Tool 03 — Video
Where did this clip come from?The free AI video detector built on container forensics.
A free AI video detector built on container forensics. We scan for encoder atoms, AI generator markers (Sora 2, Runway Gen-4, Pika, Kling, Veo 3, Hailuo) and C2PA provenance, then sample frames so you can judge with your own eyes.
Aipurity is a free AI video detector and deepfake checker that reads the container’s actual evidence — and hands you the frames — instead of a made-up 97% confidence score.
- Runs 100% in your browser
- Clips are never uploaded
- No fake confidence percentages
Drop a file, or click to choose
Analysed locally in your browser. Nothing is uploaded, ever.
What this AI video detector actually reads.
AI generator markers
Sora 2, Runway Gen-4, Pika, Kling, Veo 3 and Hailuo exports sometimes carry their names in container metadata or the byte stream. When present, that’s a direct confession.
Container atoms & encoder tags
MP4 “©too” atoms and FFmpeg version strings record the writing software. A fresh Lavf tag with no camera metadata means the original provenance is gone.
C2PA provenance
OpenAI attaches Content Credentials to Sora downloads, Google marks Veo output with SynthID, and the EU AI Act forces machine-readable marks on AI video from August 2026.
Sampled frames for your own eyes
We pull three frames across the timeline. Human review of hands, text, physics and continuity is still one of the most reliable checks.
How to tell if a video is AI-generated.
- 01
Scan the container locally
Head and tail of the file are scanned byte-by-byte for generator fingerprints, encoder atoms and provenance boxes. Nothing is uploaded.
- 02
Decode and sample
The browser decodes the video and captures frames at 12%, 50% and 88% of the timeline for visual inspection.
- 03
Verdict from evidence only
“Detected” means the file names an AI tool. “Inconclusive” means the trail is clean or stripped — the case for almost anything off a social platform.
Generators we look for
What evidence each major AI video generator can leave — and the caveat that applies to every AI-generated clip.
| Generator | Evidence it can leave | Caveat |
|---|---|---|
| OpenAI Sora / Sora 2 | C2PA Content Credentials manifest and visible watermark on app downloads | Platform re-uploads and screen recordings strip both |
| Google Veo 3 | SynthID watermark; container marks on some export paths | SynthID needs Google’s own verifier; we read what the container shows |
| Runway Gen-4 | Metadata markers on direct exports, version-dependent | A clean export leaves no trail — verdict stays “inconclusive” |
| Kling 2.x | Brand markers in some export paths | Chinese-platform re-shares almost always arrive stripped |
| Pika 2 | Markers vary by plan and export route | Watermark-free tiers often ship clean files |
| Luma Dream Machine | Container metadata on direct downloads | Re-encodes erase it like everything else |
| Hailuo (MiniMax) | Occasional brand strings in the byte stream | When present, that’s a confession; absence proves nothing |
| Avatar & newer tools (HeyGen-style, Wan, Seedance…) | C2PA or encoder trail when present; brand markers vary widely | We read whatever the container shows — unknown tools with clean exports come back “inconclusive” |
| Any generator, after TikTok / YouTube / WhatsApp | Usually nothing — platforms re-encode and strip metadata | Trace the earliest source file you can find and check that instead |
Who runs this check
Newsrooms & fact-checkers
A clip lands before the event is confirmed. Container evidence plus sampled frames give you something citable in minutes, not a black-box score.
Trust & safety teams
Triaging user reports of synthetic video — the encoder trail sorts “re-encoded platform copy” from “fresh generator export” instantly.
Comms & executive protection
A fake clip of your CEO “announcing” something moves markets faster than a denial. Check the file, then respond with evidence.
Families & group chats
The WhatsApp forward that looks almost real. Ten seconds of container reading beats forwarding it to fifty people.
Researchers & educators
Teaching media literacy with real evidence — show students what a C2PA manifest and a stripped container actually look like.
Red flags to check in the sampled frames
We pull three frames precisely so a human can look. These are the details that still betray generators most often:
Hands, teeth and jewelry
Finger counts, earring asymmetry and teeth boundaries still glitch across frames — especially between our 12% and 88% samples.
Text, logos and signage
Background lettering that warps, misspells itself, or changes between frames is a classic generator artifact.
Physics: shadows, reflections, liquids
Light sources that disagree with shadows, mirrors showing the wrong scene, liquid that moves unnaturally.
Continuity across the three frames
Clothing details, room layout and object positions should evolve consistently. Discontinuities between distant frames are cheap to spot and hard for generators to prevent.
Visual tells fade with every model generation — treat them as supporting evidence next to the container reading, not as proof.
What a verdict looks like
“Detected” — the container confessed
A generator marker (Sora, Runway, Kling…) or a C2PA manifest was found in the file. The exact bytes and fields are listed for you to verify.
“Inconclusive” — the trail is clean or stripped
The honest verdict for almost anything downloaded from a social platform. It means no evidence survived, not that the clip is real.
Your own eyes, assisted
Three sampled frames plus the red-flag checklist above — because on video, human review of physics and continuity still beats every pixel classifier in the wild.
Why we refuse to fake a percentage
A 2025 peer-reviewed comparison measured a leading commercial detector at 44.51% accuracy on AI video — a coin flip — even while it scored 94% on still images. Frame-level video detection is an arms race the defense is currently losing, and a fabricated confidence score is worse than no answer. That’s also why the industry is moving to provenance: Google now verifies its own AI video via SynthID in the Gemini app, and OpenAI ships C2PA in Sora downloads. Reading those marks is exactly what this check does.
