An AI hallucination is output from a language model that sounds convincing but is factually wrong or made up, such as a source or figure that does not exist. It is the best-known risk of working with a language model, and also the least visible: a hallucination arrives in the same fluent, confident tone as a correct answer, so the form gives nothing away. If you do not check the substance of the output, you only notice the difference when someone else does.
The cause lies in how the model works, not in a malfunction. A language model predicts the next word based on patterns and has no built-in check against reality. In professional work, that produces three recurring forms: made-up sources, made-up figures and made-up technical details. All three look plausible, and that is exactly why they slip through.
Spotting them starts with checking the content against an independent source. Preventing them starts earlier, with the question you ask and the check you build in afterwards as a fixed step. You cannot rule them out completely, but you can reduce the risk considerably.
How much damage a hallucination does depends on what you do with the output. An internal note calls for a different check from advice that goes to a client. At firm level, the question is not whether someone is paying attention, but whether that check is recorded somewhere and can be demonstrated.
What is an AI hallucination?
An AI hallucination is output from a language model that is grammatically correct and convincingly worded, but factually wrong or entirely made up. The model presents that output with the same certainty as an answer that is correct, with no signal that anything is wrong. The error lies in the content, never in the form.
The term comes from psychology, where hallucinating means perceiving something that is not there. For a language model, the comparison only half works, because there is no perception and no intent. A recurring objection in the professional debate is therefore that confabulation is the more precise word: filling gaps with plausible-sounding content.
Hallucinations occur in every large language model, including ChatGPT, Claude and Gemini. There is no reliable percentage per model, because the likelihood varies widely by field, question type and answer length. Specific, factual questions about names, dates, quotes and figures carry the highest risk of fabrication.
Not every mistake a language model makes is a hallucination. With more complex tasks that have several conditions, a model sometimes follows part of the instructions and ignores the rest without saying so. That is an instruction error, not a fabrication. A hallucination concerns the factual content of the answer; an instruction error concerns the task that came before it. Both need checking, but in different ways.
Why does AI hallucinate?
AI hallucinates because a language model does not look up facts but predicts the most likely next word based on patterns in its training data. There is no checking step between that prediction and reality. The result fits the pattern, which is not the same as being correct. The same mechanism sits at the core of generative AI, which builds text through probability rather than from a source.
Four factors increase the likelihood of a hallucination. If the training data covers a topic only partly or not at all, the model fills the gap with an assumption. The vaguer the question, the more room the model has to guess. In long conversations and large documents, a model is more likely to lose the thread and mix up information from earlier in the conversation. And language models are trained to be helpful, so when the choice is between giving no answer and producing something anyway, the answer usually wins.
That last factor explains why a hallucination so often fits exactly what you hoped to hear. If you understand how AI works at word level, you also see why a model prefers to give a complete answer rather than half an answer with a gap in it. The model does not distinguish between what it draws from data and what it fills in.
Which types of AI hallucination do you come across at work?
In professional work, three types of AI hallucination do the most damage: made-up sources, made-up figures and made-up technical output. They share one problem: all three look exactly like a correct answer.
Made-up sources and quotes
Made-up sources are the hallucination that undermines trust fastest. During the period I used ChatGPT intensively, it regularly cited studies, reports and quotes that did not exist when I checked. A fabricated source also often comes fully formed: an author, a year, an institution and a title that sounds logical for the topic.
If you then ask the model whether the source really exists, it often just confirms that it does. That confirmation is the same kind of prediction as the original source and adds nothing to your certainty. Only opening the source counts as a check.
Made-up figures and facts
Made-up figures are particularly dangerous because a number looks reliable and is rarely recalculated. A recurring pattern I saw with ChatGPT was that, when given a keyword, it added a search volume unprompted, even though it has no connection to a keyword tool such as Ahrefs or SEMrush. If you have no access to that data yourself, you accept such a figure with no reason to doubt it.
Simple facts go wrong just as easily. A model can state with complete confidence that Sydney is the capital of Australia, when it is Canberra. With general knowledge like that, the error stands out quickly; with a figure from your own field, it often does not.
Made-up code and technical details
In programming, a hallucination produces code that looks correct but does not work. I regularly came across non-existent Python functions, WordPress hooks that are not documented anywhere and invented CSS properties. Usually, this only comes to light when you run the code, which costs time in the development process.
Models differ in how often this happens. When I switched from ChatGPT to Claude, I noticed that Claude is less quick to make up sources and figures, a difference that also shows up in the Claude vs ChatGPT comparison on factual questions. Checking remains necessary with both models.
How do you spot an AI hallucination?
You spot an AI hallucination by its content, never by its form, because the tone is just as confident as with a correct answer. Five checks bring most fabrications to light before the output goes any further.
- Ask for the source and actually open it, instead of asking whether the source exists.
- Ask the same question again in a new conversation and compare the two answers.
- Recalculate figures and check that percentages and totals add up.
