How to Tell If a Video Is AI Generated: 7 Checks That Work

Treat it as a case, not a quiz. Each check below produces a signal that is weak on its own and useful in combination; a clip that fails three of them is probably synthetic, a clip that fails one is probably compressed. Guides on how to tell if a video is AI generated that promise one giveaway (six fingers, no blinking) are describing models from several generations ago.

• Continuity: scrub around cuts and fast motion for identity, object and background drift.
• Hands, text and physics: the three things current models still fumble under stress.
• Audio: cadence, breaths, room tone and lip timing, listened to without the picture.
• Channel and account signals: upload cadence, disclosure labels, description and history.
• Detector tools: a probability from more than one model, never a verdict from one.
• Provenance: C2PA Content Credentials, SynthID and platform labels, where they exist.
• Context: earliest upload, independent coverage, and whether the claim even needs the video to be fake.

Why build your own real-or-AI reference set?

The fastest way to calibrate your eye is to make a synthetic clip that mimics footage you shot yourself. Record ten seconds on a phone (a hand picking up a labelled bottle, a person turning to speak), then prompt the same scene in the LazyKiwi AI video generator and play the two side by side. The Kling v3 Omni page lists 3 to 60 second clips at 720p or 1080p from a text prompt or a single image, which is enough range to match most phone footage; the 5-second 720p render was priced at 829 credits on the button when we captured it. The live catalog (2026-09-05) lists ten video models, Kling v3 Omni, Veo 3.1 Lite, Hailuo, MiniMax H3, Seedance 2.5, Wan 2.7, Happy Horse, Grok Imagine Video, Hunyuan and LTX, so a reference pair can be rendered on the same family the suspect clip most resembles.

If you would rather test your eye on someone else's set, the MIT Media Lab built a public site for exactly that: its Detect Fakes project lets you view manipulated and unaltered videos and see whether you can tell them apart, per MIT Media Lab (2026). Do that once before you judge a viral clip; most people are less accurate than they expect.

Single frames lie in both directions: a generated frame can be flawless and a real frame can be a smear of motion blur. What generators still struggle with is keeping the world consistent across time, so the useful signs are the ones you can only see by scrubbing.

Which frames should you scrub?

• The frames on either side of every cut: identity drift hides at edit points.
• Any fast gesture: step through at quarter speed and count fingers on the way in and the way out.
• Every occlusion: a hand crossing a face, a person passing behind a pole; things change while hidden.
• Any pass over text: a sign, a shirt print, a phone screen.
• The last second: many generators lose coherence at the end of a clip, which is why creators trim it.

Close your eyes first. Cloned or synthesised speech tends to keep an even pace with breaths in the same places, flatten emotional peaks, and sit in a room tone that never changes when the speaker moves. Real speech speeds up, trips, and picks up the room: a door, a chair, a change of reverb when the head turns.

Lip timing is only evidence when the audio and picture are actually aligned. Screen recordings, wireless earbuds and platform transcodes introduce a constant offset on real footage, so a fixed lag proves nothing. What you are looking for is the offset that changes, or specific mouth shapes that never match their sounds: a closed-lip 'm' on an open mouth, a 'th' with no tongue, repeated across the clip.

A useful reality check: several models in the LazyKiwi workbench render picture only. The Veo 3.1 Lite page lists 16:9 or 9:16 output at 720p or 1080p, durations from 4 to 60 seconds and a 1,000-character prompt, and audio is not part of this tier here. Any voice on a clip made this way was laid over it in an editor, and the room tone will belong to a different room than the one in frame. That mismatch is one of the more reliable audio tells.

• Listen once with the picture off and mark where the pace or breathing sounds metronomic.
• Compare the voice with a verified recording of the same person; pronunciation habits are hard to clone.
• Follow one physical event (a cup set down, a door) and check that the sound arrives with the image.
• Note whether the room tone shifts when the camera or speaker moves; a static bed under a moving shot is a flag.
• Look for lip shapes that mismatch the same sound repeatedly, not a constant lag.

YouTube gives you one signal the clip itself cannot: a disclosure. Its policy requires creators to disclose when they use AI to meaningfully alter or generate photorealistic content, with three named cases (a real person made to say or do something, altered footage of a real event or place, a realistic scene that did not occur), and the content is then labelled as AI generated or altered for viewers, per YouTube Help (2026). The absence of a label proves nothing, because the disclosure is set by the uploader.

How to tell if a YouTube video is AI generated in one minute

• Open the description and expand it: look for the altered-or-synthetic label, a 'how this content was made' note, tool credits or source links.
• Watch the first cut and the last second at quarter speed with the checks from the visual section.
• Mute it, then unmute it with your eyes closed; a stock-sounding narration over silent B-roll is the common pattern.
• Search a distinctive line from the title or transcript; an earlier upload elsewhere changes the question from 'is it AI' to 'is it stolen'.

