AI hallucinations explained: why models make things up and how to catch it

An AI hallucination is output that reads as confident and correct but is false. The model did not look something up and misread it; it composed a plausible answer out of patterns, and the answer happened not to match reality. A citation to a paper that does not exist, a statistic with no source, a quote nobody said, a feature a product never shipped, a clause that is not in the contract you pasted: all hallucinations.

They are the single biggest reason professionals get burned by AI, and they are also the most manageable risk once you know where they cluster.

What a hallucination is, and what it is not

It helps to be precise, because the word gets used loosely.

A hallucination is: an invented fact, source, number, name, date, or detail presented as true. The defining feature is that the model produced it with the same fluency and confidence it uses for things that are correct.

A hallucination is not: an opinion you disagree with, a draft in the wrong tone, an answer based on outdated training data that was true at the time, or a mistake caused by your ambiguous prompt. Those are real problems with different fixes.

The distinction matters because hallucinations are not marked. There is no formatting difference between "Paris is the capital of France" and "A 2023 industry study found a 34 percent drop in churn," a sentence invented for this paragraph exactly the way a model would invent it. Both arrive in the same confident voice.

Why it happens

A large language model generates text by predicting the next likely token, one at a time, based on patterns learned during training. It has no database of facts and no lookup step. When the question touches something the model saw thousands of times, prediction and truth line up. When it touches something rare, absent, or newer than its training data, the model still has to produce the next token, and the most likely one is whatever makes the sentence sound right.

Three things make this worse:

• It learned what good answers look like. Training data is full of well-formed citations, precise-sounding statistics, and authoritative summaries. The model learned the form of a good answer extremely well, so its guesses come out formatted like facts.
• It was rewarded for being helpful. Post-training pushes models toward complete, confident responses. "I don't have reliable information on that" is a response vendors have had to teach deliberately, and it still loses out to a plausible guess more often than you would like.
• It cannot check its own work against the world. Unless a tool connects it to search or to your documents, it has nothing but its patterns to consult. Asking "are you sure?" just triggers another round of prediction.

For the deeper mechanics, read How large language models work. For the working version: the model is fluent first and factual second, and every design choice you make should account for that order.

Where hallucinations are most likely

Risk is not evenly spread. It concentrates in a handful of situations, and you can learn to recognize them on sight.

Citations and sources

The most famous example is real. In 2023 a federal judge in New York sanctioned two lawyers who had filed a brief citing court cases that ChatGPT had invented (the case was Mata v. Avianca). The citations looked like every other citation. None of them existed. Every professional who cites anything should assume the same thing can happen to them until they have opened the source.

Numbers and statistics

Ask "what percentage of small businesses use AI?" and you will get a number. It will be round, it will sound reasonable, and it may be attributed to a real-sounding organization. Whether a real survey produced it is a separate question the model cannot answer. The safe pattern is to find the number yourself, paste it with its source, and let the model do the writing.

Niche and specific facts

Models are strongest on well-documented topics and weakest on the long tail: a regional regulation, the founder of a 30-person firm, the exact spec of a part number, what a particular clause in your industry's standard contract means. The less written material exists about a thing, the more the model is guessing, and the guess is delivered in the same voice as its best-supported answer.

Recent events

Every model has a training cutoff, and without a search tool it does not know what happened afterward. Worse, it may not know that it does not know, so it describes the world as of its cutoff as if that were today. Prices, executives, product features, laws, and deadlines all go stale. If a fact could have changed in the last year, treat it as current only when a search-grounded tool shows you a dated source.

Warning signs to train yourself on

None of these prove a hallucination, but each should trigger a check:

• Specificity without provenance. A precise figure, date, or name that arrived with no source and that you did not supply.
• Perfectly formatted references. Author, year, journal, page numbers, all present and tidy. Real references found by search tend to come with a link; invented ones tend to come with confidence.
• Round, convenient numbers. Statistics that land neatly on 30, 50, or 75 percent deserve extra suspicion.
• Agreement with your framing. You asked "why did revenue fall because of the price change?" and it explained why, without questioning whether it did.
• Confident answers to obscure questions. If you would struggle to find the answer in ten minutes of searching, be suspicious that the model produced it in two seconds.
• Details that appear in a summary but not in your memory of the source. Especially names, amounts, and deadlines.

