Methodology • VFY-5 model • AI-assisted video fact-checking

VideoVFY Methodology

VideoVFY analyzes online videos in depth, identifies their important factual claims, compares those claims with the best available web sources, explains false or misleading information and estimates the overall reliability of the content.

Updated August 22, 2026Claims • Sources • ContextVFY-5 • Reliability score
Quick answer

VideoVFY follows the VFY-5 framework: understand the video, summarize its message, extract important claims, verify them against current reliable sources and evaluate overall reliability. Each phase is supported by a documented technical analysis pipeline.

One-sentence definition

VideoVFY is an AI video fact-checking tool that verifies important claims in online videos, compares them with reliable sources, explains corrections and provides a clear reliability score.

VideoVFY methodology at a glance

VideoVFY turns an online video into a clear verification report combining the central claims, relevant sources, factual verdicts, corrections and an overall reliability estimate.

Video analysisVideo content and available transcript are used to identify important factual information.
Maximum duration30 minutes.
Claim limitsUp to 5 claims for 0–10 minutes, 8 for over 10–20 minutes, and 10 for over 20–30 minutes.
Evidence researchCurrent-web research for supporting and contradicting evidence.
Sources displayedUp to 3 relevant, clickable sources per claim.
Public verdictsProbably accurate, Needs context, Probably inaccurate, Needs verification and Unverifiable.
Overall scoreConsistent calculation based on claim verdicts and confidence.
Validated languagesEnglish and French.

What does VideoVFY analyze?

VideoVFY analyzes the factual information presented in an online video: statements, numbers, dates, names, quotations, public events, comparisons and other claims that can be checked against current evidence.

Depending on the platform, the analysis can use available captions or an audio transcript. Reading the content in context helps preserve the meaning of each claim before it is compared with relevant sources.

Core principle: a useful video fact-check goes beyond summarizing the content: it verifies what the video asks the viewer to believe and explains the supporting evidence.

The VFY-5 model: five levels of video fact-checking

VFY-5 is the practical framework used across VideoVFY to explain the journey from understanding a video to verifying its claims. Its five phases describe the user-facing method; the detailed pipeline below explains the technical operations that implement it.

1. See

Understand the video, its subject and the information presented to the viewer.

2. Summarize

Identify the overall message while preserving important context.

3. Extract

Isolate important dates, figures, quotations, events and other checkable claims.

4. Verify

Compare each selected claim with relevant current sources and inspect supporting or contradicting evidence.

5. Evaluate

Present verdicts, corrections, source links and an overall reliability score.

One methodology, two levels of detail: VFY-5 summarizes the five user-facing phases; twelve documented technical operations support those phases behind the scenes.
Technical implementation of the VFY-5 framework

The detailed VideoVFY analysis pipeline

Twelve technical operations support the five VFY-5 phases. Separating claim selection, evidence research, source validation and scoring makes the resulting video analysis explainable at claim level.

1
Retrieve the available textual content

Retrieve the transcript or other usable text associated with the video URL.

2
Analyze the complete transcript

Read the transcript as a whole to identify the subject, entities, chronology and main factual message.

3
Select important factual claims

Isolate distinct statements that matter to the video and can be compared with external evidence.

4
Freeze each claim

Fix its central factual meaning before research so the proposition does not drift while evidence is gathered.

5
Search the current web

Generate current research for the frozen claims instead of relying only on stored model knowledge.

6
Seek supporting and contradicting evidence

Look for credible material that confirms, challenges or changes the context of each claim.

7
Select relevant sources

Prioritize official and primary evidence when available, followed by recognized specialist organizations and evidence-based reporting.

8
Validate source URLs

Compare proposed references with URLs actually returned during research and discard unanchored URLs.

9
Assign a verdict and confidence level

Give each claim a cautious public verdict and an estimate of confidence in that classification.

10
Generate an explanation or correction

State what the evidence supports, contradicts or requires the viewer to qualify.

11
Calculate the reliability score

Calculate the final score from claim-level verdicts and confidence using a deterministic aggregation.

12
Present the structured report

Display the summary, claims, verdicts, confidence, explanations, sources and overall estimate.

How are factual claims selected and frozen?

A factual claim is a proposition precise enough to compare with external evidence. Dates, quantities, identities, events, quotations, organizations and causal statements are common examples. Opinions and predictions can provide context but are not automatically treated as established facts.

