AI Hallucinations Explained: Why Chatbots Make Things Up

Ask a chatbot for five sources on a narrow subject. You may get five citations that look perfect, and two of them may not exist. The authors are real. The journal is real. The paper never was. This failure has a name, hallucination, and it is the main reason to keep your guard up.

The word suggests a glitch, as if a patch could remove it. It cannot. Fabrication comes from the same machinery that writes the good sentences. A model incapable of inventing anything would also struggle to compose a paragraph you had not already given it.

That does not leave you helpless. Hallucination is predictable. It clusters around certain kinds of question, it leaves tells on the page, and it drops sharply when the tool reads real documents before answering. This guide covers all three.

What an AI hallucination actually is

A hallucination is output that a model presents as fact when it is false, unsupported or wholly invented. Common forms include a made-up citation, a misattributed quote, a wrong founding date and a biography that merges two people. The content is wrong. The delivery is not.

That second point is what catches people out. Fabricated text arrives in the same steady voice as accurate text. There is no hedge, no stumble, no change of tone. The writing itself gives a reader no signal at all.

Why some researchers prefer the word confabulation

Psychologists use confabulation for something close to this. A person with a gap in memory fills it with plausible detail and believes the result. There is no intent to deceive, and the story holds together.

That word fits a language model better than "hallucination", which implies false perception. A model has no senses and no beliefs. It has a gap, and a very strong habit of filling gaps smoothly.

Errors that are not hallucinations

The term is more useful when it stays narrow. Several common complaints belong in other categories.

• Stale knowledge: training stops at a knowledge cutoff date, so recent events are missing rather than invented.
• Misread questions: an ambiguous prompt can produce an accurate answer to the wrong question.
• Arithmetic slips: a model working through a calculation can drop a digit halfway.
• Requested fiction: a short story, a mock press release or a sample dataset is compliance, not fabrication.
• Contested claims: where experts disagree, a confident answer is a poor summary rather than an invention.

Reserve the word for confident invention. That is the failure that carries no warning label, and the one worth building habits around.

Why a fluent model invents

Three forces combine here. How the model generates text, how it was scored during development, and how much it can physically store. None of them is a bug.

Prediction is not retrieval

A language model writes by predicting the next token, a word or word fragment, over and over. It chooses likely continuations from patterns learned in training. That loop is traced step by step in how AI chat assistants work.

Nothing in the loop is a lookup. The model holds no index of documents and no table of facts to consult. Where a fact appeared often in training, the likely answer and the true answer usually match. Where it appeared rarely, likeliness carries on alone.

Citations show this plainly. The model has seen millions of references, so it knows their shape exactly: author, year, title, journal, volume, pages. Asked for one it does not hold, it builds the shape and fills it in.

A chatbot is not lying to you. Lying needs a known truth to work against, and the model only has a sense of what usually comes next.

Scoring has rewarded the confident guess

Models are measured on benchmarks, which are large sets of test questions with known answers. A right answer scores. An honest "I do not know" scores nothing. Under those rules guessing wins, exactly as in a multiple-choice exam with no penalty for a wrong tick.

Human feedback pushes the same way. Reviewers who rank draft answers tend to prefer complete, helpful ones over hedged ones. A model tuned on those rankings learns to sound sure of itself.

Compression blurs the rare facts

Training squeezes patterns from trillions of words into a fixed set of parameters, the numeric weights inside the model. That squeeze is lossy, like a heavily compressed photograph. Frequent facts stay sharp. Rare ones blur into their neighbours, so one author's book drifts to a contemporary and one company's milestone lands on a rival.

Sampling adds a final wobble. Assistants choose among likely next tokens with a controlled dose of randomness, which is what stops the prose reading like a form letter. Ask the same question twice and the path can differ, sometimes away from the correct answer.

Which questions carry the highest risk

Fabrication is not spread evenly across a conversation. It gathers where the true answer is specific, rare and checkable. The table below sorts the common cases.

Other AI tools fail differently, which is worth knowing. Speech-to-text swaps similar-sounding words, and the damage is counted by a public metric explained in our guide to word error rate. A generative assistant has no equivalent single score, because an invented paragraph has no correct version to measure it against.

