AI Hallucinations Explained: How to Fact-Check a Chatbot
Picture an answer like this from a chatbot: a small town in Japan holds a festival on March 3, the mayor’s name is included, and a local newspaper article is quoted. Now suppose the town is real, but the festival date is wrong, the mayor retired two years ago, and the quoted article was never written. Nothing in the answer’s tone would warn you. Invented details arrive in the same confident shape as true ones. That skepticism is a core skill in everything we cover in the AI automation series.
This behavior has a name: hallucination. OpenAI’s own help center defines it as a model producing responses that are not factually accurate, including wrong dates, fabricated quotes, and citations to sources that were never written. Anthropic says the same about Claude: it can write things that look correct but are very mistaken, including quotes that appear authoritative without being grounded in fact. Google builds a verification button into Gemini precisely because its answers need checking. This is not a defect one vendor has and another has solved. It is a property of how every large language model currently works.
The good news: you do not need to understand transformer architecture to protect yourself. You need a habit. This guide explains why chatbots invent facts, which claims are most likely to be false, and a four-pass workflow for checking an answer before you rely on it. The standard here is the same one this site applies to itself: claims should trace to a source, as laid out in our editorial policy.
Key takeaways
- Hallucinations are plausible but false statements, not bugs that appear occasionally. OpenAI’s research argues they are a predictable result of how models are trained and graded: standard evaluations reward a confident guess over an honest “I don’t know,” so models learn to guess.
- The most dangerous claims are the specific ones: dates, statistics, quotes, citations, names of papers and people. General explanations are usually safer than precise-sounding details.
- Built-in tools help but do not finish the job. ChatGPT can search the web and cite sources, Gemini has a Double-check button that color-codes statements, and Claude can cite the pages it used. Each of these still requires you to open the source.
- A reliable check has four passes: extract the checkable claims, run the built-in verification, open the primary source yourself, then mark each claim verified, corrected, or dropped.
- Match the effort to the stakes. A dinner recipe needs a skim. A medical, legal, or financial claim, or anything you plan to publish, needs the full workflow.
Table of contents
What an AI hallucination actually is
OpenAI’s research page on the topic gives the cleanest definition: hallucinations are plausible but false statements generated by language models. Two words in that sentence do the work. “False” means the statement does not match reality. “Plausible” means nothing about how it sounds warns you. A hallucination is rarely written in hesitant language. It arrives with the same confident structure as a true sentence, which is exactly why it works.
OpenAI’s help center lists what this looks like in practice: incorrect definitions, dates, or facts; fabricated quotes, studies, and citations; and overconfident answers to ambiguous or complex questions. Anthropic’s support documentation adds that Claude can produce quotes that look authoritative and convincing without being grounded in fact, and warns users not to treat the model as a singular source of truth.
One example from OpenAI’s own testing is worth remembering. When asked for the title of a specific researcher’s PhD dissertation, a widely used chatbot confidently produced three different titles, none of them correct. Asked for that researcher’s birthday, it gave three different dates, all wrong. The model was not malfunctioning in those moments. It was doing what it is built to do: producing the most likely-sounding continuation. When the true answer is rare or absent from training data, “likely-sounding” and “true” come apart.
Why chatbots make things up
The short version: because guessing is rewarded.
OpenAI’s September 2025 research summary argues that hallucinations persist partly because of how models are evaluated. Most benchmarks grade a model on accuracy: the percentage of questions it answers exactly right. Under that grading, a model that guesses when unsure scores better over thousands of questions than a model that admits uncertainty, because a guess occasionally lands while “I don’t know” always scores zero. OpenAI compares it to a multiple-choice test where leaving an answer blank guarantees zero points: students learn to guess, and so do models.
The same research notes that errors are worse than abstentions, and OpenAI’s own Model Spec says models should indicate uncertainty rather than state confident falsehoods. But the scoreboards the industry optimizes for still mostly reward the guess. Until that changes, the person using the chatbot carries the verification burden.