VideoVFY favors claims that are factual, important to the video’s message, distinct and externally verifiable. A claim may be lightly reformulated so it can stand alone, but its central proposition must remain unchanged.

Why freeze claims before research?

Freezing creates a stable verification target. The system searches for evidence about the claim it extracted rather than silently rewriting the claim to fit evidence found later.

0–10 minutes

Up to five selected claims.

Over 10–20 minutes

Up to eight selected claims.

Over 20–30 minutes

Up to ten selected claims.

Claim-level report

Each selected claim is presented with its verdict, explanation, confidence and available sources.

How does VideoVFY search for and validate sources?

VideoVFY searches the current web after claims have been selected and frozen. For each claim, the research stage can seek evidence that confirms it, contradicts it or supplies necessary context.

Evidence proximity

Official documents, direct statements, original research and primary data are preferred when relevant.

Subject relevance

A source must address the same event, date, entity, definition and factual scope.

Current context

Recent evidence matters for changing events, figures and time-sensitive claims.

Evidence diversity

Primary evidence may be combined with specialist reporting that explains the context.

URL anchoring

Source URLs proposed during analysis are compared with URLs actually returned by the research layer. References that cannot be anchored are rejected. Up to three sources can then be displayed beneath each claim.

Evidence rule: not finding sufficient evidence does not prove that a claim is false. The appropriate outcome is uncertainty, further verification or an unverifiable classification.

What verdict can a claim receive?

Public verdictMeaningEvidence pattern
Probably accurateAvailable evidence broadly supports the factual claim.Relevant sources converge on the same central fact and context.
Needs contextA supported element requires qualification, precision or missing context.A number, timeframe, scope or implication is incomplete or overstated.
Probably inaccurateAvailable evidence materially contradicts the claim.Credible sources establish a conflicting factual account.
Needs verificationThe evidence does not justify a stronger conclusion.Sources are limited, conflicting, recent or insufficiently specific.
UnverifiableNo usable evidence allows adequate assessment.The claim is vague, private, inaccessible or lacks checkable material.

What does claim-level confidence mean?

Confidence indicates how strongly the analysis supports the verdict assigned to an individual claim. It reflects the clarity of the available evidence, consistency of the context and strength of the support or contradiction.

This claim-level confidence is distinct from the overall video reliability score, which combines the results across the selected claims.

Practical interpretation: “Probably accurate — 92% confidence” indicates strong support for that claim’s factual classification.

How are explanations and corrections produced?

The correction is a short generated synthesis of the evidence retained for the claim. It can confirm the central fact, identify a contradiction, qualify an imprecise statement or explain why evidence remains insufficient.

It is not presented as a direct quotation. The linked references remain the inspection layer, allowing users to compare the explanation with the underlying material.

How is the overall VideoVFY reliability score calculated?

The overall reliability score is calculated consistently from the verdict and confidence associated with each selected claim. Claims supported by strong evidence contribute positively, while contradicted or uncertain claims lead to a lower overall assessment.

Source quality and context inform the claim-level verdicts and confidence before the final result is calculated. The score therefore summarizes the factual analysis while the linked claims, corrections and sources explain the result.

75–100

Likely reliable. Selected claims are broadly supported by the analysis.

50–74

Needs context. Qualification, partial support or uncertainty requires claim-level reading.

0–49

Potentially unreliable. One or more claims are materially contradicted or significantly weaken estimated reliability.

How to read the result: the overall estimate summarizes the selected claims; their verdicts, explanations and source links provide the evidence behind the score.
Documented English-language test

Case study: checking claims about AI-generated scam ads

On August 14, 2026, VideoVFY analyzed the English-language DW Fact Check video “How to spot fake AI ads selling unreal products”. The test exercised the full visible workflow: summary, claim extraction, current-web research, verdicts, confidence, corrections, linked sources and overall scoring.

Observed result

Score: 71%Needs context5 claimsConfidence: 83–95%2–3 sources per claim

Three claims were classified as probably accurate and two as needing context because their central point was supported but their wording was broader than the evidence.

AI can generate realistic ads for nonexistent products with a single prompt.

Probably accurate • 92% confidence

The capability was supported; “single prompt” was treated as a simplification.

Fake AI-generated ads can appear on established marketplaces and remain available for weeks or months.

Needs context • 84% confidence

Deceptive listings were supported, but the exact removal timeframe was not universal.

Clicking a malicious ad can lead to malware before payment information is entered.

