A detector score can be a useful reason to look closer, but it cannot prove a video is fake, establish who made it, show intent, or decide whether its caption is true. Careful verification combines the exact clip, source context, available media quality, corroboration, and a visible record of uncertainty.
Quick answer
- Preserve the direct URL, account, caption, date, and full clip before it changes.
- Find the earliest available version and compare runtime, framing, soundtrack, and context.
- Check independent corroboration—not only whether a face looks unusual.
- Run a cautious AI-video screening check, then read result state, confidence, likelihood, checked signals, and limitations together.
- Escalate identity, legal, financial, safety, or public-interest claims to qualified human review.
Separate the questions
| Question | Useful evidence | It cannot settle alone |
|---|---|---|
| Is this an original or altered copy? | Earliest upload, file history, provenance. | Whether the attached claim is true. |
| Are signals suspicious? | Screening, frame and motion review. | Intent, identity, or legal authenticity. |
| Is the post accurate? | Reporting, records, witnesses, full context. | Whether every pixel is generated. |
Preserve source and context
Record the direct URL, account, caption, upload date, duration, and visible labels. Note missing evidence: a private, cropped, shortened, screen-recorded, or heavily compressed post carries less usable information. NIST advises considering provenance, editing history, compression, clarity, other angles, and source purpose.NIST: Is This a Deepfake?
Search for the original creator, a longer version, an official source, livestream, press footage, or independent angle. The link-versus-upload guide explains why public access can still limit screening.
Use provenance and screening carefully
Content Credentials can provide tamper-evident asset-history information; they do not judge whether a real-world claim is factual.C2PA: Content Credentials explainer Their absence does not prove AI generation.
Read a screening report as evidence, not a verdict. The evidence report guide explains why likelihood, confidence, checked and unchecked signals, and limitations must travel together. NIST’s Guardians program also evaluates detection under changing conditions such as compression, so performance needs ongoing assessment.NIST: Guardians of Forensic Evidence
Inspect patterns and corroborate
Watch several continuous moments, not one paused frame. Check faces, hands, reflections, text, shadows, and movement, while remembering that compression, translation, dubbing, clipping, and replacement audio can look unusual without proving a deepfake. Read why compressed video is harder to verify before drawing a conclusion.
For a consequential claim, look beyond the file: independent reporting, official statements, public records, full recordings, and other footage often provide stronger context. “This repost is too short or processed to support confidence” can be the most accurate conclusion.
Frequently asked questions
Can a detector prove a video is a deepfake?
No. It provides likelihood-based screening signals from accessible evidence, not proof of intent, identity, factual truth, or legal authenticity.
Does a low score prove a video is real?
No. A low-confidence or limited result can mean the input did not support a fuller review.
What should I do with a high-stakes clip?
Preserve its context, locate primary sources and seek qualified human review before an accusation or consequential decision.
Bottom line
Combine source context, earliest available media, provenance when available, cautious screening, and independent corroboration. A detector should tell you what to check next—not turn a score into an accusation.