To fact-check AI-written content against primary sources, pull out every checkable claim, find the original page or document that states each one, read it yourself on the day you check, and mark each claim supported, contradicted or unverifiable. Then fix or cut whatever is not supported, and keep a record of what you read.
- Google's guidance on generative AI content calls it critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing (Google, Search Central, read 30 October 2026).
- In a 2023 audit of four generative search engines, on average 51.5% of generated sentences were fully supported by their citations (arXiv:2304.09848, read 30 October 2026).
- Penlid's methodology page says a claim is supported only when a source states that specific fact, and that a source on an adjacent statistic is recorded as unverifiable (Penlid, Methodology, read 30 October 2026).
- Google's generative AI guidance shows a last-updated date of 1 October 2026 and its people-first content guidance one of 5 October 2026 (Google, Search Central, read 30 October 2026).
Why does AI-written content need a fact-check?
AI-written content needs a fact-check because a language model predicts likely wording from its training data instead of retrieving facts, so a fluent sentence can still be wrong. Google's guidance on generative AI content makes the point in its section headed Focus on accuracy, quality, and relevance
, where it says generative AI outputs may contain inaccuracies, also known as hallucinations. It calls it critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing, and says the review also applies to metadata such as title elements, meta description elements, structured data and image alt text. It does not say who does the checking or give a method; the method below is ours.
NIST's Generative AI Profile calls the same failure confabulation: false or erroneous content stated with confidence. A 2025 paper says large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty, and that such hallucinations persist even in state-of-the-art systems. We give no error rate for current models. Our reasoning: a rate measured on one model and task would not tell you how often your draft is wrong, and you cannot tell which sentence is wrong without checking it.
Whether Google penalizes AI-generated content is a separate question, covered in does Google penalize AI-generated content. This guide is about whether the content is true.
What counts as a primary source?
In this guide, a primary source is the page, document or dataset where a claim starts, published by whoever made the claim or measured the result; anything that reports it later is secondary. NIST's profile describes high-integrity information as information that can be linked to the original source or sources with appropriate evidence. Our reading: a fact-check is that link back to the source, made one claim at a time.
| Kind of claim | Primary source to open | Secondary, so not enough alone |
|---|---|---|
| What a company says about its own policy or product | Its own documentation or help page | An article or AI answer summarizing it |
| A statistic or research finding | The paper, report or dataset that produced it | A news story quoting the figure |
| A price, plan or feature | The vendor's pricing or product page, read that day | A review or comparison site |
| A date, version or change | The page's own date line or the publisher's changelog | A search snippet |
| A quotation | The document that contains the words | A page that repeats the quote |
If no primary source turns up, the claim is unverifiable, so cut it or reword it to what you can support.
Why is a citation from the AI not a check?
A citation attached to an AI answer points to a source; it does not show that the source says what the answer claims. In a 2023 audit of four generative search engines, on average 51.5% of generated sentences were fully supported by their citations and 74.5% of citations supported their sentence, so about one citation in four did not support its sentence (100 − 74.5 = 25.5, our arithmetic). Those are 2023 results for Bing Chat, NeevaAI, Perplexity.ai and YouChat, not a rate for any current model or for your drafts.
NIST adds that outputs may include confabulated logic or citations that purport to justify or explain the system's answer, which may further mislead people into inappropriately trusting it. A subtler failure is the adjacent statistic, which in our reading is a real source with a real number about a different quantity. Penlid's methodology puts the rule this way: A source discussing an adjacent statistic is recorded as unverifiable, not supported.
So open every link the AI supplies and look for the sentence that states the claim.
How do you fact-check an AI draft, step by step?
Fact-check an AI draft in six steps: freeze the text, list its checkable claims, find each claim's primary source, read it and copy the supporting words, give every claim a verdict, then fix the draft and record the check.
- Freeze the draft. Save the version you are checking, with the date, so a later edit cannot leave you unsure which words were verified.
- List the checkable claims. These are statements a source could confirm or contradict. Our methodology page names
statistics, dates, prices, attributions, product facts
; we would add quotations and statements about what a policy says. Include the title, meta description, structured data and image alt text. - Find the primary source yourself. Search for where each claim started, using the table above. Treat any link the AI gave you as a lead: a link that fails to load, or loads a page that does not state the claim, is no support.
- Read the live page and copy the supporting words. Check the number, unit, date, scope and any qualifier such as 'may' or 'usually'. Recompute any figure the draft derives, such as a percentage; a derived figure is your claim, not the source's.
- Give each claim a verdict. Supported: a source states this specific fact. Contradicted: a source states something that conflicts with it. Unverifiable: you found no source that states it, including when you found only an adjacent statistic.
- Fix the draft, then record the check. Keep supported claims, correct contradicted ones to the source's wording, and cut or reword unverifiable ones. Our reasoning: one error suggests others from the same draft, so re-read the sentences around each contradicted claim. Who signs off next is covered in what a human should approve before AI-drafted content is published.
What should the record of a check contain?
