Can AI fact-check a message? What it can and cannot establish

Start with the exact claim, not the answer

Before asking a tool whether something is true, write the claim in one sentence. Include the person or institution, the action, the date and the place when those details matter. A message saying that an office is closed today is different from a message saying that it closed last year. If the answer addresses the wrong version, even excellent research will not settle your question.

Separate statements that can be checked from opinions, predictions and requests. A forward may combine a real announcement with an unsupported explanation of why it happened. Ask the tool to list those separately. This makes it easier to notice when one verified detail has been used to make the whole message sound established.

A fluent explanation is not a record of what happened

Language models can produce convincing but incorrect details and references. Search tools can reduce some gaps by retrieving material, but the final explanation still needs checking. A sentence that sounds cautious can be wrong; a sentence that sounds certain can be unsupported. Treat tone as writing style, not as a measurement of evidence.

A useful follow-up is: Which source supports this specific sentence, and what remains uncertain? Then open that source yourself. If the tool supplies a title that cannot be found, or a link to a page that never makes the claimed statement, remove that item from your evidence rather than asking the same tool to reassure you.

Check the link and the relationship to the claim

A working citation is only the beginning. Read the relevant passage with the surrounding paragraph. Check whether the cited person is making an allegation, reporting an observation or presenting a verified record. Those are different kinds of information.

Consider this hypothetical example: a post says every school in a state is closed tomorrow. Its citation describes a holiday for one district last month. The page may be authentic and the summary grammatically clear, but it does not support the statewide claim. Note the mismatch in scope, place and time rather than simply marking the source as good or bad.

Ask what the tool actually had access to

A tool may receive a screenshot, a short transcript or a preview image when you expected it to inspect a full video. Missing sections can change the meaning. An automatic transcript can also confuse names, numbers, negation or overlapping speech. Compare the claimed quotation with the original recording when possible.

For a long or inaccessible item, identify the gap explicitly: the complete recording was unavailable, the key sentence was inaudible, or only the caption could be read. Do not fill the missing portion with a plausible story. A clear account of what was examined is more useful than an impressive answer based on an assumed input.

Uncertainty is a useful result

Not finding support is different from finding evidence against a claim. A new local event may not have an accessible public record yet. A poorly worded search may also miss relevant information. The appropriate conclusion may be that the available material does not settle the question.

Keep a short note of the evidence needed next: the dated notice, the full recording, the original dataset or an independent account. That gives you a stopping rule. Repeating the same question until one model says yes creates apparent agreement without adding evidence.

A practical review before you forward

You do not need to investigate every forwarded item indefinitely. If you cannot support the consequential part, pause sharing it as fact. If you share an unresolved question with someone who can help, preserve the uncertainty and avoid repeating personal accusations.

• Does the answer address the exact claim in the message?
• Can I open the source and find the supporting passage?
• Do the date, location and people match?
• Are several links repeating the same original account?
• Does the conclusion acknowledge material that could not be checked?

Where ForwardCheck fits

ForwardCheck is an AI-assisted beta that helps examine forwarded content and show the sources behind a report. It is not an independent newsroom, and an AI-generated report is not a human investigation. Read its methodology and the actual sources before relying on a result.

Use a report to organise your next check, especially when the claim or evidence is complicated. Do not turn its verdict into an accusation about the sender, or treat a label as a measured guarantee of accuracy. If an important source is missing or the claim was misunderstood, that is a reason to question the report.

Common questions

If two AI tools agree, is the claim verified?

No. They may rely on the same material or make the same inference. Independent supporting evidence matters more than the number of matching answers.

Does a citation prove an AI answer is correct?

No. Open it and check that it supports the exact claim, including its date, location and scope.

Does NOT PROVEN mean false?

No. It means the available evidence did not establish or rule out the claim. Evidence against a claim is a different finding.