A chatbot has just handed you a court case that does not exist, a quote nobody said, or a statistic with a decimal point of false precision. That is a hallucination, and it is the single most common way AI tools go wrong in real work. This guide gives you a fast triage routine to catch them, explains why they happen (the answer is more specific than "AI is unreliable"), and walks through the four fixes that reduce them most, in order of effectiveness. It ends with which models currently do best and when you should stop trusting the chatbot altogether.
Why this answer looks wrong: five causes, likeliest first
If you are looking at a suspicious answer right now, work down this list:
- You asked the model to recall something specific from memory. Names, dates, citations, prices and version numbers are the worst case. Paste the source into the chat and ask again.
- Web search or citations were off. Most assistants have a search or "browse" toggle. Turn it on and ask for links to every claim.
- The question had no good answer, and the model guessed anyway. Add "If you are not sure, say so" to the prompt and ask for a confidence note per claim.
- You used a general chatbot for a document question. Use a source-grounded tool such as NotebookLM, or a "chat with your files" mode, so the answer must come from the text.
- The task was outside what any model can know. Private data, events after the model's training cutoff, or niche technical details. No prompt fixes this; verify elsewhere.
Symptoms: fake citations, invented quotes, confident wrong numbers
Hallucinations rarely look like errors. They look like the most fluent paragraph in the conversation. IBM's definition is a useful one: output that is nonsensical or inaccurate but presented as if it were real. The symptoms cluster into a few recognizable shapes:
- Fabricated sources. A plausible author, a plausible journal, a plausible year, and no such paper. Legal citations are notorious; a Stanford RegLab study found general-purpose models hallucinated on 69 to 88 percent of specific questions about federal court cases.
- Invented quotes. The model produces a sentence a real person "said" that matches their views and never left their mouth.
- Wrong specifics inside a right answer. The explanation is fine; the price, date, or command flag inside it is not.
- Confirmation of your own mistake. Ask "why did X happen in 2019" when it happened in 2021, and many models will explain the 2019 version.
- Consistency failures. Ask twice and get two different numbers. That is diagnostic: a model that knows the answer gives the same one.
The common thread is that the model is optimizing for a plausible continuation of the text, and a plausible continuation of "According to a 2022 study" is a study.
Why it happens: models are rewarded for guessing
For years the standard explanation was "training data is noisy." That is true but incomplete, and a paper from OpenAI in September 2025 gives a sharper account. In Why Language Models Hallucinate, Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang argue that hallucination has two roots. The first is statistical: during pretraining a model cannot tell a true rare fact from a plausible false one, so some errors are inevitable for facts that appear once or never in its data.
The second root is the one you can do something about. Most benchmarks score a model on accuracy alone, so "I don't know" earns zero points and a guess earns points some of the time. As OpenAI's summary of the paper puts it, models behave like students on a multiple-choice exam who have learned that a blank answer never scores. Training and evaluation reward confident guessing, so confident guessing is what you get.
The authors' proposed fix is to change how existing benchmarks are scored, so abstaining is not penalized relative to a wrong answer, rather than adding yet another hallucination test. Until that happens across the industry, the practical lesson for you is the mirror image: when you prompt, make it explicitly acceptable to abstain.
A related and slightly alarming data point: more reasoning does not automatically mean fewer hallucinations. On OpenAI's own PersonQA benchmark, TechCrunch reported from the system card that o3 hallucinated on 33 percent of prompts against 16 percent for the earlier o1, largely because it made more claims overall. Confidence and capability are not the same axis.
Quick checks: the 60-second verification routine
You do not need to fact-check everything; you need to fact-check the load-bearing specifics. This routine takes about a minute per answer.
- Circle the specifics. Names, numbers, dates, titles, URLs, commands, and anything in quotation marks. Everything else is explanation and can be judged on whether it makes sense.
- Search the exact string. Put a claimed title or quote in a search engine in quotation marks. A real paper or quote surfaces immediately; a fabricated one returns nothing or only other chatbot output.
- Open the link. If the model gave a URL, click it. A surprising share of hallucinated citations use a real domain with a made-up path.
- Ask the model to re-derive it. In a fresh conversation, ask the same question. A different answer means at least one of them is wrong.
- Ask for the source of each claim. Then check whether the source actually contains the claim, not just whether it exists. Models frequently attach a real paper to a claim the paper does not make.
If a claim fails step 2 or 3, treat the whole answer with suspicion, not just that sentence.
Fix 1: Give the model the source instead of asking it to remember
This is the single most effective change. A model asked to recall a fact from training is doing a lossy retrieval from a compressed memory. A model asked to read a document you pasted is doing a task it is good at: finding and restating text in front of it.