A record of a check should hold, for each claim, the claim as the draft words it, the source URL, the exact words you found there, the date you read the source and the verdict. Our methodology page describes a similar discipline: checkable claims are extracted from every draft and verified individually against live sources, each with a stored record of verdict, confidence, source URL and the exact supporting quote
. A supported verdict requires a source stating that specific fact
, and A draft with any contradicted claim is held for human review regardless of its overall score.
These fields and rules are Penlid's own; we do not claim they are a standard.
The date matters because pages change: Google's generative AI guidance shows a last-updated date of 1 October 2026 and its people-first content guidance one of 5 October 2026. Our reasoning: a model that writes from training data may reflect an older version of a page, so read the live one. NIST's profile, a companion to a framework intended for voluntary use, suggests deploying and documenting fact-checking techniques to verify the accuracy and veracity of what generative AI systems produce. Our reasoning: the same habit of writing down what you did, and when, also helps if you want to measure a change later; see what a content operations ledger should record.
Example (illustrative numbers)
These counts are invented to show the arithmetic; they are not data from any draft or any site. Suppose an AI draft runs 1,380 words and you list its checkable claims: 14 statistics, 9 dates, 12 attributions and 11 product or policy facts. That is 14 + 9 + 12 + 11 = 46 claims, about one every 30 words (1,380 ÷ 46 = 30).
You check each claim against its primary source and record a verdict.
| Verdict | Claims | Share of 46 | What you do |
|---|---|---|---|
| Supported | 31 | 31 ÷ 46 = 67.4% | Keep it, with its record |
| Unverifiable | 9 | 9 ÷ 46 = 19.6% | Cut it, or reword it to what a source states |
| Contradicted | 6 | 6 ÷ 46 = 13.0% | Replace it with what the source states |
| Total | 31 + 9 + 6 = 46 | 67.4 + 19.6 + 13.0 = 100.0% | Hold the draft |
About two thirds of the claims are supported, but 6 are contradicted. Under a rule like Penlid's, which holds a draft with any contradicted claim for human review, this draft is held; the supported share does not offset the contradicted claims.
Now suppose you correct all 6 contradicted claims, cut 5 of the 9 unverifiable claims and reword the other 4 to what their sources state. The draft then has 46 − 5 = 41 claims. Only 6 + 4 = 10 of them are new text, so the re-check starts with those 10 and the sentences around them, and you record the date you read each source.
What this cannot tell you
- A supported claim is only as current as its source. The verdict holds for the page as you read it on the date you recorded, and a primary source can itself be wrong or out of date.
- Accurate claims do not make an accurate argument. A draft can be right sentence by sentence and still mislead through what it leaves out (our reasoning), so this check does not replace an editor's read.
- Not every statement has a source. Opinions, advice and forecasts have no primary source to find (our reasoning); label them as such or cut them.
- It does not tell you how Google will treat the page. Google's guidance asks for fact-checking for accuracy and trustworthiness, and its people-first guidance asks whether content has easily-verified factual errors; neither page says that checking a draft raises rankings or protects a site from spam enforcement.
- The 2023 citation figures are not a current error rate. They describe four engines in 2023, so they do not tell you how often a current model, or your own drafts, get a claim wrong; only checking the drafts does (our reasoning).
Questions people also ask
How do you fact-check information from an AI?
Treat each factual statement in the answer as a lead, not a finding. Find out who first made the claim, open their page and look for the sentence that states it. If all you find is a nearby figure about something else, or nothing, count the claim as unverifiable and cut it or reword it.
Can an AI fact checker check AI content for you?
It can do part of the work, but the check itself is reading the source. Anthropic's documentation describes asking the model, after it drafts, to find a supporting quote for each claim and to retract any claim it cannot find a quote for. The same page says these techniques significantly reduce hallucinations but do not eliminate them entirely, and advises always validating critical information, especially for high-stakes decisions. Our reasoning: a tool's answer is one more set of claims to verify.
Can a fact-checking website verify AI-written content?
Only as a lead, in our reasoning. A fact-checking site reports on claims already in circulation, so it may not cover the statistic, date or price in your draft, and it is a secondary source. Use it to find where the claim began, then check the claim against that page.
How do you check whether content is AI-generated?
That is a different question from whether the content is accurate, and this guide does not cover detection. A sentence a person wrote and one a model wrote need the same check against a primary source. If you need to know how your own drafts were made, record it when you create them; whether to tell readers is a separate choice, covered in how to disclose AI assistance on a page.
Sources
- Google Search's Guidance on Generative AI Content on Your Website, Google, Search Central, read 30 October 2026.
Focus on accuracy, quality, and relevance
- Creating helpful, reliable, people-first content, Google, Search Central, read 30 October 2026.
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST, read 30 October 2026.
- Why Language Models Hallucinate, arXiv, read 30 October 2026.
- Evaluating Verifiability in Generative Search Engines, arXiv, read 30 October 2026.
- Reduce hallucinations, Anthropic, read 30 October 2026.
- Methodology: how we measure, and what we can't, Penlid, read 30 October 2026.
statistics, dates, prices, attributions, product facts
verdict, confidence, source URL and the exact supporting quote
requires a source stating that specific fact. A source discussing an adjacent statistic is recorded as unverifiable, not supported. A draft with any contradicted claim is held for human review regardless of its overall score.