How FactHeck fact-checks a video or post

The five-stage pipeline

Every submission runs through the same automated pipeline. Each stage is independent and its output is recorded, so a verdict can always be traced back to the claims and evidence that produced it.

1. Ingestion. We fetch the public post from the source platform (TikTok, Instagram, YouTube, X and others), including the video or image and any caption text. Photo posts skip the audio steps and go straight to visual analysis.
2. Transcription. For video and audio, speech is transcribed so the spoken claims become text we can examine.
3. Claim extraction. A language model reads the transcript, caption, and on-screen/visual content and pulls out the discrete factual claims being made: the specific, checkable assertions, separated from opinion and rhetoric.
4. Evidence retrieval. For each claim we search for independent, authoritative sources and gather the most relevant evidence for and against it.
5. Verdict. Each claim is rated against the evidence, with a short explanation and links to the key sources used. The claim-level verdicts are then summarised into an overall assessment of the post.

Our rating scale

Each individual claim is rated on a seven-point scale, from fully supported to contradicted by the evidence:

• True: well supported by reliable evidence.
• Mostly True: accurate in the main, with minor caveats.
• Mixed: contains both accurate and inaccurate elements.
• Misleading: technically defensible but framed to create a false impression.
• Mostly False: largely contradicted by the evidence.
• False: contradicted by reliable evidence.
• Unverifiable: there is not enough reliable evidence to rate the claim either way.

The post as a whole receives an overall rating (Reliable, Mostly Reliable, Mixed, Questionable, or Unreliable) reflecting the balance of its individual claims. A claim we cannot verify is labelled as such rather than being treated as false: absence of evidence is not evidence of falsehood.

Verdict rubric

This section defines each label precisely so you can interpret results consistently.

Fact-check verdicts

• True — The claim is well supported by reliable, independent evidence. The core assertion is accurate and the framing does not distort its meaning.
• Mostly True — The claim is accurate in its main substance, but contains minor inaccuracies, outdated figures, or missing context that a careful reader should be aware of.
• Mixed — The content contains a combination of accurate and inaccurate elements that cannot be cleanly separated. Some parts check out; others do not.
• Misleading — The claim is technically defensible or even literally true, but is framed, cherry-picked, or presented out of context in a way that creates a false overall impression. The deception is in the framing rather than the facts themselves. This is distinct from Mostly False, where the underlying assertion is itself incorrect.
• Mostly False — The central assertion is largely contradicted by the evidence. There may be a kernel of truth or an accurate peripheral detail, but the main claim does not hold up. This differs from Misleading (where the framing is the problem) and from False (where the claim is contradicted without significant qualification).
• False — The claim is directly contradicted by reliable evidence. There is no meaningful accurate core. Use this label when the assertion is wrong, not merely spun or exaggerated.
• Unverifiable — There is not enough reliable, publicly available evidence to rate the claim in either direction. This is not a finding of falsehood. It means the evidence is absent, inconclusive, or too conflicting to support a verdict. Claims about private events, unpublished data, or topics where no authoritative source exists are candidates for this label.

The boundary that matters most: Misleading vs Mostly False vs False. A misleading claim may be literally accurate; a mostly false claim has an incorrect core assertion; a false claim is flatly contradicted. Unverifiable is never a synonym for false — absence of evidence is not evidence of absence.

AI-detection verdicts

• Likely AI — Multiple independent dimensions of the analysis (visual artifacts, temporal consistency, contextual plausibility, audio-visual alignment) flag the content as consistent with AI generation or synthesis.
• Partially AI / AI Edited — The content shows signs of partial AI involvement: elements that appear synthetic or manipulated alongside elements that appear genuine. This includes AI-edited footage, AI-generated inserts, or deepfake overlays on otherwise real video.
• Likely Real — The analysis found no significant indicators of AI generation or manipulation across the dimensions examined.

Each AI-detection result also carries a confidence level (high / medium / low) that reflects how many independent signals agreed and how strongly. A low-confidence result means the analysis was equivocal; treat it as a prompt for closer scrutiny rather than a definitive finding. High confidence means multiple independent dimensions reached the same conclusion, but it still does not constitute proof — see the limitations section below.

Sources and evidence

We prioritise primary sources and recognised, independent authorities: official statistics and government bodies, peer-reviewed research, and established news organisations, over anonymous or partisan posts. The specific sources used for each claim are listed with the verdict so you can check our working and reach your own conclusion. We link out to those sources directly; we do not ask you to take our word for it.

AI-generated content detection

Separately from claim-checking, FactHeck can assess whether an image or video shows signs of being AI-generated or manipulated. This is a probabilistic signal, not a certainty: detection of synthetic media is an evolving field, and a result should be read as "consistent with" rather than "proof of" AI generation.

Limits and known failure modes

FactHeck's verdicts are automated assessments produced by AI language and vision models working over retrieved evidence. They can be wrong, and users should treat them as a structured starting point for their own judgement rather than a final determination.

Known failure modes include:

• Context blindness. Models may miss irony, satire, or in-group references that change the meaning of a claim. A satirical post can be processed as a sincere assertion.
• Source availability. The pipeline can only assess evidence that is publicly accessible at the time of checking. Breaking developments, paywalled research, or specialist knowledge not yet reflected in indexed sources may be absent from the evidence set.
• Evidence weighting. Models may over- or under-weight a source based on how prominently it appears in retrieved results rather than its actual authority on the specific topic.
• Claim extraction errors. The extraction stage may miss claims, conflate distinct claims, or separate a claim from context that would change its verdict.
• Transcription errors. Accented speech, low audio quality, or non-standard terminology can introduce errors in the transcript that propagate through subsequent stages.
• AI-detection false positives and false negatives. Detection of AI-generated or manipulated media is probabilistic and an evolving field. Real content can be flagged as synthetic (false positive), and convincingly produced AI content may not be flagged (false negative). AI-detection results are especially susceptible to both error types and should not be relied upon as sole evidence of authenticity or inauthenticity.

All results — fact-check and AI-detection alike — are informational only. They do not constitute legal, professional, or editorial determinations. If a result matters materially, verify it independently through primary sources.

The role of AI, and its limits

FactHeck's pipeline is automated and uses large language models. That makes it fast and consistent, but it is not infallible. Models can misread context, miss sarcasm, over- or under-weight a source, or be limited by what evidence is publicly available at the time of checking. Verdicts are a well-sourced starting point for your own judgement, not the final word. When the evidence is thin or contested, we say so rather than manufacturing false confidence.

How we use AI in our guides

Our how-to and explainer guides are drafted with AI assistance and then reviewed and edited for accuracy by a named author before publication. A human author is accountable for the final content of every guide. This is separate from the fact-checking pipeline described above, which produces the per-claim verdicts.

Independence

FactHeck is operated by SAFEGUARDAI (MIRA) LTD. Verdicts are produced from the evidence by the pipeline described above; they are not for sale and are not influenced by the people, brands, or platforms a claim happens to involve.

Corrections

We will get things wrong sometimes. When we do, we fix it and say what changed. See our corrections policy. If you believe a verdict is mistaken, email hello@factheck.com with the link and what you think is wrong.