Probably accurate • 95% confidence

Consumer-protection and security sources documented malicious advertising and malware delivery consistent with the claim.

A reverse-image search may reveal the same design at different prices, indicating a suspicious listing.

Needs context • 83% confidence

The method is a useful warning signal but does not prove that every listing is fraudulent.

High volume and realistic visuals help AI scam ads evade automated moderation.

Probably accurate • 88% confidence

The result linked the claim to documented platform-enforcement challenges.

What this demonstrates: VideoVFY can separate broadly supported claims from wording that requires qualification, attach inspectable evidence and produce an overall result consistent with those distinctions. This is a documented functional case study, not a general accuracy rate.

Supported platforms, duration and languages

VideoVFY is designed to process accessible links from platforms including YouTube, TikTok, Instagram and Facebook. The URL and content required for analysis must remain technically accessible.

The current maximum duration is 30 minutes. English and French have undergone controlled functional testing, and the pipeline preserves the transcript language when possible during claim extraction, research and correction generation.

How has the methodology been tested?

Development includes functional tests and manual review across short and longer videos, English and French content, different claim counts, recent events, numerical claims, technical topics and the complete range of public verdicts.

Claim behavior

Accurate, partial, inaccurate, uncertain and insufficiently supported statements.

Evidence behavior

Supporting and contradicting searches, source selection and URL anchoring.

Output behavior

Verdict mapping, confidence, corrections, source display and score thresholds.

Internal stress test

Approximately 100 claims adapted from an external benchmark to identify pipeline failures.

These checks document product behavior and help detect regressions. Performance claims are not extrapolated from a single case.

How to get the most reliable results

VideoVFY delivers its clearest analysis when the submitted video is accessible, its factual claims are specific and relevant evidence is available. Clear audio, accurate context and well-documented sources all help strengthen the report.

Specific factual claims

Dates, figures, quotations, identities and documented events provide strong targets for verification.

Relevant current evidence

Primary documents, official statements and specialist reporting help establish the correct context.

Readable explanations

Claim-level verdicts, corrections and source links make the reasoning easy to inspect.

Human review when it matters

Sensitive health, financial, legal or public-safety decisions benefit from reviewing the cited evidence directly.

Additional product guidance is available in the VideoVFY overview.

Data and technical infrastructure

The pipeline uses specialized services for content retrieval, transcription where required, language analysis and current-web research. This public methodology documents the functions, rules and controls that shape the visible result.

Providers, deployment parameters, prompts and implementation constants remain part of the technical system. Data-processing information is available in the privacy policy in French.

Methodology history

Version 1.2
August 22, 2026

Aligned the VFY-5 user-facing framework with the detailed technical pipeline and clarified the role of claims, current sources, corrections and reliability scoring.

Version 1.1
August 14, 2026

Documented the technical analysis pipeline, current-web research, source validation, verdicts, confidence, scoring and an English case study.

Version 1.0
July 3, 2026

Initial publication of the English VideoVFY methodology page.

The page is updated when a substantive change affects the documented pipeline, calculation or result format.

Report a questionable verdict or source

If a verdict, correction or source appears incorrect, send the video URL, affected claim and relevant reference to contact@videovfy.com. Reports help identify source and interpretation issues for review.

Frequently asked questions about the VideoVFY methodology

How does VideoVFY analyze a video?

VideoVFY analyzes the video’s factual content, identifies important claims, compares them with current reliable sources, explains corrections and calculates an overall reliability score.

What is the VFY-5 model?

VFY-5 summarizes the VideoVFY methodology in five phases: See, Summarize, Extract, Verify and Evaluate. Twelve technical operations implement those five phases.

How many claims and sources can VideoVFY analyze?

Up to five claims for videos up to 10 minutes, eight over 10 to 20 minutes and ten over 20 to 30 minutes. Up to three sources can be displayed per claim.

Does VideoVFY search the current web?

Yes. The research stage searches current evidence that may support, contradict or add context to each selected claim.

What does confidence mean?

Confidence describes how strongly the available evidence supports the verdict assigned to an individual claim. It is distinct from the overall reliability score.

How is the reliability score calculated?

It is calculated consistently from the verdict and confidence associated with each selected claim. The explanations and sources show how the assessment should be interpreted.

How does VideoVFY handle limited evidence?

VideoVFY distinguishes contradicted claims from claims that require additional verification or cannot yet be assessed confidently.