You are the reference image

YouTube launched photorealistic AI avatars for Shorts this week. The feature uses Google's Veo to generate video clips of you, doing things you never did, in places you never went, saying things you never said. All from a selfie and a voice recording.

The setup takes a minute. You record your face reading a few prompts. You speak into the microphone so the model captures your voice. YouTube builds a photorealistic avatar. Then you type a prompt: "me explaining quantum physics on a space station" or "me hiking through a rainforest" or whatever the algorithm is rewarding this week. The model generates an 8-second clip of a person who looks and sounds like you, performing the scene you described. Multiple clips can be strung together. You only do the selfie once.

Two billion potential YouTube users just became their own reference images.

The face you uploaded

This series wrote about Val Kilmer's face being used as a reference image for a film he never shot, processed by a model that has no concept of mortality. The meaningful distinctions between tribute and exploitation lived entirely outside the model. The tool did not know what it was pointed at.

YouTube's avatar feature is the same workflow with the consent problem solved and a new one opened. You uploaded your face voluntarily. The permission chain is clean. The model still does not know what it is pointed at. But now you are pointing it at yourself, and the prompt that follows is an empty text box that YouTube fills with "describe a scene."

Describe a scene. The same casualization that this series has documented from every angle. The structured cinematographic vocabulary does not appear anywhere in the interface. The avatar is generated at 720p for up to 8 seconds with whatever defaults Veo applies. The lighting is whatever the model decides. The framing is whatever the model decides. The "performance" is a statistical average of human-shaped movement derived from training data, wearing your face like a costume.

The face works. The performance does not. That was the finding several articles ago, and nothing has changed. A photorealistic version of your face mounted on a body doing things no human director asked it to do. The model heard "hiking through a rainforest" and produced something that looks vaguely like hiking in a place that looks vaguely like a rainforest, performed by a body that moves like a stock-footage extra. Your face is the most specific input in the entire pipeline. Everything else is defaults.

Four products, one model, four vocabularies

Veo 3.1 now lives in four Google products simultaneously. Google Flow gives filmmakers structured controls for scene composition, camera movement, and iterative refinement. Google Vids gives Workspace users directable avatars for quarterly presentations at 720p with a ten-generation monthly ration. Gemini gives chatbot users a conversational prompt interface. YouTube Shorts gives creators a selfie button and an empty text box.

Same weights. Same architecture. Same training data. Four interfaces. Four levels of creative control. Four radically different relationships to the output.

Flow is the only one that approaches the model as a filmmaking tool. Vids approaches it as a productivity shortcut. Gemini approaches it as a chat convenience. YouTube Shorts approaches it as a selfie filter that learned to hallucinate.

The model does not deteriorate across interfaces. The vocabulary does. Each product strips another layer of intentionality from the interaction. By the time the same model reaches a YouTube creator, the interface has removed every mechanism for expressing a specific creative decision except the text box. And the text box says "describe a scene."

The safety slide

YouTube said the feature gives users "an easier way to include themselves safely and securely in videos." The SynthID watermarks are there. The C2PA metadata is there. The disclosure labels are there. The avatar cannot be used by anyone else. Auto-deletion after three years of inactivity. This is thorough for the consenting creator.

The question the safeguards do not address is whether a photorealistic generative avatar of a real person, performing actions prompted by a text box, looks different from a photorealistic deepfake of the same person performing the same actions. One was consensual. One was not. The output is visually identical. The watermark is metadata that survives casual viewing the way nutrition labels survive a bag of chips: technically present, functionally invisible.

Europe is excluded from the rollout entirely. The regulatory geography continues to track the legal boundary. CapCut launched Seedance in Southeast Asia and Latin America, not the US or Europe. YouTube launches avatars globally except Europe. The markets with the strongest personal data protections are the markets where the feature does not ship. The legal boundary and the access boundary follow the same line. Again.

The input changed

Every prior article in this series assumed the filmmaker stood behind the camera. The creative decisions about what to shoot, how to light it, where to place the subject all addressed the world in front of the lens. The filmmaker was the author. The output was the work.

YouTube's avatar feature rearranges this. The filmmaker is the subject. The creative decisions about their own face, body, and voice have been delegated to a model whose defaults were documented across this entire series: beauty bias, center framing, safe lighting, convergent aesthetics. The model is now applying all of those defaults to you.

When a filmmaker generates a landscape with Veo, the beauty bias produces a landscape that looks better than what was described. When a creator generates themselves with Veo, the beauty bias produces a version of them that looks... different. Smoother. Better lit. The optimization target (perceptual quality, visual appeal) is now pointed at a real person's face. Whether that is flattering or uncanny depends entirely on how close the statistical average lands to the person who uploaded the selfie.

The vocabulary still applies. A structured prompt that specifies lighting, environment, camera angle, and color palette will produce a more intentional avatar clip than "me in a coffee shop." The model still responds to precision. The interface still does not encourage it. The gap between what the tool can do and what the interface invites you to say is wider here than anywhere else, because the subject of the generation is the person holding the phone, and they do not know they are allowed to ask for more.

This is the furthest the absorption trajectory has traveled. Standalone tool to chatbot to editing timeline to agent to productivity app to selfie button. Each step shrank the prompt field. This one enlarged the person. The input is your face. The prompt is an afterthought. The output carries your likeness and the model's taste in equal measure.

The vocabulary does not care whose face is in the frame. It works the same way it always has. Precision in, precision out. Vagueness in, defaults out. The defaults are just wearing your face now.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never uploaded a selfie to a generation model and suspects this positions him firmly in the minority.