There are two more mundane causes worth knowing. First, knowledge cutoffs: a model’s training data ends at a point in time, so questions about recent events invite confident answers built from stale information unless the model searches the web. Second, training data is uneven. A model has seen millions of pages about famous topics and almost nothing about your local regulation, your niche software version, or a specific person’s biography. The rarer the fact, the more likely the model fills the gap with a plausible invention.
The claims most likely to be wrong
Not every sentence in a chatbot answer carries the same risk. A practical fact-check starts by sorting the answer into claim types, because the failure rate differs enormously by type.
Dates, statistics, and version numbers. These are the classic failure. They are precise, they sound authoritative, and they are exactly the kind of detail that is sparse in training data. Treat every unlinked number a chatbot gives you as unverified until you see it on a primary page.
Quotes and citations. Both OpenAI and Anthropic explicitly warn about fabricated quotes and references to sources that do not exist. A chatbot can produce a citation with a realistic title, a real-sounding author, and a plausible journal name, where the paper itself was never written. The format of a citation is easy to imitate; the existence of the paper is the thing to check.
Current events and fast-moving facts. Prices, plan limits, feature lists, office holders, software behavior. Anything that changes monthly is at risk from knowledge cutoffs. OpenAI’s help center notes that without search tools, answers come from training data with a fixed endpoint.
General explanations of established concepts. These are the safest content a chatbot produces, because the underlying material is abundant and stable. “What is caching” will usually be fine. “Does plugin X still do Y in the October update” will not be.
A useful rule of thumb: the more a claim would require the model to have memorized one specific line from one specific page, the more likely it is invented.
A four-pass fact-check workflow
This workflow assumes nothing about which chatbot you use. It works the same for ChatGPT, Gemini, Claude, or any other assistant, and it scales from a thirty-second skim to a full verification depending on what the answer is for.
Pass 1: Extract the checkable claims
Read the answer once and list only what could be true or false: the dates, names, numbers, quotes, citations, and direct recommendations. Skip the framing and the hedges. A typical chatbot answer contains five to fifteen genuinely checkable claims buried in prose that sounds authoritative but asserts little. Your list is the unit of work. Everything after this pass operates on the list, not on the answer as a whole.
Pass 2: Run the built-in checks
Each major chatbot now ships some form of self-check. They are worth the thirty seconds they cost, as long as you understand what they actually tell you.
Gemini has a Double-check button. Per Google’s help page, it uses Google Search to color-code the response: green highlights mark statements where Search found content likely similar, orange marks statements where it found content likely different or found nothing relevant, and no highlight means there was not enough information to evaluate. Two limits matter: the feature is not available for responses generated with extensions, and it skips content like code and markdown tables. A green highlight means “similar content exists,” not “this is true,” and Google says the linked page is not necessarily the one Gemini used.
ChatGPT, depending on your plan, can search the web and cite its sources, and OpenAI’s help center recommends checking those citations by visiting the links directly. Claude can work with web search results, and Anthropic’s advice is identical: review the cited sources, because the original pages contain context the model’s summary may have dropped or misread.
Think of pass 2 as a filter. It catches obvious fabrications fast and costs almost nothing. It cannot confirm anything by itself, because the model grading the answer shares the same weaknesses as the model that wrote it.
Pass 3: Open the primary source
This is the pass that actually kills hallucinations, and the one most people skip. A primary source is the original thing: the official documentation, the paper itself, the vendor’s live pricing page, the government page, the person’s own website. Not a blog post about the thing, and definitely not another chatbot’s summary of the thing.
For each claim on your list from pass 1, find the primary source and read the relevant lines yourself. Check three things: does the source exist at all, does it say what the chatbot claimed, and does it say it now, or is the page a stale snapshot? If a chatbot cites a paper, search the title. If the title does not resolve to a real publication, everything built on that citation collapses at once.
Pass 4: Verdict per claim
Mark every claim on your list with one of three outcomes. Verified: a primary source confirms it. Corrected: the source exists but the detail was wrong, so fix the detail and keep going. Dropped: no source backs it, so the claim leaves your work entirely. The value of this pass is that it converts a vague feeling of “the answer seemed fine” into a per-claim record you can defend later, including to yourself.
