You have a draft. An AI wrote most of it, and it reads well: confident, fluent, structured. That confidence is exactly the problem. A language model sounds the same whether a claim is solid or invented, so the fix is not reading more carefully. It is checking things in a fixed order every time, the same way a newsroom fact-checker works a piece before it runs, rather than trusting your gut about which sentences "feel" risky.
Here is that order: pull every checkable claim out of the draft first, sort them by how much damage a wrong one would do, verify the highest-risk ones against a primary source, then verify names, dates, and numbers last, since those are the easiest to get wrong quietly. Do this before you touch tone or style. A fact-check on a sentence you later cut is wasted time; a rewrite of a sentence built on a false claim is wasted twice.
Why "read it carefully" does not work here
Most editing instinct is tuned to catch awkward phrasing, not confident falsehoods. AI-written text does not signal uncertainty the way a human writer under-informed on a topic usually does, with hedges, vague language, or a request for more time. A model will state a fabricated statistic in the exact same register as a well-supported one. That is the whole problem in one sentence, and it is why this needs a checklist instead of a vibe check.
Step 1: Pull out every checkable claim
Go through the draft once, changing nothing, and mark every sentence that makes a claim someone else could independently confirm or deny:
- Statistics and percentages
- Named studies, reports, or sources
- Direct or paraphrased quotes attributed to a person
- Dates and version numbers
- Company names, job titles, and "as of" claims
- Anything phrased as a fact rather than an opinion
If a sentence has none of these, it is probably framing or transition text and does not need a fact-check pass. This step alone usually cuts what looked like a 1,500-word verification job down to a dozen sentences.
Step 2: Rank the claims by how much a wrong one would cost
Not every unverified claim carries the same risk. A wrong claim about a competitor's pricing or a health, legal, or financial recommendation can do real damage. A wrong claim about a general industry trend, stated loosely, usually cannot. Spend your limited verification time on the claims that would embarrass you or mislead a reader if they turned out to be false, not on every single one equally.
Step 3: Verify the source actually exists
This is the step people skip, and it is the one that catches the most damage. Before you check whether a cited study says what the AI claims, check whether the study exists at all.
- Search the exact title in quotation marks. If nothing comes up, that is already a warning sign.
- Check for a DOI or a direct link, and follow it. A citation with no way to locate the source is not a citation.
- Confirm the publication venue is real and that the volume, issue, or date lines up with that source's actual publishing history.
A model can also cite a real source and then attribute a claim to it that the source never makes. Existing is step one. Saying what the AI claims it says is step two, and both have to pass.
Step 4: Verify the statistic itself, not just its source
Once you know a source is real, check what it actually reports. Look at the original number, not a summary of it, and note the year, the sample, and any caveat the original source attaches. A statistic that was true in 2019 and is quoted as current, or one that applies to a narrow group and is presented as universal, is a fabrication of context even when the number itself is real.
Step 5: Check names, titles, and dates last
These errors are the easiest to introduce and the hardest to notice, because they look plausible. A person's job title may have changed since the AI's training data was current. A company may have been acquired or renamed. Confirm current titles and affiliations directly on the person's or company's own page, not on a secondhand summary, and treat anything time-sensitive as needing a fresh check regardless of how confident the sentence sounds.
What this looks like on a real draft
Say an AI-written paragraph claims: "A 2024 study found that 73% of small businesses using AI tools reported time savings, according to a report from the National Small Business Association." Before you touch the wording:
- Search for that exact report title. Does it exist under that name, from that organization?
- If it exists, does it actually report 73%, or is that number closer to a different figure in the source, or from a different survey entirely?
- Is 2024 the correct year, or has the AI attached a plausible-sounding year to an undated figure?
If any one of these fails, the sentence does not get softened. It gets cut or rewritten from a source you have actually confirmed.
What no source can replace
Some of what a fact-check catches is not a false statistic. It is a claim with no error in it that still cannot be confirmed anywhere, meaning it should be treated as unverified rather than accepted because it sounds reasonable. "Unconfirmed" and "false" are different findings, and a careful process keeps that distinction rather than collapsing everything into "checked" or "not checked."
