How to Fact Check AI Generated Content: Map Claims to Dated Primary Sources
Make this article actionable
Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.
Quick answer
To fact check AI generated content, stop reading the report as prose and start reading it as a list of claims. Break the draft into individual assertions, tag each one by type (number, date, quote, causal claim, prediction), then try to attach a dated primary source to each. Anything you cannot attach gets marked unresolved rather than quietly deleted or left standing. The goal is not a "clean" report — it is a report where every sentence's status is visible to the reader.
This works because AI research drafts fail in predictable ways: they blend real facts with plausible-sounding filler, they flatten dates, and they state contested findings as settled. A claim-by-claim pass catches all three.
Turn the useful parts into next steps
Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.
Step 1: Convert the draft into a claim list
Read the report once without editing. Then go back and number every sentence that asserts something about the world. Ignore transitions, framing, and your own commentary.
A claim is anything a reader could disagree with on factual grounds:
- Quantities — "the market grew 40% between 2021 and 2024"
- Dates and sequences — "the standard was revised before the pilot launched"
- Attributions — "researchers at X found that Y"
- Definitions — "this term refers only to Z"
- Causal statements — "the change caused the decline"
- Predictions — "adoption will double by 2027"
Split compound sentences. "The 2019 study found a 12% effect and was later replicated" is two claims with two different verification paths.
Keep the split mechanical. You are not judging accuracy yet — you are building the queue.
Step 2: Classify each claim by how it can be verified
Not every claim needs the same evidence. Sorting first saves hours.
| Claim type | What counts as verification | Typical failure mode |
|---|---|---|
Statistic | The dataset, report, or filing the number came from, with its publication date | Number is real but attached to the wrong year or population |
Date / sequence | A dated document, release note, or archive record | Two events merged into one timeline |
Named attribution | The actual paper, post, or transcript, plus the person's role at the time | Real person, real topic, invented finding |
Definition | A standards body, textbook, or the source that coined the term | Definition drifts to a broader or narrower meaning |
Causal claim | Study design that supports causation, not just correlation | Correlation restated as cause |
Prediction | A named forecaster and a stated horizon | Prediction presented as established fact |
The right-hand column is where most AI drafts break. Treat it as your review checklist, not as a description of any specific model.
Step 3: Attach dated primary sources
For each claim, work outward from the most authoritative source you can actually open:
- Primary — the original dataset, filing, standard, transcript, or paper.
- Secondary — reporting or review that cites the primary source.
- Tertiary — summaries, encyclopedias, aggregator pages.
Prefer primary. When you can only reach secondary, record that limitation in the table instead of hiding it.
Date everything. A source's publication date is part of the evidence, not metadata. A 2018 figure cited in a 2026 report is not wrong — it is just old, and the reader deserves to know. Record both the source date and the date you checked it.
Record the exact locator. Page number, section heading, table name, timestamp. "It's in the report somewhere" is not a citation, and you will not find it again in three weeks.
If you are running this pass inside a research workspace, the useful pattern is to keep the claim list and the source list side by side so each claim carries its own evidence trail — that is the workflow the AI research agent is built around, and you can paste a draft in and ask it to return a claim table rather than a rewritten report.
Step 4: Resolve contradictions
Contradictions are the highest-value finding in the whole pass. Do not average them away.
When two sources disagree, work through these questions in order:
- Are they measuring the same thing? Different definitions, populations, or geographies produce different numbers that are both correct.
- Are the dates different? A revision or a later dataset can legitimately supersede an earlier figure. Say which one you are using and why.
- Is one source downstream of the other? If both trace back to the same original, you have one source, not two.
- Is the disagreement methodological? Survey vs. administrative data, self-reported vs. observed. Name the method.
- Is it genuinely unresolved? Then say so, and describe what evidence would settle it.
Write the resolution as a sentence the reader can audit: "Source A reports X for 2022 using survey data; Source B reports Y for 2023 using administrative records. We use B for the 2023 figure and note the definitional gap."
Step 5: Mark unresolved statements explicitly
Every claim ends in one of four states. Use the same labels everywhere so a reader can scan for them.
- Verified — primary source located, dated, and consistent with the claim as written.
- Partially verified — the direction is right but a number, date, or scope detail does not match.
- Unresolved — no source found, or sources conflict without a resolution.
- Removed — the claim could not be supported and was cut.
Unresolved is a legitimate output. A report that says "we could not confirm this figure" is more useful than one that asserts it. What you must not do is leave an unsupported claim in the body with no marker, because the reader will read it as verified.