In practice: paste the article, the spec sheet, the email thread or the PDF into the chat before asking. Then phrase the question so that the answer must be located in the text ("Using only the document above, what does it say about...?"). Ask for the sentence or section the answer came from. If the model cannot point to one, it did not find it.
This works because grounded summarization is the task where models hallucinate least, by a wide margin. Vectara's hallucination leaderboard, which measures how often a model introduces unsupported claims when summarizing a document it was given, has its best models in the low single digits. The same models asked open-ended factual questions with no source in front of them score far worse, as the o3 numbers above show.
Fix 2: Ask for uncertainty and "I don't know"
Because the model's training rewarded guessing, your prompt has to explicitly un-reward it. A few phrasings that measurably help:
- "If you are not confident, say so rather than guessing."
- "For each factual claim, add a confidence level: high, medium or low."
- "It is fine to answer 'I don't know' or 'I cannot verify this.'"
- "List anything in your answer that I should verify independently."
These will not stop a model from being wrong, but they change the shape of the wrongness. You get flagged claims instead of smooth prose, and flagged claims are what the 60-second routine is designed to catch. Be aware that some models over-hedge when asked this way; if every sentence comes back "medium confidence," narrow the instruction to specifics only.
Fix 3: Turn on web search, citations or reasoning
Most consumer assistants can now browse, and browsing changes the task from recall to reading. When search is on, ask for a link on each claim and then apply step 3 of the routine. A link the model actually retrieved is worth more than one it "remembers."
Reasoning modes (extended thinking in Claude, "think longer" in ChatGPT, thinking modes in Gemini) help with problems that have a checkable structure: math, code, logic, multi-step planning. They help less with pure fact recall, and as the o3 example shows they can make a model more willing to assert. Use reasoning for reasoning problems and search for factual ones; do not treat "thinking" as a truth switch.
Fix 4: Use a grounded tool for document questions
For anything that lives in your own files, use a tool built to answer only from those files. Google's NotebookLM is the best-known example: it is designed, in Google's words, to give grounded answers with in-line citations pointing back to the specific passage in your uploaded sources, and to decline questions the sources do not cover. ChatGPT Projects, Claude Projects and enterprise "chat with your documents" tools work on the same principle.
The trade-off is scope. A grounded tool will not fill gaps with general knowledge, which is exactly the point. If you need general knowledge as well, use the grounded tool for the document facts and a normal chat for the rest, and keep track of which is which.
Which models hallucinate least right now
The honest answer is that it depends on what you measure. The same generation of models can score under 5 percent on grounded summarization and over 30 percent on open-domain recall, as the Vectara and PersonQA figures above show. There is no single "most truthful model."
For grounded summarization, which is the task most closely matching Fix 1, the Vectara leaderboard as of this writing puts several widely available models in a tight band:
Scroll sideways to see the full table →
| Model | Hallucination rate (summarization) |
|---|---|
| Gemini 2.5 Flash-Lite | 3.3% |
| Phi-4 | 3.7% |
| Llama 3.3 70B Instruct | 4.1% |
| Mistral Large (Nov 2024) | 4.5% |
Two caveats. These numbers move with every leaderboard refresh, so check the source before quoting them. And a low score here tells you the model sticks to a document you gave it; it says nothing about how it behaves when you ask it a trivia question. If your use case is the trivia question, the relevant benchmark is closer to PersonQA, and the numbers there are much higher for every vendor.
For a plain-language guide to what these models can and cannot do, and how their paid tiers compare, see our comparison of ChatGPT Plus, Claude Pro and Google AI Pro.
When to stop trusting the chatbot entirely
Some jobs should not be done by a general chatbot at all, no matter how you prompt it:
- Legal, medical and financial specifics. Citations, dosages, statutory text and tax thresholds. The error rates in these domains are high and the cost of an error is higher.
- Anything you cannot check. If you have no way to verify a claim, you have no way to know whether it is a hallucination, and the model has no way to tell you.
- Events after the training cutoff without search on. The model will often produce a confident answer anyway.
- Numbers that will be reused. A hallucinated figure that goes into a spreadsheet propagates. Get it from the primary source.
In these cases the model is still useful as a drafting and explaining tool. It is not useful as a source. The rule of thumb that survives every model generation: let AI write the sentence, but never let it supply the fact.
Sources and further reading
- Why language models hallucinate (OpenAI)openai.com
- Why Language Models Hallucinate (arXiv 2509.04664)arxiv.org
- OpenAI's new reasoning AI models hallucinate more (TechCrunch)techcrunch.com
- Hallucinating Law: Legal Mistakes with Large Language Models are Pervasive (Stanford Law School)law.stanford.edu
- Hallucination Leaderboard (Vectara, GitHub)github.com
- What Are AI Hallucinations? (IBM)ibm.com
- LLM Hallucinations: A Guide (Lakera)lakera.ai
- Learn about NotebookLM (Google Help)support.google.com