To fact-check AI-assisted writing before publishing, run a fixed six-step workflow. Identify each factual claim, and separate supported claims from unsupported ones. Verify sources and citations, check context and freshness, and give high-risk claims extra scrutiny. Then decide for each one: publish, revise, or remove. The order matters: skipping the separation step is why most AI fact-checking fails.
This is the procedure our other posts on the topic point to. It assumes the slow step has moved from writing to checking. We make that case in why AI made writing faster but publishing slower, and this page gives you the process.
Six steps, in a fixed order, from claims to a decision
Six steps, always in this order, and each one produces what the next step uses. The workflow is the same for a newsletter and for a whitepaper. Only the triage weighting, which claims you check hardest, changes.
Identify claims → separate supported from unsupported → verify sources → check context & freshness → scrutinize high-risk claims → decide: publish, revise, or remove. If you do only one thing differently from today, make it step two. Most people jump from here is the draft straight to spot-check a few things. That is not a workflow.
Editing is not fact-checking
No, and mixing them up is the root failure. Editing improves how the piece reads: flow, clarity, structure, voice. Fact-checking confirms the piece is true: claims, citations, attributions, currency. A draft can be well edited and entirely wrong. AI made this worse: a model's output is fluent by default, so it clears the editing bar with ease. That fluency tells you nothing about whether it clears the truth bar. If your pre-publish process is a read-through, you are editing. Fact-checking is a separate pass with a different question.
Editing gets faster as you learn a topic. Fact-checking does not, because a confident fabrication looks exactly like a confident truth until you check the source. Experience makes you a faster reader. It does not make you better at spotting errors in your own draft.
The six-step fact-check workflow
Run the steps top to bottom, and do not start verifying claims until you have separated them. Otherwise you spend your effort on the easy ones and run out of attention before the risky ones.
Step 1: Identify each factual claim
Go through the draft and mark each sentence that asserts something checkable: a statistic, a date, an attribution, a causal claim, a definition, a superlative. Opinions and framing are not claims. If a sentence could be true or false in the world, it is a claim, and it goes on the list.
Step 2: Separate supported from unsupported
For each claim, ask one question: does the draft carry a source for this, or did the model just assert it? Split the list in two. Unsupported assertions are the highest-risk group, because nothing flagged them. In a read-through, they look identical to supported ones.
Step 3: Verify sources and citations
For each cited claim, confirm three things: the source exists, its details (authors, title, year) match, and the source supports the claim. The full method covers the 2026 failure modes that a working link no longer catches. It is in how to check if AI citations are fake.
Step 4: Check context, caveats, and freshness
A claim can be sourced and still misleading. It can lose its caveat, be stretched past its sample, or be out of date. Confirm the claim still holds today. Then confirm the draft kept the qualifier that made the source's version defensible.
Step 5: Scrutinize high-risk claims differently
Some claims would embarrass you, cost a client, trigger a correction, or carry legal or medical weight. Those get a second, independent check, which is different from the same check done twice. Risk decides how much verification a claim earns. Its order in the document does not.
Step 6: Decide to publish, revise, or remove
Each claim ends in one of three states: verified and kept, weakened to exactly what the evidence supports, or removed. There is no probably fine. Record the decision, and the verification becomes something you can show as well as something you did.
Each type of claim fails in its own way
Different claim types fail in different ways, so match the check to the type. That way you avoid checking everything the same shallow way.
| Statistic / number | Trace to the primary source, not the article that quoted it. Confirm the figure, the population, and the year. A number with no methodology nearby is a flag. |
| Citation / reference | Existence. A character-level metadata match. And a source that supports the specific sentence. Resolving the link is not enough in 2026. |
| Attributed quote | Find the quote in the original source verbatim. Confirm the speaker, the wording, and that the surrounding context does not reverse its meaning. |
| Causal claim | Confirm the source claims causation, not correlation. Confirm the draft did not upgrade a hedged finding into a firm one. |
| Definition / technical fact | Check against current primary docs. Version-specific and API facts age fast. Almost-right is wrong here. |
| Superlative / first / only | The hardest to defend, and the most screenshotted. Ask for an explicit, current, citable basis. If there is none, cut the superlative. |
Three mistakes behind almost every published error
Three mistakes account for almost every published error from a team that thought it had a process.
Letting the model verify itself
Ask the model that wrote the draft whether the draft is true, and you get a closed loop: its check shares the blind spots of its draft. A different vendor's model is a real independent check, and the same one is not. This is the core of whether one AI can reliably fact-check another.