Container forensics, in plain words
The terms this page keeps using — defined once:
Container & atoms
An MP4/MOV file is a container of “atoms” — structured boxes holding the video plus bookkeeping like which software wrote it (the “©too” atom). Generators and editors leave their handwriting here.
Encoder tag (Lavf/FFmpeg)
The signature of the software that last wrote the file. A fresh FFmpeg tag with zero camera metadata means the file was rebuilt — whatever history existed before is gone.
C2PA Content Credentials
A cryptographically signed provenance manifest. OpenAI attaches it to Sora downloads; the EU AI Act makes machine-readable marking mandatory for AI media from August 2026.
SynthID
Google’s invisible watermark, used on Veo output. Only Google’s own verifier reads it — we report the open evidence the container itself shows.
Re-encode / transcode
What every platform does to every upload: decode and re-compress. It shrinks files — and incidentally destroys metadata, markers and manifests. The single biggest reason honest verdicts are “inconclusive”.
Frame sampling
Pulling stills at 12%, 50% and 88% of the timeline. Cheap, local, and surprisingly powerful: continuity errors between distant frames survive re-encoding better than any metadata does.
On a live video call right now?
This tool checks files, not live streams — and real-time face swaps on calls are exactly how the Arup fraud worked. If you’re on a call that feels wrong: ask the person to turn their head fully sideways, pass a hand in front of their face, or stand up — real-time swaps still glitch on occlusion and profile views. Then verify the request through a channel you already trust, not one provided in the call. Record the call if policy allows; the recording becomes a file you can check here afterwards.
Got it from TikTok, YouTube or WhatsApp?
Every major platform re-encodes uploads, and WhatsApp forwards are compressed again on every hop — camera metadata, generator markers and C2PA manifests rarely survive. An “inconclusive” verdict on a platform-laundered clip is the tool being honest, not broken. The sampled frames still help: look at hands, jewelry, text in the background, and whether physics stays consistent across the three frames. And for anything that matters, procedure beats detection: the deepfake video call that cost Arup $25M in Hong Kong succeeded on urgency, while a 2024 attempt to impersonate Ferrari’s CEO by phone failed the moment an assistant asked a personal question only the real CEO could answer. Call back through a known channel.
Honest limits
What it can’t tell you.
Every major platform re-encodes uploads, erasing container evidence — a clip that survived a platform comes back “inconclusive” no matter how it was made. Frame-level detection is an arms race: 2024–26 benchmarks measured accuracy collapsing to near chance on in-the-wild video.
Common questions.
How can I check if a video is AI-generated for free?+
Use the checker above — free, in-browser, no upload. It scans the container for generator markers (Sora 2, Veo 3, Runway Gen-4, Kling, Pika), verifies C2PA provenance, and samples three frames for your own review. For platform-laundered clips expect “inconclusive” — that honesty applies to every deepfake video detector, ours included.
How accurate are AI video detectors in 2026?+
Independent testing puts even leading commercial tools near coin-flip on AI-generated video (44.51% in one 2025 peer-reviewed comparison), far below their still-image scores. That’s why this check reads container evidence and provenance marks instead of pretending to a percentage.
Can you detect Sora 2 or Veo 3 videos?+
Fresh exports often carry evidence — Sora downloads include C2PA Content Credentials, Veo output carries SynthID — and we read what the file shows. Copies that went through a platform or screen recording usually lose it, and the honest verdict becomes “inconclusive”.
Can you detect a deepfake face swap?+
Not from pixels, honestly — in-the-wild benchmarks show specialized detectors barely beat chance on fresh generators. We read the container’s history and give you frames to inspect.
The clip came from TikTok. Why is everything “inconclusive”?+
Platforms strip and re-encode every upload, destroying camera metadata and AI markers — the same applies to YouTube, Instagram and WhatsApp forwards. Trace the earliest source you can find and check that file instead.
Do AI videos always have a watermark?+
No. C2PA and SynthID adoption is growing — and EU AI Act marking obligations apply from August 2026 — but marks can be stripped by re-encoding, and plenty of generators still ship clean files. Absence of a watermark proves nothing in either direction.
Don’t TikTok and YouTube label AI videos automatically?+
They try — both platforms roll out AI labels built on creator self-disclosure plus metadata reading (TikTok uses C2PA). Coverage is inconsistent: stripped files sail through, and mislabeled edge cases go both ways. Platform labels are a useful hint, not a verdict — checking the actual file you received is still on you.
Can I paste a YouTube or TikTok link?+
No — by design. URL-paste tools have to fetch and process the video on their servers, which is exactly the upload we refuse to do. Download the clip (or obtain the original file) and drop it here; nothing leaves your browser. Bonus: the downloaded file is also the honest thing to test, since it’s the bytes you actually received.
Why is there no API?+
The check is browser-first by design — clips never touch a server. Batch checks and API access exist in the Pro plan for teams that need volume.
Does it work on a phone?+
Yes — the container scan and frame sampling use the mobile browser’s own decoder, so a clip saved from WhatsApp or your camera roll checks the same way it would on desktop. Nothing leaves the device.
Is the video uploaded anywhere?+
No. The byte scan and frame capture use your browser’s own decoder.
Sources & further reading
- Comparative evaluation of commercial deepfake detectors (PMC, 2025) — 44.51% video accuracy finding
- Google: verifying AI-generated video with SynthID in the Gemini app (2026)
- C2PA — Coalition for Content Provenance and Authenticity
- EU AI Act, Article 50 — transparency obligations (applicable 2 Aug 2026)
- Reuters — Ferrari executive deepfake voice-call attempt foiled by a verification question (2024)