How to Verify Political Videos

Video verification combines the fundamentals of image verification with additional techniques for motion, audio, and synthetic media. As generative AI lowers the cost of fabricating footage, the priority is not a single magic detector but provenance — establishing where a video came from and who first published it.

This guide assumes you have read the digital verification fundamentals and applies the same provenance-first mindset to moving images.

Step 1 — Find the original source

Identify who first posted the video and where. Because the same clip is re-uploaded endlessly, tracing the earliest copy is essential.

• Capture representative keyframes and run them through reverse image search.
• Look for the earliest upload across platforms and archives.
• Separate the viral account from the originating account.

Step 2 — Break the video into frames

Extract keyframes and verify them exactly as you would still images — reverse search, geolocate, and check context. Remember that miscaptioned real footage is far more common than deepfakes: a genuine clip from another time or place, relabelled for the current election.

Step 3 — Assess for synthetic media

For suspected deepfakes and other synthetic media, look for tell-tale artefacts:

• Lip-sync that drifts out of alignment with the audio.
• Unnatural or absent blinking, and warping around the hairline, teeth, or ears.
• Lighting and shadows inconsistent with the scene.
• Audio whose acoustics do not match the visible environment.

Detector tools can help, but they are probabilistic and easily fooled. Weigh them alongside provenance and context rather than treating them as a verdict.

Step 4 — Corroborate

Triangulate the footage against independent evidence:

• Other recordings of the same moment from different angles.
• Official statements, schedules, and credible reporting.
• A reconstructed timeline of when and where events occurred.

Step 5 — Preserve and log

Archive the video and its hosting page, save the file, and where possible record cryptographic hashes for chain-of-custody. Document every source and observation so your conclusion is reproducible and defensible.

Worked example: the "leaked" concession clip

A short clip appears to show a candidate conceding before polls close. Keyframe reverse search finds no earlier copies, but the audio's room acoustics don't match the visible outdoor setting, and the lip movements drift on specific phonemes. Rather than declare it "AI," the honest conclusion is: no verified provenance, multiple synthetic-media indicators, treat as unverified and likely manipulated — pending the original source.

Common mistakes

• Judging authenticity from compression artefacts alone.
• Ignoring the audio track, which often reveals manipulation.
• Over-trusting a single AI detector's confidence score.
• Amplifying the clip while "just asking questions" about it.

Key takeaways

• Provenance beats detection; find the earliest source first.
• Verify keyframes like images, and always check the audio.
• Deepfake signals are cumulative — no single artefact is proof.
• Preserve everything to counter the liar's dividend.

Next steps

Deepen your understanding of the technology in AI and Elections, and keep the Verification Handbook close for reference.