Fact-checking with Claude: build a traceable report
Checking a text with Claude: Fact-Checker's six verdicts, and why AI-generated content cannot serve as a source.
By Educasium
Check a text with Fact-checker →
You are about to publish an article, a LinkedIn post or a sales deck drafted with the help of AI, and a question nags at you: what if a figure or a claim in the text is simply wrong? Going back to the tool that produced the text and asking "is this actually true?" answers nothing — you would be asking the source of the doubt to judge its own reliability.
Fact-Checker structures a different method. It pulls every verifiable claim out of a text, checks each one against identified primary sources, then returns a precise verdict for each — true, mostly true, mixed, unverifiable, false or outdated — rather than a general approval of the text as a whole. The value of that report depends entirely on the sources actually consulted, never on how confidently the verdict is stated.
This article details those six verdicts, explains why content generated by an AI can never serve as evidence for checking other AI-generated content, and shows how to build a report that someone else could redo from scratch.
Summary
- Why an isolated verdict is not enough
- Fact-Checker's six verdicts
- Why AI-generated content cannot serve as a source
- Checking a text step by step
- What Fact-Checker does not do
- Training to build verifiable content
- Frequently asked questions
Why an isolated verdict is not enough
A verification report should let someone else repeat exactly the same check and reach the same conclusion; an isolated verdict, however confidently stated, does not allow that unless it names the source consulted and the exact passage that justifies it. Fact-Checker therefore organizes every claim around three fixed elements — the extracted statement, the evidence located and the resulting verdict — so the report stays checkable afterward, not merely convincing while you read it.
The full text and its reference date bound the review
Supplying only the excerpt that worries you, without the full text or its reference date, makes it impossible to tell information that has since become true apart from information that was false from the start. A figure that was accurate when written can become outdated six months later without any error having been made at the time; that is exactly the difference the "outdated" verdict exists to express, provided the text's reference date is known from the outset.
Checking a summary and checking an external claim are different operations
If the text under review summarizes meeting notes or a transcript, two different checks are possible: confirming that the summary stays faithful to the source document, or confirming that an external claim quoted inside that summary is itself accurate. Both operations need different evidence — the original document in the first case, an external primary source in the second — and conflating them produces a report that looks complete without being so.
Fact-Checker's six verdicts
Fact-Checker returns a precise verdict for every extracted claim, chosen from six fixed categories rather than a simple true/false binary that would hide the in-between cases — which are, in fact, the most common ones in a real text.
| Verdict | Meaning | Recommended action |
|---|---|---|
| True | The claim matches exactly what the consulted primary source shows | Keep the claim as worded |
| Mostly true | The claim is broadly accurate, with a minor nuance or rounding | Adjust the wording to include the missing nuance |
| Mixed | Part of the claim is confirmed, another part is contradicted or absent from the source | Split the two parts and handle each separately |
| Unverifiable | No accessible source can confirm or rule out the claim | Flag the claim as unverified, without concluding on its truthfulness |
| False | The primary source directly contradicts the claim | Correct or remove the claim before publication |
| Outdated | The claim was accurate at an earlier date but no longer is | Update the figure or date, or remove the claim if no recent data exists |
Why "unverifiable" never equals "false"
The absence of accessible evidence proves nothing about a claim's truthfulness: a source can be paywalled, taken down, or simply missing from a first search pass, without the claim it might have confirmed being false as a result. Fact-Checker therefore keeps "unverifiable" as a category distinct from "false," precisely so that a lack of evidence never turns into evidence of error — a frequent and yet deeply misleading confusion in a verification report.
Why AI-generated content cannot serve as a source
Fact-Checker explicitly treats any AI-generated content as an unusable source for checking another claim that is itself potentially AI-generated — a rule that describes a real limit of the tool, not a minor technical footnote. A chatbot that confirms a figure is only rephrasing a statistical likelihood about what sounds like a plausible answer; it does not consult reality to confirm it, and presenting that as evidence would mean validating a doubtful claim with a tool that shares the exact same structural weakness.
What the pack actually checks, and what it does not claim to do
Fact-Checker does not "detect" falsehood autonomously: it checks a claim against sources you supply or against primary sources it can name and consult — a piece of legislation, a published study, an official release, statistics from an identified body. Presenting the pack as a general falsehood detector would overstate what it actually does; describing it as a tool that matches claims against named evidence is exactly what it does, no more and no less.
Without web access, the check stays document-based
Without web search enabled, Fact-Checker can only check a text against the documents you explicitly attach to the conversation — it then cannot confirm or rule out a claim about recent events. In that setup, ask for a document-based comparison limited to the supplied materials rather than a check of current events, which the pack cannot materially perform without outside access.
