"This isn't even him speaking, it's his AI avatar. There's a cut scene after nearly every sentence, and his intonations sound like AI." That's the most-liked recent comment (16 likes) under a Nate Herk video about AI tools, and it shows where viewers are: watching a talking head and wondering whether the person was generated. Under a Dan Martell video, another commenter writes that they don't think the host records his videos anymore and suspects an AI model (13 likes). We can't verify either accusation, and this post won't referee them. Two comments don't prove a trend, but they show the situation: viewers deciding by eye. The research below suggests how far eye-level guessing is from reliable, and the checks after that use better evidence than the face.
What an AI avatar is
An AI avatar is a computer-generated presenter: a face, a voice, and lip movement produced from a script instead of a camera. The products come in two flavors. A stock avatar is a pre-made presenter licensed from an avatar platform. A cloned avatar is built from footage of a person, sometimes the customer themselves.
The buyers are businesses: training videos, product explainers, and localized versions of the same video in dozens of languages. HeyGen offers a free plan (3 videos a month, no credit card, per its pricing page) with a paid Creator plan at $29 a month. General video models generate the same output shape: Google's Veo 3, announced May 20, 2025 per CNBC's report, generates clips with dialogue and sound included.
The same tools can produce a company's training video or a fictional "expert" recommending supplements. What a viewer gets in both cases can look similar, which is why the question moved from "could a video be generated?" to "was this one?"
The signs that still work
Each sign below is a hint, not proof. Generated video has improved quickly, and last year's visible artifact may be patched in the next model. Phones and compression also make human faces look strange. Treat the signs as reasons to check further, not as answers.
Hands and teeth. Hands can reveal inconsistencies: fingers that merge or change count between shots, jewelry that moves, grips that don't match the object. Teeth can change shape or alignment between nearby shots.
Sentence rhythm. The anchor comment's "his intonations sound like AI" describes a pattern that shows up in some generated clips: level energy from the first word to the last, with similar pause lengths between sentences. Human speakers speed up, trail off, and breathe audibly, though some presenters are coached into near-uniform delivery too. A voice with unusually uniform pacing is worth a closer listen, not an answer by itself.
Cuts that hide the mouth. Editors cut for pacing all the time, and there's nothing suspicious about a cutaway. The pattern worth rewinding for is specific: a cut landing mid-sentence, repeatedly, exactly where the lip movement would be visible.
Backgrounds and physics. Compare fixed background edges and objects across nearby frames. Edges that shimmer, objects that shift position, or a scene that changes without an editing reason are clues worth a source check.
What the research found
A University of Florida research team (Ebner, Pehlivanoglu, Zhu, and Cahill) ran two studies comparing human and machine detection. In the first, more than 2,200 participants and two machine-learning models evaluated 200 facial images. Humans performed at chance, about 50 percent, while the algorithms reached 79 and 97 percent. In the second, nearly 1,900 participants watched 70 short videos. The pattern reversed: humans identified AI-generated video about 63 percent of the time, while the tested models performed near chance.
On video, the viewers beat the software in that test. A plausible reason: motion gives viewers information that a still image doesn't. The other takeaway matters as much: 63 percent means people were wrong about a third of the time. Confident is not the same as correct.
The label is the first check
YouTube requires creators to disclose AI-generated or altered content that seems realistic. Creators set the 'AI use' option in YouTube Studio at upload, and the video then carries an "AI generated or altered" style label for viewers. Per YouTube's disclosure page, the requirement covers content that makes a person appear to say or do something they didn't, alters footage of a real event or place, or generates a realistic scene that didn't occur. Minor edits (beauty filters, lighting, scripts, thumbnails) don't require disclosure.
The label is the fastest first check, but it depends on creators following the policy. In January 2026, the BBC reported on weight-loss videos circulating on Facebook with AI-generated "doctors" who, per the report, don't work at the hospital they appear to represent. A label is a strong positive signal when it's there. Its absence is not proof either way.
Checks that beat watching
Check the person's public history. A presenter with little verifiable history gives you less to check: search the name, look for the organization, colleagues, and posts from previous years. The BBC report says the "doctors" in those videos don't work at the hospital they appear to represent, so the history check stops at step one.
Treat the face as unverified even when it looks human. Full Fact documented deepfakes of well-known UK doctors spreading menopause misinformation through a network called Wellness Nest, per The Guardian's December 2025 report. The doctors named said they had not made or endorsed the videos. A human-looking face doesn't establish who is speaking or whether they said the words.
Check the claim, not just the clip. A generated video with a false claim and a genuine video with a false claim fail the same check. We covered the four moves in How to Fact-Check AI Output: A Practical Workflow: name the claims, find the source, search the claim itself, drop what you can't confirm.
Match the stakes to the effort. For entertainment, watching is enough, and a wrong guess costs nothing. When the video asks for money, health decisions, or trust in an identity, run the trail checks first. Our low-stakes and high-stakes split in Is ChatGPT Safe to Use? What Beginners Need to Know applies to video the same way it applies to typing.
FAQ
What is an AI avatar?
An AI avatar is a computer-generated presenter built from a script: a face, a voice, and lip-synced movement rendered by software instead of recorded by a camera. It can be a stock presenter licensed from a platform or a clone made from footage of a specific person. Channels use them to produce videos without a camera, a crew, or the presenter's schedule, which is why the format shows up everywhere from corporate training pages to anonymous advice channels.
What's the difference between an AI avatar and a deepfake?
A deepfake is video or audio that puts a specific person's likeness on something they didn't say or do. An AI avatar is the presenter technology itself, usually built with the subject's involvement or with a stock face. The overlap: cloning someone's likeness without their consent turns an avatar into a deepfake.
Can software detect AI-generated video?
In the University of Florida video study, the machine-learning models tested performed near chance while human viewers reached about 63 percent. Detector performance tends to drop when a new generator ships, because each model produces its own kind of artifact. Treat any single detector score with caution, and don't outsource the whole check to one tool.
Do I have to disclose AI video on YouTube?
If you post there, the 'AI use' setting in YouTube Studio covers it: YouTube requires disclosure for realistic AI-generated or AI-altered content, and the video carries a label for viewers. Beauty filters, lighting fixes, AI-assisted scripts, and thumbnails don't need disclosure. Platforms other than YouTube set their own rules, so check the current policy where you post.
Is it illegal to post an AI video without labeling it?
It depends on the conduct and the jurisdiction. A synthetic video used for fraud or impersonation may violate existing laws, and in the EU the AI Act's transparency rules (Article 50, applying from August 2, 2026) require disclosing deepfakes to viewers, per the European Commission's guidance. Platform policy is the check a viewer can run today, which is why the label check works better than a legal one.
Want to go further?
The same source-first method works on text, images, and demos, not just video: Is the AI Doing It, or Is a Tool Doing It? How to Read an AI Demo covers what a demo's result can and can't be attributed to. For the next video you question, run the public-history check before trusting the face.