Do not begin with a detector. Begin by preserving the claim, locating the first accountable source and checking whether independent evidence supports it. A detector can be one input, but it should not decide whether you trust, share or report a suspicious video, audio clip or image.

Appearance alone cannot settle whether a clip is authentic. NIST’s synthetic-content report surveys several approaches, including provenance and detection. The workflow below treats technical signals as leads to check against the original source and independent evidence.

1. Preserve the context before it disappears

Save the post URL, account name, timestamp, caption and surrounding replies. Take a screenshot of the page and, where permitted, retain the original file rather than only a screen recording. The exact claim matters as much as the pixels: a genuine clip can be paired with a false location, date or translation.

2. Find the earliest accountable source

Trace the asset back to the first uploader you can identify. Is it an official account, a named journalist, a public institution or an anonymous aggregator? Search distinctive phrases from the caption, and use reverse-image or key-frame search where relevant. Treat search results as leads and compare them against the original publication. This is an editorial verification workflow, not a guarantee that any single tool detects manipulation.

3. Check the event, not just the file

Look for independent reporting, official statements, schedules, public records and other contemporaneous material. If a public figure is allegedly speaking, check their verified channels and whether credible outlets reporting from the location have corroborated it. One repost that cites another repost is not independent confirmation.

4. Inspect provenance when it exists

Content Credentials can reveal signed information about a file’s source and changes. Treat an intact credential as useful evidence, then inspect who signed it and what it records. Do not infer that missing credentials prove synthetic origin: many authentic files have no provenance record, and an absent record does not explain why provenance is unavailable.

5. Use technical tools as leads, not final answers

Visual anomalies, audio glitches and detector scores can help determine what to investigate next. They are not reliable substitutes for corroboration. A responsible conclusion may be “unverified” rather than “fake,” especially while a news event is developing.

6. Report harm through the right channel

If the content impersonates someone, facilitates fraud, contains non-consensual intimate material or may violate law or platform rules, report it through the platform’s reporting flow and include the preserved context. In the EU, the Digital Services Act requires platforms to provide notice-and-action mechanisms for illegal content and to notify the reporter of the final decision. The Commission’s separate user-rights guidance describes explanations and complaint options.

For imminent personal, financial or physical risk, contact the relevant local authority or the affected organisation directly. Do not amplify a suspected deepfake merely to ask whether it is real. The most useful public response is a specific one: state what is known, name what remains unverified and link to the original evidence.

Sources