Going deeper on each step
This is the order of operations. Each step has more detail behind it than fits in one pillar guide:
- To understand why this checking is necessary at all, see what an AI hallucination actually is.
- Step 3 above, checking whether a source exists, is covered in full in why AI invents sources and citations and how to check whether a study an AI cited actually exists.
- Step 4, checking the statistic itself once the source is confirmed real, is covered in verifying a statistic before you publish it.
- Step 5, checking names, titles and dates, is covered in full in verifying names, titles and companies in an AI draft and why AI gets dates and recent events wrong.
- A quote is its own specific check: see checking a quote an AI attributed to someone.
- If the draft includes any calculation, see checking AI output for numbers and math errors, and for tracing a citation back to its true origin, primary source vs secondary source.
Sources, links and prompting
- Whether asking the model for citations in the first place actually helps: asking AI for its sources, what works and what doesn't.
- Why those citations keep arriving as dead URLs no matter how the prompt is worded: why "cite your sources" prompts still produce fake links.
- The batch procedure for testing a draft's links quickly: how to check whether a URL an AI gave you is real.
- Prices deserve their own check, because they go stale faster than anything else: verifying what AI says about pricing and plans.
Doing this when you cannot check everything
- Deciding which claims get your limited time: which claims to verify first.
- What this realistically costs, and when writing it yourself is faster: how long fact-checking an AI draft actually takes.
- Whether a second model helps, and the one direction in which it does: cross-checking one AI against another.
- Turning all of this into something repeatable you can hand to someone else: building a verification checklist for your own site.
Claims that need their own treatment
- Features a model says a product has, and the removed-feature trap: verifying AI claims about a product's features.
- Rules and regulations, where jurisdiction is the check nobody runs: verifying what AI says about laws and regulations.
- Where to stop entirely, because the reader could be harmed: health and money claims, and when not to publish at all.
- Summaries of long documents, where the failure is omission and cannot be fact-checked: verifying AI summaries of long documents.
When the check does not come back clean
- Cut, hedge, or attribute, and how to tell which: what to do when you cannot verify a claim.
- When editing costs more than starting over: when to throw the whole AI draft away.
- Keeping a record of what you checked, and what you knowingly did not: how to log what you verified.
- The honest process for fixing something already published: writing a corrections policy for AI-assisted content.
Once it is accurate, it still has to be readable
Verification is the first stage, not the whole job. What comes after it, turning a checked draft into something worth publishing, starts at turning an AI first draft into writing worth publishing. The two stages run in that order for a reason: editing prose before knowing which claims survive is work spent on sentences that may not last.
This site's own Editorial Policy commits to the same standard applied here: if something has not been checked, it gets said plainly rather than implied otherwise.
A short verification checklist
- Every statistic traced to a source you can open and read
- Every named source confirmed to exist under its stated name
- Every quote checked against something the person or document actually said
- Every date and title checked against a current, primary source
- Anything you could not confirm, marked as unconfirmed and either cut or clearly flagged
FAQ
Does fact-checking mean an AI draft is unreliable by default?
It means an AI draft is unverified by default, which is a different thing. Treat it the way you would treat a first draft from a junior writer who has not yet been told which claims need a source. Some of it will hold up. None of it should be assumed to.
How long should fact-checking actually take?
It depends entirely on how many checkable claims the draft makes and how high the stakes are. A short opinion piece with two statistics takes minutes. A data-heavy explainer with a dozen cited studies takes considerably longer, and that time is the cost of publishing it responsibly.
What if I cannot find a source at all?
Treat the claim as unconfirmed, not as true. Either cut the sentence, rewrite it as a stated estimate with no false precision, or hold the piece until you can verify it. Publishing an unconfirmed claim as fact is the exact failure this process exists to prevent.
Can I ask the AI itself to check its own sources?
You can, but treat the answer with the same skepticism as the original claim. A model that invented a citation once can restate the same invented citation confidently when asked to double-check itself. Independent verification, outside the tool that made the claim, is the only step that actually counts.
Is this different from normal editing?
Yes. Editing improves how a piece reads. Fact-checking confirms whether what it says is true. A well-edited, confidently written paragraph can still be built on a completely invented source, which is exactly why this is a separate pass, done before the polish.