AI hallucinations: why AI invents facts

AI hallucinations in one sentence

An AI hallucination is a factually incorrect statement that a language model formulates with absolute confidence. Typical examples are invented sources, false numbers, non-existent judgments, and quotes that were never said.

The tricky part is not the error itself, but the missing warning signal. A human who does not know something hesitates or says so. A language model formulates equally fluently in both cases. Therefore, hallucination is the greatest quality risk in professional AI use, and not one that can simply be configured away.

Faithfulness and factualness: two types of wrong

Imagine a translator who perfectly translates a recipe, but it says salt instead of sugar. The translation is faithful to the original, but factually wrong. Exactly this distinction is important with AI.

Fraunhofer separates two types of errors. Faithfulness means fidelity to the provided source. Factualness describes the factual correctness in the real world. An answer can follow the attached source exactly and still be wrong. It can also be right without being in the source. If you only check one level, you only find half the errors.

Four manifestations

TÜVIT distinguishes four types of AI hallucinations.

• Sentence contradiction: One sentence contradicts another within the same answer.
• Prompt contradiction: The answer deviates from your instruction, for example regarding the desired format.
• Factual contradiction: The statement contradicts established knowledge about the world.
• Random hallucination: A detail appears that does not belong to the question at all.

Why the term is controversial

The term hallucination actually describes a sensory illusion. It attributes a human experience to the computer that it does not have. The word confabulation circulates as a more precise alternative. This means that a memory gap is filled with invented material.

These errors also exist outside of texts. Image models draw hands with six fingers, object recognition assigns patterns incorrectly, and transcriptions invent word fragments. This article covers language models because that is where the greatest risk arises in everyday work. The image part is covered in generating AI images.

The causes: five levels

AI does not hallucinate because of a software bug. Invented facts arise from five simultaneously acting properties of the technology.

1. The model knows nothing, it formulates

A language model does not store facts in a queryable database. It has merely learned which word sequences likely fit together. When asked for a study, it therefore generates something that looks like a study reference. It combines a plausible author, a year, and a journal title. Whether this combination really exists is unimportant for the generation process. The technical basis for this is explained in how AI works.

2. Knowledge is compressed, not stored

Training data is condensed into model parameters, and this compression is lossy. Frequent facts survive this process very well. Rare details, however, blur. This affects marginal figures, small companies, niche products, and precise numbers. Models are therefore reliable for general knowledge and unreliable for special details. This is exactly the opposite of what users usually expect.

3. Gaps are filled instead of marked

Language models are optimized to provide a helpful answer. From training with human feedback, they have learned that complete answers are rated better than admitting ignorance. This reward structure creates a systematic incentive to guess. OpenAI describes this mechanism in the publication "Why Language Models Hallucinate" from 2025. As long as leaderboards measure this way, this incentive remains.

4. Randomness in formulation

During text output, the model does not always choose the most likely next building block. It draws randomly from the best candidates. This ensures linguistic variety, but also deviations. The more creative the setting is, the higher the risk of invention becomes. Two identical prompts therefore rarely deliver two identical answers.

5. The context itself is flawed

Even a source connection does not completely solve the problem. If the retrieved documents are contradictory or outdated, a wrong answer with a correct-looking source reference arises. This is dangerous because the footnote creates a trust that the content does not deserve. An analysis by Imperial College London, Stanford, and the Internet Archive classified around 35 percent of new websites as AI-generated by mid-2025. If you use the open web, you often rely on model outputs. How you bind answers to your own sources instead is shown in the article on RAG.

Where hallucinations occur particularly often

Hallucinations follow clear patterns. Certain task types are real risk zones.

How you recognize a hallucination

Recognizing a hallucination does not mean guessing. There are clear signals in the text and three reliable checks.

Four warning signals in the text

• The source reference is unusually smooth. Author, year, title, and journal fit together perfectly, as if they came from a form field.
• A number is noticeably round. 30 percent, every third person, a doubling. Real measurements are usually unround in reality.
• A proper name stands without context. A person, an institute, or a product suddenly appears without the origin being explained.
• The level of detail does not fit the question. You ask very generally and immediately get a file number or an exact date delivered.

Three checks that work

1. Open the source and look for the passage. Do not just check if the link exists. You must find the statement in the document. If the passage is missing, the reference is incorrectly cited.
2. Independently cross-check proper names and numbers. Use a second channel with access to the original data. Look in a register, an official gazette, or a specialized database.
3. Ask the same question cold again. Open a new chat without prior history. If the details deviate now, the model guessed beforehand.

The common misconception: asking the model

Asking the model if it is sure is not a real check. The self-assessments of language models have little to do with actual correctness. A model very willingly confirms an invented source upon request. The most famous court case on this topic failed exactly because of this.

Documented cases and what they teach

Real cases show how quickly invented facts slip into professional work. They also show where human control failed.