Checking a text step by step
Five steps structure a check that stays verifiable afterward by someone else.
Step 1: Supply the full text, its reference date and any already-known sources. Attach the whole document, not only the excerpt that worries you, along with its drafting date and any source already identified for the claims it contains.
Step 2: Prioritize the claims that would change a decision. A long text often contains more verifiable claims than there is time to check all of them; focus the effort on the ones that would actually influence a decision if they turned out to be wrong.
Step 3: Request evidence, source and date for every prioritized claim. Require the exact passage that justifies each verdict, along with the source's URL or reference and its publication date where available — a verdict missing these three elements cannot be checked afterward.
Step 4: Explicitly separate false, unverifiable and unchecked. These three categories describe different situations and must appear separately in the final report; merging them erases a distinction the report exists precisely to preserve.
Step 5: Keep the list of unchecked claims if time ran short. If not every claim could be verified, that list must accompany the final report rather than quietly disappear — an incomplete but honest report remains preferable to one that looks complete without being so.
What Fact-Checker does not do
Fact-Checker never confirms a claim based on other AI-generated content, including its own earlier answers in the same conversation — this rule directly protects the reliability of the final report, it is not excessive caution. It also never turns missing evidence into a verdict of falsehood: an inaccessible page or an untraceable source leaves the check incomplete, never concluded against the claim.
The pack also does not replace a legal service when the claim at stake carries contractual consequences, nor a statistical service when the question concerns the methodology of a cited study. In the feedback we receive from professionals who regularly publish AI-assisted content, the most common difficulty is not finding a source — it is accepting that a claim stays "unverifiable" rather than forcing a decisive verdict that would feel more reassuring without being more accurate.
Training to build verifiable content
Knowing how to tell a primary source apart from a rewording, recognizing when a claim stays unverifiable rather than false, and structuring a report someone else could redo — these are skills built through practice before they become reflexes. Educasium's Master Claude training devotes modules to content verification for marketing teams and self-employed professionals who regularly publish AI-assisted content. For a self-employed professional under the relevant scheme, the FIFPL covers part of the cost of a Qualiopi-certified training under the 2026 criteria set at €300 per day and €900 per year, with e-learning capped at 50% of the daily rate.
Frequently asked questions
Can Fact-Checker automatically detect a lie in a text?
No, and presenting the pack that way would overstate what it actually does. Fact-Checker checks every extracted claim against primary sources you supply or that it can name and consult; it has no autonomous ability to detect falsehood independent of a verifiable source existing for every claim it checks. A verdict returned without an actually consulted source should be treated as a hypothesis for you to verify yourself, never as a confirmation.
Can a chatbot's answer be used as evidence in the report?
No. An answer generated by an AI, whatever the tool, is never an acceptable primary source for Fact-Checker, even when it comes from an earlier conversation with Claude itself. The pack requires an identifiable primary source — a piece of legislation, a published study, an official release, statistics from a named body — never a rewording produced by another AI system. AI-generated content, even when accurate in substance, remains text with no verifiable trail: it is exactly the kind of source Fact-Checker consistently rules out, regardless of how plausible it looks.
What should you do if no source can confirm a claim?
Mark it as unverifiable rather than false, and keep it in the list of unresolved points that accompanies the final report. The absence of accessible evidence proves nothing about the claim's actual truthfulness; it only signals that no consulted source settled the question, which remains useful information in itself for whoever must decide whether to publish. That marking stays valid until a new reliable source turns up, and it should be flagged explicitly to anyone who reuses the text.
Does the pack work without internet access?
Yes, but with reduced scope: without web search enabled, Fact-Checker can only check the text against documents explicitly attached to the conversation, unable to confirm or rule out a claim about recent events. In that setup, ask for a document-based comparison limited to the supplied materials rather than a check of current events, which the pack cannot materially perform without outside access. Turning web search back on before processing a text that references recent events avoids mistaking this technical limit for a genuine absence of available sources.
A verification report only has value if someone else can redo it identically from the same sources. That is exactly what the Fact-Checker verification pack structures, carefully separating what is confirmed, what is false and what simply remains unverifiable.
Once corrections are approved, improving the text's readability without touching its substance follows a different method: our guide to rewriting with Humanizer details how to work on style while fully preserving the nuances the verification restored.
Training 100% fundable via OPCO/FIFPL. Qualiopi-certified programme. To train your team on content verification with AI as part of our Master Claude training, contact Educasium and specify your status (employee, self-employed, business owner) and your goal.