows that AI users become more efficient but less aware of their own mistakes; see AI does not eliminate the Dunning-Kruger effect, it shifts itThe better the tool, the more vigilance tends to decrease, which is precisely when the remaining errors go unnoticed.
F.A.Q
What is a hallucination in AI?
This is a response produced by a generative AI that is false, invented or not based on available sources, but presented confidently as a fact: a reference that does not exist, a false figure, an invented quote.
Why does ChatGPT invent answers?
Because it predicts the most plausible text, without needing to consult any factual basis, and because its training has long rewarded a confident answer rather than an admission of ignorance. The phenomenon affects all assistants, not just ChatGPT.
Will the hallucinations disappear?
They decrease significantly with each generation of models, especially with web search and evolving reasoning methods, but no developer promises their complete disappearance. They are inherent to the very nature of language models. Therefore, it's essential to organize verification rather than waiting for a perfect version.
Is giving a document to the AI enough to prevent hallucinations?
This greatly reduces them, but doesn't eliminate them: a summary can still add nuance or conflate two points. Ask the model to quote the exact sentence that justifies each statement, and check a few points by sampling.
Which AI hallucinates the least?
Rankings change with each model release and depend on the type of task. In practice, how the tool is used matters more than the tool itself: an average assistant connected to your documents and properly logged makes fewer mistakes than a cutting-edge model queried from memory.
Who is responsible for a widely disseminated hallucination?
The person and the organization that disseminate the content. The courts have reminded lawyers sanctioned for fabricating legal precedents, such as Air Canada for its online assistant, of this. This is why the verification rule must be included in the company's AI charter.
Sources
- Kalai, Nachum, Vempala and Zhang, Why Language Models HallucinateOpenAI and Georgia Tech, September 2025.
- PPC Land, New research explains why language models hallucinate, September 7, 2025.
- University of North Carolina Law Library, Handling Hallucinations: A Reflection on the Last Three YearsJune 2026 (Mata v Avianca case, Damien Charlotin's base).
- LawSites, Court imposes sanctions on lawyers who filed bogus cases after relying on ChatGPT, June 2023.
- American Bar Association BC tribunal confirms companies remain liable for information provided by AI chatbot (Moffatt v. Air Canada), February 2024.
- Associated Press, Deloitte to partially refund Australian government for report with apparent AI-generated errors, 7 October 2025
- Damien Charlotin, AI Hallucination Cases Database.