How to spot deepfakes and AI-generated video in news (2026 edition)
Synthetic video is now indistinguishable from camera footage in still frames. Here's what still gives it away — and the free tools you can use to verify any clip in under 90 seconds.
Two years ago, deepfake detection was a parlour trick: weird hands, melting backgrounds, eyes that didn't blink. In 2026 most of those tells are gone. The detection problem has shifted from the visual layer to the provenance layer — where did this clip come from, who first posted it, and is there a verifiable original?
Here's the workflow professional verification desks now use, simplified for ordinary readers.
1. Start at the source, not the clip.
Before doing any forensic work, find the earliest known posting of the clip. Run a reverse image search on a clean still frame using Google Images, TinEye, or Bing Visual Search. If the “earliest” appearance is a viral repost account with no original camera footage attached, the clip is suspect by default.
2. Look for camera-honest artefacts.
Real cameras produce specific physical artefacts: motion blur on fast-moving edges, rolling-shutter wobble on phone cameras, lens flare that moves with the light source. Modern synthesis is getting better at faking these, but it still struggles with compound motion — a person speaking *while* the camera pans *while* something else moves through the frame. Slow the clip to 0.25x and watch for hands that pop into place or hair that re-renders during transitions.
3. Check the audio independently.
Voice cloning is now the weakest layer of most deepfakes. Listen for: unnatural breath patterns (real people inhale mid-sentence; clones often don't), consistent room tone across cuts (synthesised audio usually has a sterile, evenly mixed background), and lip-sync that drifts on stressed consonants like “p” and “b”. Free tools like ElevenLabs’ AI Detector and PlayHT detector classify a clip in seconds.
4. Reverse-search the alleged location.
Most political deepfakes claim a specific setting (“Senator X at the Davos panel”). Verify that setting independently. Was Senator X actually at Davos that week? Was the panel real? News organisations now publish “event manifests” that list every public appearance — cross-reference before you re-share.
5. Use C2PA provenance where available.
Adobe, Microsoft, OpenAI, and major camera makers now embed C2PA credentials in real images and AI-generated images alike. Drag the file into contentcredentials.org and you'll see if it has a verified origin chain. Real news photos shot on modern cameras will carry C2PA stamps. AI-generated content from compliant tools will also carry them (and identify themselves as synthetic). Content with no C2PA from a major outlet is increasingly a red flag.
6. Triangulate with cross-platform reactions.
Deepfakes spread asymmetrically. The first 100 reposts are usually from accounts with similar profiles (matched ideology, low original-post ratio, recently created). Real footage spreads heterogeneously across the political spectrum within hours. Prism's Cross-Platform Snapshot is built for this — paste the clip's headline and see whether the reaction pattern matches an organic story or a coordinated push.
7. Wait 30 minutes before sharing.
This is the single most powerful tool. Verification orgs like AFP Fact Check, AP Fact Check, BBC Verify, and Bellingcat typically publish first-line debunks within 30–90 minutes of a major synthetic video going viral. If a clip seems too perfect a narrative hit on your political opponents, give the professionals half an hour.
The hard truth
We've crossed a threshold where a sophisticated synthetic video is now cheaper than hiring an actor. The infrastructure of trust is shifting from “does the footage look real” to “can I verify the chain of custody from camera to my screen.” That's a permanent change. The skills above are not optional any more for anyone who reads news online.
Related: How to spot AI-generated images · How to verify a viral video · How to fact-check a news article.