- Watch out for precision without a reference, such as an exact number or a verbatim quote with no traceable origin.
- Put every summary next to the original document before you pass it on.
Summaries deserve extra attention here. During the period I used ChatGPT, it sometimes summarised documents in a way that sounded plausible but differed in substance from the source. If you only read the summary and do not go back to the original, you take on incorrect information without noticing anything.
A second warning sign is an answer that matches exactly what you hoped to hear. If something sounds too complete or too neat, that is a reason to check more carefully, not less. I use one question as a filter myself: would I publish this under my own name?
How do you prevent AI hallucinations?
To prevent AI hallucinations, you reduce the likelihood in advance and catch the rest afterwards with fixed checks. The first intervention lies in the question itself, because a well-defined question with context and criteria leaves the model less room to guess. That makes asking AI a question a checking step before there is any output.
The second intervention is requiring a source. Ask for verifiable references, and for factual questions use a model with web access or a set-up in which the model draws on your own documents instead of its training data. The model then takes the information from a source you can open.
The third intervention is the second reader. I bring the four-eyes principle from my years at major banks into AI work, because nothing goes out without a second check. Sometimes I use Claude as a second reader of ChatGPT output, or the other way round, so that one model tests the other’s assumptions. A second model does not replace human editing; it only filters out the most obvious errors.
The fourth intervention makes the rest traceable. In my own workflow, I log prompts and outputs, so I can check afterwards which question led to which answer and where an error crept in. Without that trail, a hallucination remains an incident; with it, it becomes a pattern you can catch in the next prompt.
When is an AI hallucination a real risk?
An AI hallucination only becomes a real risk when the output ends up somewhere without a person stepping in. One thing I take from banking practice is that not every task has the same risk profile. An internal note you will read through yourself needs less checking than a published article, a quote or advice that goes straight to a client.
That hierarchy determines how much checking a task deserves. The higher the stakes, the tighter the check, and the earlier that check has to come in the process rather than at the end. Responsibility for the output lies with the professional who uses it, not with the model.
Mata v. Avianca shows how hard that can hit: in that case, a New York judge imposed sanctions on two lawyers on 22 June 2023. Their court filing cited six rulings that ChatGPT had made up, complete with case numbers and passages, and the court imposed a fine of $5,000. Since then, the case has become the international reference point for lawyers using AI and for the question of how far a lawyer’s own duty to research extends when using AI.
For accountants, notaries and lawyers, the bar is higher than for marketing copy for the same reason. Professional rules place responsibility for the content on the professional, whatever tool they used.
How do you build hallucination checks into your organisation?
Checking for AI hallucinations only becomes reliable once it is recorded as a working agreement rather than an individual habit. A colleague who always checks and a colleague who never does both deliver work under the same firm name. Without an agreement, nobody knows which of the two did the work in front of you.
Such an agreement has three parts: which tasks AI may touch, which check goes with them and who carries out that check. Link those three to the risk class of the task, so that an internal summary does not go through the same process as a draft deed or an advice letter. Record what happens if the check finds something, because a check without a follow-up step is a formality.
Since 2 February 2025, this has also been a matter of law. Article 4 of the EU AI Act requires organisations to take measures to support the AI literacy of employees who use AI systems, and recognising incorrect output is directly part of that. Since 27 July 2026, it has been a best-efforts obligation: what counts are the measures an organisation takes, such as recorded agreements and training, not the individual knowledge of one enthusiastic colleague.
Offices that want to turn this from a habit into a way of working usually put in place an internal AI policy, a prompt library for each core task and fixed checkpoints. In the AI Literacy Course, these are exactly the elements that make the difference between a team that uses AI and a team that knows when not to trust the output.
Frequently asked questions about AI hallucination
After the explanation above, a few practical questions about AI hallucination often remain.
Does it help to give AI permission to admit uncertainty?
Yes, explicitly giving AI permission to admit uncertainty reduces the likelihood of a hallucination. Anthropic lists this in its documentation as the first measure: give Claude room to say it has too little information for a reliable judgement. Put a sentence to that effect explicitly in your prompt, so the model can say when it cannot give a reliable answer. An honest ‘no answer’ is more useful for professional work than a fabrication in a convincing tone.
Do AI hallucinations disappear with newer models?
No, newer models also produce AI hallucinations. OpenAI wrote in September 2025 that hallucinations remain a challenge for all large language models and that accuracy will never reach 100 per cent, because some questions cannot be answered. According to OpenAI, the standard tests for models reward guessing over admitting uncertainty. Checking therefore remains necessary, whichever model or version you use.
Is an outdated answer also an AI hallucination?
An outdated answer is not necessarily an AI hallucination, but it is often the result of a model’s knowledge cut-off. Every model is trained on data up to a certain point; Anthropic, for example, states a reliable knowledge cut-off of June 2026 for Claude Opus 5.5 and Sonnet 5.5. If you ask about something after that without web access, you get outdated information or the model fills the gap with a guess, and the latter is a hallucination.