2023 to 2024: court files and chatbot information

The invented jurisprudence catalog. In the proceedings against Avianca before the Southern District of New York, lawyers filed a brief in 2023. This relied on several court decisions, none of which existed. A chatbot had generated them including file numbers and confirmed them upon request. The court sanctioned the lawyers. Lesson: asking a model about the correctness of its own output does not work.

The expensive chatbot advice. A chatbot from Air Canada described a fare rule to a customer that did not exist at all. The Civil Resolution Tribunal in British Columbia decided in February 2024 that the company must answer for this information. Lesson: what your AI system says to the outside world, you say.

The error in the product demo. In the public Bard demo in February 2023, the AI incorrectly assigned an astronomical first image to the James Webb Telescope. Experts corrected this within hours. Lesson: hallucinations catch even the manufacturers themselves in perfectly prepared situations.

Since 2025: hallucinations in reports and studies

The current cases often concern consulting firms. Deloitte Australia delivered an audit report in 2025 with invented source references and an incorrectly reproduced court judgment. KPMG took a report on agentic AI offline in 2026 because the Financial Times had found invented case studies in it. EY withdrew a study on cyber risks in the same year because many footnotes pointed to non-existent sources.

These three cases hit the core of daily work. The errors did not arise primarily in the model, but because the checking step before it was missing. They happened in exactly the documents that B2B teams produce every day.

Continuously updated: the case database

Court decisions with hallucinated citations are collected in a public database. It appears at damiencharlotin.com and records these cases worldwide. This list is more reliable than any article because it does not age.

No external study tells you how high the error rate is in your own work. Only your own sample answers that. Take ten typical technical questions from your field. Ask them with and without web research, open every mentioned source, and count how often the statement is actually in it.

Countermeasures: what really helps

Hallucinations cannot be avoided with a single trick. You need a combination of measures, sorted by their effect.

1. Provide sources instead of querying. Give the model the relevant documents. Instruct it to answer exclusively from them. "What does contract law say about X?" becomes "Answer based on the attached contract. If the answer is not in it, say so." Models can read much better than remember.
2. Force web research with source obligation. Activate the research mode for current topics. Insist on evidence and then actually open this evidence. Deep research with AI shows where the limits lie.
3. Explicitly allow uncertainty. One sentence in the prompt noticeably changes the behavior. "If you are not sure about a detail, mark it as uncertain instead of guessing." Writing better prompts provides further patterns.
4. Break down tasks. Long assignments increase the risk because the model has to bridge gaps. Small steps with an intermediate check are significantly more reliable.
5. Outsource calculations. Numbers belong in code execution or in a table. They have no place in continuous text.
6. Cross-check by a second model. This finds some of the errors, but only if the source is in the context. Without a source, two models simply evaluate each other as plausible.

A QA process for the team

Individual tricks are not enough when several people work with AI. What really works is a tiered process with clear stages and responsibilities. Step 1: Classify content by risk. Not every text needs the same depth of checking.

Step 2: Define checking rules. Three rules cover the majority. Every number needs a source. Every source is opened. Every proper name is cross-checked. The corporate AI policy shows how you formalize this.

Step 3: Name responsibility. For every published piece of content, there is exactly one person responsible by name. Whoever approves, is liable. The Air Canada case shows that false information from your own system is binding. The EU AI Act training requirement demands that the people who are responsible for AI outputs know their limits exactly.

Step 4: Collect errors. Keep a short list of hallucinations that have occurred. This makes the risk zones in your own field visible and provides good material for training. The table from step 1 can be adopted directly as a template. Simply add a column for the responsible person and the date.

Residual risk: plan for it honestly

Hallucinations do not disappear. They are a property of the technology and not a temporary state. Even source-supported systems are measurably off. An investigation by the European Broadcasting Union with the BBC in October 2025 found at least one clear flaw in 45 percent of the checked chatbot answers. Grounding lowers the error rate, but it does not eliminate it.

An example from drug discovery shows that inventions are not always a problem. There, models are intentionally allowed to propose new molecules, which are then tested in the laboratory. The checking authority is firmly built in, while it is often missing in everyday office life.

A simple decision rule follows from this. Use AI where an error is noticed and correctable. Drafts, structures, and summaries of available sources are ideal. Everything where an unnoticed error lands directly with the customer or in public needs a human in between. Drawing this line is a management decision. If you need help with this, we will gladly support you in setting up AI processes with a checking step.

This topic is not to be confused with the question of whether a text comes from an AI. For this, there is the article detect AI content and our analysis of AI text detectors.

Conclusion

AI hallucinations arise because language models formulate instead of looking things up. They lose rare knowledge during compression and prefer to answer rather than remain silent. A hallucination is not a malfunction, but a normal side effect of how they work.

The effective countermeasure is not the better model, but the better approach. Provide sources, allow uncertainty, break down tasks, and outsource numbers. Know the warning signals and establish a tiered checking process in which a human answers for every piece of content. You can find the technical basics for this under Large Language Models: basics.