Built-in verification tools, compared
| Tool | What it does | What it does not do |
|---|---|---|
| ChatGPT web search and deep research | Looks up current information and attaches cited sources you can open; deep research compiles multi-source cited answers (per OpenAI’s help center) | Citations still need opening; a cited link is a lead, not proof |
| Gemini Double-check button | Color-codes statements against Google Search results: green for likely similar content, orange for likely different or none found (per Google’s help page) | Not available with extensions; skips code and markdown tables; green means “similar content exists,” not “verified” |
| Claude with web search | Answers with cited sources; Anthropic instructs users to review the originals because summaries can drop or misread context | The citation list is only as good as the pages it found; authoritative-looking quotes can still be ungrounded |
When you can relax, and when you cannot
Fact-checking everything a chatbot says would make the tool useless, so calibrate. The right question is: what happens if this specific claim is wrong?
Low stakes, skim only: brainstorming names for a project, getting a recipe, understanding a concept you will verify by doing. If an error costs you nothing, a careful read is enough.
Medium stakes, run passes 1 and 2: a blog outline, a tool shortlist, a learning plan. Errors here cost time, and the built-in checks catch the worst of it.
High stakes, run all four passes every time: anything medical, legal, financial, or safety-related; anything you will publish under your name; any number that will appear in a decision document. Anthropic’s own guidance says users should carefully scrutinize high-stakes advice, and that standard transfers directly. If you automate parts of your work with AI, which we cover in our beginner’s guide to AI automation, build the verification pass into the workflow itself rather than trusting the output at the end.
Common mistakes when fact-checking AI
Asking the same chatbot “are you sure?” A model that hallucinated once will often confirm its own invention with equal confidence. Verification has to leave the chat window.
Checking whether a link exists but not what it says. A real URL next to a claim does not mean the page supports the claim. Models attach plausible links to statements the linked page never makes. Open it and find the actual sentence.
Treating a green highlight or a citation list as a verdict. Google’s own documentation says green means similar content exists, and it explicitly notes the linked page may not be the one the model used. Built-in checks are leads.
Verifying only the claims that surprised you. Hallucinations are most dangerous when they align with what you expected, because nothing triggers your doubt. The boring claims in the middle of the list deserve the same pass 3 as the surprising ones.
AI hallucinations: quick answers
Which chatbot hallucinates the least?
This guide does not rank them, because the honest answer changes with every model release and depends heavily on the topic. OpenAI states that GPT-5 hallucinates significantly less, especially when reasoning, but that hallucinations still occur. Every major vendor’s own documentation tells users to verify important answers, which is the more durable fact to build habits on.
If I ask for sources, does that solve it?
It helps, and you should always ask, but a citation is only as good as your decision to open it. Fabricated references to non-existent sources are a documented failure mode, so an unopened citation proves nothing.
Are hallucinations a sign the model is broken?
No. They are a predictable output of systems trained to produce likely-sounding text and graded in ways that reward guessing. OpenAI’s research frames them as a fundamental challenge for all large language models, not a defect unique to one product.
Can I trust an answer that sounds detailed and specific?
Specificity is not evidence. The dissertation example from OpenAI’s research produced three different specific, confident, wrong titles. Detail affects how an answer feels, not whether it is true.
Sources
- OpenAI, “Why language models hallucinate” (September 2025) – definition of hallucinations as plausible but false statements, the evaluation-incentive explanation, and the dissertation and birthday examples: https://openai.com/index/why-language-models-hallucinate/
- OpenAI Help Center, “Does ChatGPT tell the truth?” – hallucination examples including fabricated citations, knowledge cutoffs, confidence versus reliability, and the advice to verify quotes, data, and references: https://help.openai.com/en/articles/8313428-why-does-chatgpt-generate-incorrect-or-biased-information
- Google Gemini Apps Help, “Double-check responses from Gemini Apps” – how the Double-check button works, the meaning of green and orange highlights, and its limits with extensions, code, and tables: https://support.google.com/gemini/answer/14143489
- Claude Help Center, “Claude is providing incorrect or misleading responses. What’s going on?” – Anthropic’s description of hallucination, the warning about authoritative-looking ungrounded quotes, and the instruction to review cited sources and scrutinize high-stakes advice: https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-on
