How to Verify a Video Before You Share It

How to Verify a Video Before You Share It

To verify a video, begin with the claim and original source, not with a search for visual glitches. Preserve the URL and context, find the earliest available version, test the stated date and location, inspect provenance, seek independent corroboration, and use technical detectors only as supporting signals. If the evidence remains weak, do not share the clip as fact.

The threshold should rise with the harm a wrong share could cause.

First, state what the video is supposed to prove

"Is this real?" is too broad.

A video may be an authentic recording attached to the wrong date. It may show a real event in another country. It may be a staged scene described as spontaneous. It may be generated fiction accurately labeled by its creator. It may contain a real person whose audio was replaced.

Write one testable claim:

"This video shows the mayor of City A accepting cash from Contractor B on July 26, 2026."

Now the work has parts. Does the person match? Is the place City A? Was it recorded on that date? Is the exchange what the caption says? Is the audio original? Who published it, and how did they obtain it?

flowchart TB
    A["Video plus caption"] --> B["Write the exact claim"]
    B --> C["Preserve source and versions"]
    C --> D["Find earliest available origin"]
    D --> E["Test time, place, edit, and provenance"]
    E --> F["Seek independent corroboration"]
    F --> G{"Evidence state"}
    G --> H["Supported"]
    G --> I["Refuted"]
    G --> J["Unresolved"]

The output is an evidence state, not a feeling.

1. Preserve the source before it changes

Record the page URL, platform, account, visible caption, upload time, access time, and the identity claimed by the uploader. Save screenshots of the surrounding post when lawful and safe. If you download the file, keep the original download separate from clips or screenshots you create.

Do not redistribute harmful media merely to prove you saw it. Intimate imagery, child sexual abuse material, graphic violence, classified material, active threats, and other dangerous or illegal content require specialized handling. Use platform reporting, counsel, investigators, or emergency services as appropriate.

For an ordinary public clip, preservation helps compare later edits and deletions. It does not create a forensic chain of custody.

2. Find the earliest available version

Reposts remove context. They may crop a watermark, replace audio, change the caption, or start after the relevant event.

Search distinctive phrases from the caption. Look for account handles, logos, signs, uniforms, landmarks, and spoken names. Extract several clear keyframes and run reverse image searches. AFP's professional guide explains how keyframes can reveal earlier appearances of a video. Google News Initiative also provides reverse image search training.

The earliest version you can find is not automatically the original. It is a better lead.

Contact the uploader if the claim matters. Ask whether they recorded it, when and where it was captured, what happened before and after, whether they edited it, and whether they can provide the original file. Treat their answers as claims that may need support.

3. Reconstruct the context

Check whether the video existed before the alleged event. Compare the weather, daylight, shadows, language, architecture, terrain, road markings, transit, license plates, uniforms, and public schedules with the claimed place and time.

Look for longer versions. A ten-second clip may exclude the action that changes its meaning.

The UK House of Commons Library's May 2026 verification briefing recommends source checks, reverse search, metadata, visual clues, time, location, and corroboration. It also makes an important point: some investigations remain uncertain.

Do not turn one matching landmark into proof of the whole caption. Location evidence does not establish date, speaker identity, or causal interpretation.

4. Inspect the edit and audio

Watch the video more than once. Listen separately from the picture. Note cuts, changes in aspect ratio, repeated frames, abrupt ambience, inconsistent room tone, missing reactions, and subtitles that make stronger claims than the audible speech.

Find the same event from another angle if possible. Compare the sequence, not only a striking frame.

Transcribe material speech and search a distinctive sentence. Confirm translations with a qualified speaker when meaning matters. A plausible subtitle is not evidence that the speaker said those words.

These checks can identify questions. They are not a universal list of synthetic artifacts. Compression, stabilization, low light, rolling shutter, frame interpolation, and ordinary editing can make authentic footage look strange.

5. Check Content Credentials when they exist

Content Credentials can expose signed assertions about an asset and its history. A viewer may see who or what signed a manifest, declared creation or editing actions, ingredients, and whether validation succeeded.

The current C2PA 2.4 specification binds manifests to assets through cryptographic mechanisms. Its FAQ also explains that embedded metadata can be removed and that some implementations may recover credentials through soft bindings or external storage.

Interpret each result narrowly.

Provenance is evidence about origin and handling. Truth still needs claim-specific verification.

6. Seek independent corroboration

Search for records that do not all trace back to the same post. Another account repeating the caption is not independent.

Useful corroboration may include other recordings from different positions, official logs, public schedules, weather records, contemporaneous reporting, direct witnesses, business records, sensor data, or a response from the person depicted.

Consider incentives and access. A source can be first and still be wrong. An official source can have authority and still need scrutiny. A crowd can repeat one fabricated origin until it looks like many sources.

Write down what each record independently establishes.

7. Use detectors last

An AI or manipulation detector can be useful for triage. It may produce a score, classification, heat map, or explanation.

NIST's synthetic-content report describes content-based detection as probabilistic and emphasizes testing conditions. A detector can fail when generators, formats, compression, or edits differ from its evaluation data. False positives can harm real people.

If you use one, preserve the tool name, version, date, input file, settings, score, threshold, and published limitations. Never publish "AI-generated" as a fact based only on a score.

The question is not whether a tool sounded confident. It is whether its tested capability matches this asset and claim.

8. Decide whether the evidence supports sharing

Use three outcomes.

Supported means the available source, context, provenance, and corroboration support the stated claim at the required threshold.

Refuted means reliable evidence conflicts with a material part of the claim.

Unresolved means the evidence is incomplete, conflicting, inaccessible, or too weak for the consequence.

Unresolved is a valid answer. If you share for the purpose of seeking help, state what is unverified, avoid repeating the allegation as fact, limit unnecessary exposure, and link the source rather than laundering the claim through confident commentary.

Match the threshold to the harm

A mislabeled comedy clip and a video accusing someone of a crime should not receive the same treatment.

When doubt and harm are both high, pause.

The durable habit

Do not ask the pixels to answer every question. A realistic video can carry a false caption. A strange video can be authentic. A valid credential can describe a staged event accurately. A missing credential can belong to a genuine recording.

Start with the claim. Preserve the source. Reconstruct the context. Inspect provenance. Corroborate the event. Use technical signals carefully. Say what remains unknown.

This guide was freshly written from E070, current public verification guidance, C2PA 2.4, and NIST AI 100-4 reviewed on July 28, 2026. It is not a forensic or legal protocol. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.