A worked example

A consultant asks a chat assistant, with search turned off, "Summarize what [a mid-sized regional logistics company] announced in its most recent quarterly update." The reply is fluent: a revenue figure, a new terminal opening, a leadership change, and a quote from the CEO. It reads like a press release.

Nothing in it is verifiable. The company is private and publishes no quarterly updates. The model produced what a quarterly update usually contains, with plausible specifics filled in, because the request assumed one existed.

The same question in a search-grounded tool returns a different answer: it finds a local news item from the prior year, a job posting, and nothing resembling a quarterly update, and says so. That is the answer you want. A model admitting it found nothing is worth more than a model that found something that is not there.

The lesson is not that the first tool is bad. It is that the request asked for a niche, recent, specific fact with no source in the context window: three risk factors at once. Change the request or change the tool.

How to reduce hallucinations

You cannot eliminate them, but you can push the rate down hard with five habits.

1. Supply the facts; ask for the language. Paste the data, the source text, or the notes, and tell the model to use only that material. This is the single most effective move you can make.
2. Use grounded tools for factual questions. Search-enabled modes in ChatGPT, Claude, Gemini, and Copilot, and research tools like Perplexity, put real sources into the model's context. NotebookLM answers only from documents you upload. Grounding raises the floor; it does not remove the need to click the sources.
3. Ask for uncertainty out loud. Tell the model to label each claim as verified from the material, inferred, or unknown, and to list what you should check. Models are reasonably good at flagging their own low-confidence claims when explicitly asked.
4. Ask for quotes, not paraphrases, on anything that matters. "Quote the exact sentence from the document that supports this" makes fabrication much harder, because an invented quote will not be in the text when you search for it.
5. Verify with a source, never with the model. Asking "are you sure?" produces reassurance, not verification. Open the document, run the number, click the link.

A copy-ready prompt that forces uncertainty into the open

Use this whenever accuracy matters more than speed.

You are a meticulous research assistant who would rather say "unknown" than guess.

Task: Answer the question below using only the material I provide and, if you have it, web search. Do not draw on memory for specific facts, numbers, names, or dates.

Question: [YOUR QUESTION]

Material: [PASTE ANY DOCUMENTS, DATA, OR NOTES, OR WRITE "NONE"]

Format: Give the answer as a short list of claims. After each claim, add one label: VERIFIED FROM MATERIAL (with the exact quote and where it appears), VERIFIED FROM SEARCH (with the link), or UNVERIFIED (you could not confirm it). End with a section titled "What I could not confirm" listing anything you were tempted to state but could not support.

Constraints: If the material does not answer the question, say so plainly rather than filling the gap. If my question contains an assumption you cannot confirm, point it out before answering. Ask me up to two clarifying questions first if the question is ambiguous.

The verification checklist

Copy this into your prompt library and run it on anything AI touched before it leaves your hands.

• Citations: every link, case, or study opened and read, and it says what the AI claims it says.
• Numbers: each traced to a source I can name, or recomputed myself.
• Quotes: each found verbatim in the original.
• Names, titles, dates, deadlines: confirmed against a record.
• Summaries: spot-checked against the source document in at least three places.
• Recent events: backed by a dated source from a search tool.
• Claims about my own business: match what I actually know.
• Anything unverified: removed, or explicitly marked as unverified for the reader.

If the checklist feels heavy, remember that the point of the AI is to save you the drafting time, and that checking a draft is far faster than writing one. The full step-by-step process is in How to fact-check AI output.

Next steps

• Run the full process in How to fact-check AI output.
• Understand the mechanics in How large language models work and the glossary entries for hallucination, grounding, and knowledge cutoff.
• Choose tools that show sources: Perplexity for research and NotebookLM for your own documents.
• Build the habit into your routine with How to use AI at work, and into your team's rules with How to write a team AI policy.
• Prompts for careful research live in research and learning.