Reusable claim / source / review table
Copy this into your draft and fill one row per claim. The Status column is the deliverable.
| # | Claim (as written) | Type | Source + locator | Source date | Checked on | Status | Note |
|---|---|---|---|---|---|---|---|
| 1 | "Adoption reached 40% in 2024" | Statistic | Annual survey, Table 3 | 2025-03-11 | 2026-09-14 | Partially verified | Survey says 38%, and only for firms over 50 staff |
| 2 | "The standard was revised before the pilot" | Sequence | Release note v2.1 | 2021-07-02 | 2026-09-14 | Verified | Pilot start date confirmed in same note |
| 3 | "Researchers at X found Y" | Attribution | — | — | 2026-09-14 | Unresolved | Could not locate the paper; do not cite |
| 4 | "The change caused the decline" | Causal | — | — | 2026-09-14 | Removed | Only correlational evidence available |Two habits make this table worth keeping. First, fill Checked on every time — a verification is only as good as its date. Second, never let a row sit in Unresolved without a note explaining what you tried.
Example input
A prompt you can adapt. This is an example of how to phrase the request, not a report of a measured result:
Here is a draft research summary. Do not rewrite it.
1. Extract every factual claim as a numbered list. Split compound sentences.
2. For each claim, label the type: statistic, date/sequence, attribution,
definition, causal, or prediction.
3. For each claim, tell me what kind of primary source would verify it and
what the source would need to state.
4. Flag any claim where the draft gives a number without a year, or a
finding without a named source.
5. Return a table with columns: claim, type, required evidence, red flag.
6. Do not invent sources. If you cannot point to a real one, write
"no source supplied".
Draft:
[paste draft here]The instruction "do not invent sources" matters more than it looks. Without it, a model will often supply a confident-looking citation that you then have to verify anyway — which is slower than starting from a blank cell.
Output review checklist
Run this before the report leaves your hands.
- Every number has a year and a defined population or unit.
- Every named study, person, or organization has a locator you can reopen.
- No source is cited that you have not personally opened.
- Dates are attached to sources, not just to events.
- Contradictions are described, not smoothed over.
- Every claim carries one of the four status labels.
-
Unresolvedclaims are visible in the body, not buried in a footnote. - Causal language is downgraded wherever the evidence is correlational.
- Predictions name a forecaster and a horizon.
- The
Checked ondate is recorded for every verified row.
Common mistakes
Verifying the vibe instead of the claim. A paragraph can feel accurate while containing one wrong number. Work row by row.
Treating a secondary source as primary. If the article you found cites a study, go to the study. If you cannot, mark the claim partially verified and say why.
Accepting a citation that exists but does not say that. A real paper can be real and still not support the sentence attached to it. Open it and read the relevant section.
Letting the model supply the sources. Ask for the claim list and the evidence requirements; do the source lookup yourself, or verify every source it returns.
Deleting unresolved claims silently. Removal is fine. Silent removal hides the gap from the next person who reads the draft.
Flattening dates. "Recent research shows" is not a date. Neither is a citation year that differs from the year the data was collected.
Rewriting before verifying. Editing for tone first makes claims harder to isolate. Extract, verify, then edit.
FAQ
How long should this take? It scales with claim count, not word count. A short draft with twenty numbers takes longer than a long one with five. Budget the pass by rows in the table.
What if I cannot find any source for a claim?
Mark it Unresolved with a note on what you searched, or remove it. Do not soften the wording to make it defensible — "many observers believe" is still an unsupported claim.
Should I cite the AI tool that produced the draft? Describe how the draft was produced if your audience needs that context, but the citation burden sits with the claims. A tool is not a source.
Can I trust a source the model gave me? Only after you open it. Treat every supplied citation as a lead, not as evidence.
What about claims that are true but hard to source?
Common knowledge and definitional statements can be marked verified against a standard reference. Anything contested needs a real source or an Unresolved label.
Do I need to re-check sources later?
Yes, if the report stays live. Sources get revised and datasets get restated. The Checked on column exists so you know what is stale. If you are producing these reports repeatedly, keeping the claim table as a living artifact — rather than rebuilding it per draft — is the main reason to run the pass inside a research workspace such as the AI research agent. Credit costs for any assisted step are shown in the model selector and on the pricing page at the time you run it.
Does a verified claim mean the conclusion is right? No. Verification checks individual statements. Whether the conclusion follows from them is a separate argument you still have to make.