One case surprises careful users. A model can hallucinate about a document you supplied yourself. Under pressure to read smoothly, a summary may merge two findings or turn "suggests" into "shows". Long reports deserve a section-by-section check.

How to catch a fabrication in a minute

Verification does not have to be heavy. Most checks take seconds, and you only need them on the claims your decision actually rests on.

1. Open every source. A reference proves nothing until the page loads and supports the exact claim attached to it.
2. Test the identifier. Paste the DOI, the digital object identifier printed on most papers, into a search box. A real DOI leads to one specific work.
3. Search the quote. Put a quoted sentence inside double quotation marks and search it. Real quotes leave a trail; invented ones return nothing.
4. Ask again in a fresh chat. A consistent answer is not proof, but a changed answer is a clear signal that you have reached the edge of the model's knowledge.
5. Allow the unknown option. Tell the assistant to write "unknown" instead of guessing, and to mark low-confidence lines. Phrasing that pays off here is covered in Prompt Writing 101.
6. Match effort to stakes. A pub-quiz fact needs no check. Anything filed, published, prescribed or signed needs the full routine.

Warning signs on the page

Some answers deserve a second look before you check anything at all. These patterns are the common tells.

• A title that fits too well: a cited paper whose title answers your question word for word was often built from your question.
• Round numbers everywhere: real measurements are untidy, so unusually clean figures suggest invention.
• No hedging on a live debate: a settled tone on a contested subject usually means the detail has been flattened.
• Missing or dead links: a source list with no URLs, or with links that fail to load, is the classic pattern.
• The wrong venue: a clinical trial credited to a computing conference points at a stitched-together reference.

What actually lowers the rate

Three approaches have made a measurable difference, and two of them are available to you today.

The strongest is retrieval-augmented generation, usually shortened to RAG. The system searches a document set or the web first, then asks the model to answer from those passages and cite them. The work shifts from recall to reading, and fabrication falls sharply. Browsing modes and document upload are the consumer form of the same idea. They often sit behind a subscription, a boundary drawn in free and paid AI tools.

The second lever is your own prompt. Supplying the source text, narrowing the question and permitting "unknown" all remove the pressure to invent. The third sits with the labs: training that rewards calibrated answers, and benchmarks that score fabrication directly. A broad research survey of hallucination in large language models maps both the causes and the current defences.

Organisations increasingly treat this as a managed risk rather than an embarrassment. The NIST AI Risk Management Framework asks teams to measure, document and monitor exactly this kind of failure. Wikipedia's article on hallucination in artificial intelligence is a fair starting map of the wider literature.

Three myths worth dropping

Bigger models will fix it. Scale has cut error rates and keeps cutting them. It does not change the mechanism. A larger model still predicts text, and still meets questions its training never covered.

The model knows when it is guessing. Not reliably. Stated confidence is a writing style the system learned, not a reading from an internal gauge. The phrase "I am certain" carries no evidence whatsoever.

Only obscure topics are affected. Obscurity raises the odds, but a familiar subject with one rare detail fails too. A famous company with an unusual founding date is a typical trap.

Setting your own tolerance for error

Engineers do not ask whether a part is perfect. They ask what deviation the job allows, then measure against it. That discipline is the subject of how precision engineering controls quality. AI output deserves the same treatment.

Decide the tolerance before you start. Rough drafts can carry a high error rate, because you will rewrite them anyway. Anything published under your name carries a much tighter one. Students and researchers face the tightest limits of all. A full workflow for them appears in using AI assistants for research and study.

Measuring your own tools is easy enough. Ask an assistant three questions you already know the answers to, and make one of them obscure. A free tool such as ASKAI.FREE works for the test. Five minutes of that teaches more than any list of warnings.

Rates keep falling, generation after generation, and grounded products already invent far less than raw models once did. No serious researcher promises zero, because the mechanism that writes is the mechanism that invents. That is a limitation to work with, not a scandal to wait out.

The habits that contain it are old ones. Check the source. Match confidence to evidence. Slow down as the stakes rise. They sit alongside the wider practices in responsible AI use. More of the method waits in our topic collection on AI accuracy and across the AI Assistants hub.