Checking style but not facts
A polished read-through feels like diligence and verifies nothing. Fluency is the one thing an AI draft is sure to have. That is why it pairs so well with sycophantic agreement: both reward you for not looking closer.
Confirming a source exists but not that it supports the claim
This is the most common real-world failure. The reference is genuine and the metadata is right, but the paper does not say what the draft claims. Existence is step three of four, not the finish line.
The steps stay the same, and the triage changes by format
The six steps do not change, but the triage in step five does.
Blog posts
Lead with statistics and the studies behind them, since rivals fact-check those and readers screenshot them. Check superlatives second.
Newsletters
Lead with anything a subscriber could reply to correct. That means attributed quotes, named figures, and claims about your readers' own field. A miss here costs you a public correction to the whole list.
B2B case studies and whitepapers
Lead with customer numbers, market sizing, and competitive claims. Keep the step-six log, so editors can sign off without re-checking from scratch.
Technical explainers
Lead with version-specific behavior, and test benchmarks against current primary docs. A model's memory of an API is not a source.
Running the checks across several models makes it fast enough to keep
Steps three to five grow with the number of claims. They were the original 30-to-60-minute pre-publish cost. Doing them by hand at AI-draft volume does not scale. So most teams quietly drop them, then find the gap in public. The version that scales keeps all six steps and runs the independent checks across several models from different vendors at once. Their disagreement then points you to the claims that need a human. That argument is made in full in why AI made writing faster but publishing slower.
Notice that steps two through five are the slow part. TrueStandard runs them across four frontier models from different vendors, in parallel. In about 60 seconds, you get the separated claim list, the citations they disagree on, and a log you can replay. It does not make the publish-or-remove call for you, so step six stays yours. It removes the 45 minutes between your draft and that decision.
Frequently Asked Questions
What is the difference between editing and fact-checking AI content?
Editing improves how it reads. Fact-checking confirms it is true. A draft can pass editing perfectly and still be full of fabricated claims, because AI output is fluent by default. They are separate passes with different questions, and a read-through only does the first.
Where do most AI fact-check workflows go wrong?
At step two. People jump from the finished draft straight to spot-checking a few things, and never separate supported claims from unsupported ones. Nothing flagged the unsupported assertions, so they are the highest-risk ones, and they get skipped.
How long should fact-checking an AI draft take?
By hand, the verification steps take the better part of an hour for a long piece. That is the cost AI was supposed to remove. Run the independent checks across several models in parallel, and it drops to about a minute. Then add your judgment on the flagged claims.
Can a single tool fact-check my AI writing for me?
A single-model tool inherits single-model blind spots, and it also cannot make the publish-or-remove call. The reliable pattern is several independent models showing you where they disagree, for you to rule on. The tool does triage, and step six stays human.
Is checking with one other model enough?
It is better than self-review, because the errors are less correlated. But one alternate model is still a single point of failure. Several independent models surfacing where they disagree is the version that scales. It gives you a signal you can act on.
Keep reading
Why AI Hallucinations Are Structural
DELEGATE 52, GPT-5.5, and a Purdue impossibility proof. Three April 2026 results that move 'hallucinations are structural' from take to documented fact.
Why AI Made Writing Faster but Publishing Slower
Drafting got faster and verification did not. The work moved to the step right before your name goes on it.
What Are AI Hallucinations?
An AI hallucination is a false claim a model states with confidence. It might be a research paper that does not exist, a made-up statistic, or a wrong fact about a real person.
AI Fact-Checker False Positive Rate: We Measured Ours
We measured how often our checker flags a true claim, on 30 labelled claims. Recall is easy to advertise, since a tool that flags everything scores 100 percent on it. How often a checker flags something true decides whether you can use it. The more useful result was what our test could not tell us.
TrueStandard vs Omniscient AI
Both run several models, and both show where those models agree. The difference is the moment each one is built for. Omniscient checks what you are reading. TrueStandard checks what you are about to publish.
What we measured on this model
Each release page carries our own fabrication data for one model version, measured against the model it replaced in the same vendor's line.
Run the Workflow in 60 Seconds
TrueStandard runs the independent-verification steps across four frontier models from different vendors, in parallel. It hands you the flagged claims and a log you can show.
TrueStandard is for writers who publish AI-assisted drafts. Paid plans check every claim and citation across four AI models before you hit publish.
Free: one check a day, no signup. Pro is $20 a month for 250 drafts.
Start Verifying →