Synthetic Videos on Social Media: Verify Before You Share
Before sharing a suspicious video, check what its caption claims and whether the source supports that claim. A real recording can be old or mislabeled. An AI-generated clip can look convincing. Appearance alone does not settle either question.
A practical verification checklist
1. Preserve the post and file
Save the post URL, account, caption and upload time. Keep the best available file without editing it. A screenshot can help document a post, but it usually loses detail and file-level provenance needed for analysis.
2. Trace the earliest available source
Look for earlier uploads, the original publisher and a longer version of the clip. An account badge or high view count does not authenticate the event. Check whether the claimed person or organization published the same material through a known channel.
3. Search representative frames
Extract several clear frames from different parts of the video and run reverse-image searches. Compare landmarks, text and objects with earlier images or reporting. A match can reveal an old clip reused with a new caption. No match leaves the question open.
The InVID-WeVerify plugin provides tools for video verification. Check each search service's supported inputs rather than assuming it accepts a video upload.
4. Review the sequence and provenance
Watch the beginning, middle and end, including transitions. Note the timestamp of any changing face edge, object, reflection or audio mismatch. Compression, stabilization and ordinary editing can cause anomalies, so compare several frames and obtain a better copy when possible.
Check Content Credentials if available. Read the signer and recorded actions alongside the validation result. Valid credentials verify signed provenance claims, not the truth of a caption. Missing credentials are not proof of manipulation.
5. Use a detector as an additional check
Open the video detector for the saved file. DeepFakeCheck supports MP4, MOV and WebM up to 100 MB, with up to 10 timestamped frames sampled across the video. It does not review every frame; a short manipulation can fall between samples.
Interpret the displayed risk score as a screening signal, not a calibrated probability that the video is fake. The result cannot verify the identity of a speaker, the date or the event. Review the cited frame evidence and seek independent corroboration before making an accusation.
A limited free allowance is available without signup. Analysis content is retained for up to 7 days, then deleted. Processing time varies; an immediate result is not guaranteed.
6. Check the claim outside the file
Compare the event with independent reporting, official statements or records. Multiple posts repeating the same source are not independent confirmation. For a request involving money or access, pause and contact the claimed sender using a known number or account.
7. Decide whether the evidence supports sharing
If you cannot establish the source or context, avoid presenting the clip as confirmed. If you share it to request verification, describe the unresolved question and avoid exposing private information. Preserve the evidence if you need to report fraud or a platform-policy violation.
What a benchmark can and cannot tell you
The NIST Open Media Forensics Challenge provides a framework for evaluating media-forensics systems. A benchmark result applies to the tested system, data and conditions. It does not guarantee performance on a particular compressed social-media clip.
This article does not claim that DeepFakeCheck has participated in, been certified by or received a score from that challenge. A product's result should be assessed on its actual evidence and limitations.
Keep a short review record
Record the original URL, the best file obtained, frames searched, provenance findings and detector observations. Separate what those checks establish from what remains unknown. That record makes it easier for another person to review your conclusion.
For more detail, use the step-by-step video verification guide and the seven visual and audio checks.