How to tell if a YouTube channel is AI generated

Channel-level signals are about cadence and sameness. Several uploads a day with identical length, the same synthetic voice across unrelated topics, no on-camera person ever, thumbnails that share one rendering style, and a description with no location or name are the pattern of an automated channel. None of that is proof of generated video, and some legitimate faceless channels match most of it, so read it as a reason to run the other checks, not a conclusion. The same steps for how to tell if a video is AI generated apply to the channel's most-viewed upload, which is the one that earned the traffic.

Other platforms label differently. TikTok's policy requires people to label AI-generated content that contains realistic images, audio or video, and the company said it would test automatic labelling, per TikTok Newsroom (2023). Meta applies 'AI info' labels to video, audio and images when it detects industry-standard AI indicators or when people disclose, per Meta Newsroom (2024).

Detectors fall into three kinds: classifiers that estimate a probability from pixels and audio, provenance readers that look for a signed manifest or a watermark, and toolkits that help you do the manual checks faster. The table summarises what each vendor says on its own site as of 2026-09-06; none of these rows is a test result of ours.

Hive (2026) positions its detection as APIs for platforms with a public demo. Sensity (2026) and Reality Defender (2026) sell to enterprises, with Reality Defender advertising a free tier limited to audio and image scans. TrueMedia (2026) is the non-commercial option, and its promise of transparent accuracy metrics and uncertainty is the behaviour to want from any detector. InVID (2026) is not a detector at all; it pulls keyframes so you can reverse-search them.

Read an AI detector video score as a probability from one model trained on one distribution. It drops on compressed, filtered or upscaled footage, it cannot see a clip that mixes real and generated elements as two things, and a watermark check such as SynthID (2026) only finds marks that Google's own products embedded. Two detectors that disagree are telling you the truth: the clip is inconclusive on pixels alone.

Provenance answers a different question from detection: not 'does this look generated' but 'who made this, with what, and what changed since'. The Coalition for Content Provenance and Authenticity publishes an open technical standard for exactly that record, branded Content Credentials, per C2PA (2026). When a camera, an editor or a generator signs the file, the record travels with it and a verifier can read it.

1. 1

   Get the original file, not a screen recording

   A re-encode or a screenshot strips the manifest. Ask for the download, or use the platform's original-quality option where it exists.

2. 2

   Drop it into Verify

   The Content Credentials Verify tool (2026) inspects the file's credentials and how it has changed over time; its upload box lists video and audio formats including MP4, MOV, AVI, M4A, MP3 and WAV alongside images and PDF.

3. 3

   Read the manifest for the three facts

   Who signed it, which tool produced or edited it, and what actions were recorded. A generator's name in the tool field ends the argument; an editor's name with a 'placed' action means a composite.

4. 4

   Treat a missing manifest as silence

   The Verify page itself warns that Content Credentials are still rolling out and the file you inspect may have no information. No manifest is not evidence of a fake; a stripped manifest is not evidence of one either.

5. 5

   Cross-check the platform label

   Meta says it adds AI info labels when it detects industry-standard indicators, which is what a C2PA manifest is; YouTube and TikTok rely on uploader disclosure. A label plus a manifest is strong; a label alone is a claim.

Most wrong calls we have seen come from applying one check as if it were the whole method. These are the recurring ones, with the correction.

• Calling compression a tell: reuploads, low light and beauty filters produce mushy hands and drifting edges on real footage. A sign counts when it repeats and connects to another sign.
• Trusting a single detector: one score from one model is a probability, not a verdict. Run two, and if they disagree, say so.
• Ignoring context: the earliest upload and independent coverage decide more cases than pixels do. A real clip with a false caption is the more common fraud.
• Mistaking enhancement for generation: footage that went through an upscaler such as the AI video enhancer picks up smoothing and interpolated frames that classifiers flag. Ask whether the clip was processed before you ask whether it was generated.
• Reading a missing manifest as guilt: credentials are still rolling out, and most authentic phone video has none.
• Skipping the audio pass: a silent-model render with a voice laid over it fails the room-tone check even when the picture is clean.
• Forgetting the filters: in LazyKiwi's workbench data (August 2026) the content-safety filter was the most common failure reason for external image jobs, with 49 blocks in 30 days. Mainstream tools refuse a whole class of prompts before rendering, so a realistic clip of that class points to a tool without such filters, which narrows the provenance question rather than answering it.

How should you report what you found?

State the evidence, not the vibe. 'Likely synthetic: identity drift at two cuts, room tone static under a moving shot, no earlier upload found, two detectors above their thresholds' can be checked by someone else. Keep the URL, upload date, caption and a few frame grabs, because posts change after they get attention. Knowing how to tell if a video is AI generated is only half the skill; writing down why in a way that survives a challenge is the other half.