How to Detect AI-Generated Content in a Video or Recording
By Christopher Elley, Founder, FactHeck · Last reviewed 28 May 2026
Written with AI assistance and reviewed for accuracy by the author.
Detecting AI-generated content in a video or recording combines visual inspection with automated detection tools. For video, look for facial flicker at the edges of the face, inconsistent lighting between the subject and background, and audio that has a flat, studio-clean quality even in an apparently outdoor setting. These are signs of AI-synthesised or face-swapped footage. For audio-only recordings, listen for absent breath sounds between sentences, robotic cadence at the ends of phrases, and background noise that cuts sharply rather than fading. Free tools to test against: Hive Moderation detects AI-generated faces and voices; AI or Not covers images and short video; ElevenLabs' AI Speech Classifier checks whether audio was generated by an AI voice model. No detector is definitive on re-compressed social media content. Combine tool output with the visual and audio checklist and source verification before reaching a conclusion.
Why AI-generated video and audio are a growing verification problem
AI video generation tools (Sora, Runway, Kling, and others) now produce footage that can be indistinguishable from casual smartphone video. Voice cloning tools can replicate a person's voice with high fidelity after as little as 30 seconds of training audio. The result is a low-cost pipeline for producing fake "news clips" and fabricated quotes attributed to public figures. The Alan Turing Institute has documented both the scale of deepfake proliferation and the detection challenges it poses, while researchers at the MIT Media Lab Detect Fakes project track how detection methods keep pace, or fail to, with each new generation of synthesis models. Sensity AI (formerly Deeptrace) publishes annual statistics on the volume of deepfake content found in the wild.
What AI-generated video looks like: visual detection checklist
Watch the clip twice before reaching for any tool: once normally, and once with your attention on the hairline and ear boundaries rather than the face centre.
| Region | What to look for | Reliable on its own? |
|---|---|---|
| Hairline and ears | Temporal flickering frame-to-frame at the boundary between face and background | Fairly |
| Eyes | Reflections too symmetrical; don't shift correctly with head movement | Weak (newer models handle this better) |
| Teeth | Unnaturally uniform, slightly blurred between frames | Weak |
| Background periphery | Objects blending or distorting when the subject moves | Fairly |
| Face vs. background sharpness | Face appears sharper than background (sign of compositing) | Fairly |
| Hands | Wrong finger count, unnatural bending, fingers merging with adjacent objects | Fairly |
| Lighting | Face lit from a direction inconsistent with the background environment | Fairly |
What AI-synthesised audio sounds like: listening checklist
- Absent breath sounds: natural speech includes audible breaths between sentences; cloned audio often omits or misplaces them.
- Robotic prosody: syllable stress and sentence-final falling pitch sounds rule-generated rather than spontaneous.
- Emotional flat-lining: genuine speech has micro-variations in pitch and tempo; cloned speech tends to be more uniform throughout.
- Suspiciously clean noise floor: studio-quality audio in a claimed outdoor or candid setting is a red flag.
- Sibilant distortion: S and SH sounds are often hissy or over-compressed in voice-clone models.
- Lip-sync drift: even convincing deepfakes often show a 1 to 3 frame offset between audio and mouth movement, especially at cuts.
Try muting the clip: if the movement still looks natural without the audio, that is reassuring. If the performance only seems convincing with the audio cue, be more suspicious.
Free detection tools for video
| Tool | What it detects | Free / paid | Platform |
|---|---|---|---|
| Hive Moderation | AI-generated faces, video, and images | Free web demo | Web |
| Microsoft Video Authenticator | Deepfake face manipulation; confidence score per frame | Limited release (not public) | Enterprise only |
| InVID/WeVerify | Frame extraction, reverse search, metadata (investigative scaffolding rather than AI probability scoring) | Free | Browser plugin + web |
| AI or Not | AI-generated images and short video | Free tier | Web |
Microsoft demonstrated a Video Authenticator that scores manipulated regions frame-by-frame, but it was only released to a limited set of organisations and is not available as a public tool. Do not factor it into a workflow unless you have access.
Free detection tools for audio and voice
- ElevenLabs AI Speech Classifier: checks whether audio was generated by ElevenLabs or similar voice models; free upload via their website.
- AI or Not: covers audio alongside images.
- Resemble AI Detect (resemble.ai): voice authenticity scoring; free tier available.
Re-encoding degrades all audio detectors significantly: converting to MP3 or sharing via WhatsApp strips detail that detectors rely on. Treat audio detection results as indicative rather than conclusive, and always combine with the listening checklist above.
A step-by-step verification workflow
- Watch the clip twice: once normally, once muted. Does the movement look natural without audio cues?
- Watch again with attention on the hairline and ear boundaries, not the face centre.
- Extract a key frame (screenshot) and run it through Hive Moderation or AI or Not.
- Download the audio track and test it with the ElevenLabs AI Speech Classifier or Resemble Detect.
- Cross-reference: does the clip appear on official channels or credible news coverage?
- Reverse-search key frames via InVID/WeVerify to rule out real footage that has been recontextualised.
The limits of AI video detection: honest caveats
Detection accuracy drops sharply once footage has been re-uploaded to WhatsApp, Instagram, or TikTok, because the platform re-compression strips the fine-grained artefacts that detectors rely on. Each new generation of AI generators closes the visual gaps that today's tools exploit. The MIT Media Lab Detect Fakes project documents this arms race in detail. Hive Moderation's false positive rate on genuine content is non-trivial, so a positive result from any single tool should prompt deeper investigation, not a final judgement. No detector should be used as sole evidence in a consequential editorial or legal decision.
Reporting and escalation
If a clip appears to show a real public figure saying something damaging, do not publish or share it pending verification. Check the subject's official statements or press office first. Report the content to the platform through its reporting flow. In the UK, sharing non-consensual deepfake intimate imagery is a criminal offence under the Online Safety Act 2023. For high-stakes clips, submit to a professional fact-checker such as Full Fact or AFP Fact Check rather than relying solely on automated tools.
Frequently asked questions
Can you detect AI-generated video just by watching it?
Sometimes. Watch for flickering at the hairline and ear edges, a face that looks sharper than the background, inconsistent lighting between subject and scene, and lip sync that drifts by a frame or two at cuts. Muting the clip and watching again is a useful test. If the performance only convinces you when the audio is playing, treat that as a warning sign. Combine any visual suspicion with a detection tool before drawing a conclusion.
What are the best free tools to detect AI-generated video?
Hive Moderation offers a free web demo that covers AI-generated video and images. InVID/WeVerify is free and provides frame extraction, reverse image search, and metadata analysis, which is useful investigative scaffolding even though it does not produce an AI probability score. AI or Not covers both images and short video clips. Run at least two tools and compare; agreement between independent tools is more meaningful than any single result.
How can I tell if a voice recording is AI-generated?
Listen for absent or misplaced breath sounds between sentences, a robotic cadence at the ends of phrases, and a suspiciously clean noise floor in a clip that claims to be a candid recording. ElevenLabs' AI Speech Classifier and Resemble AI Detect are free tools that score voice authenticity. Bear in mind that re-encoding to MP3 or sharing via messaging apps significantly degrades detection accuracy.
Are AI video detectors accurate after a video has been shared on social media?
Accuracy drops significantly once a video has been re-compressed by platforms such as WhatsApp, Instagram, or TikTok. Re-compression strips the fine-grained artefacts that detectors rely on. Treat any result on a re-shared clip as indicative rather than conclusive, and corroborate it with visual inspection and source verification.
Want a second opinion in seconds? Paste the video link into FactHeck. The AI detection pipeline analyses the footage and delivers a sourced verdict on both authenticity and the claims it contains.