My AI Avatar Looks Exactly Like Me and It’s Unsettling

The Moment I Met Myself

I did not expect to feel anything. I had been writing about generative AI for over a year, testing image generators and language models on a near-daily basis. I understood how diffusion models work. I knew about GANs and neural radiance fields and all the technical machinery behind digital face synthesis. I was, I thought, thoroughly inoculated against surprise.

Then my AI avatar loaded on screen, and my stomach dropped.

It was not the cartoonish approximation I had braced for. It was me. The slight asymmetry of my jaw. The way my left eyebrow sits fractionally higher than the right. The specific distribution of stubble that only exists on my face because I consistently miss the same spot shaving. These were not generic “male, 30s, dark hair” features. They were my features, rendered with a fidelity that felt less like technology and more like surveillance.

I showed it to my wife without explanation. She looked at it for three seconds. “When did you take this?” she asked. Not “what is this?” or “who is that?” She assumed it was a photograph. When I told her an AI had generated it from twelve reference photos, she looked at me, then back at the screen, then at me again. “That’s creepy,” she said. I agreed, though I was not entirely sure why.

This is the story of spending a week generating, comparing, and thinking about AI avatars of myself — and arriving at conclusions that have less to do with technology than with what it means to have a face.

Three Platforms, Three Uncanny Valleys

I tested three of the leading AI avatar platforms in their current 2026 iterations: HeyGen Avatar IV, Synthesia Express Avatars, and Microsoft VASA-2 (research preview). Each approaches the problem differently, and the results diverge in ways that are instructive.

The input was identical for all three: twelve photographs of my face taken in different lighting conditions, plus a 30-second video of me speaking naturally. Total preparation time: about fifteen minutes.

HeyGen Avatar IV produced the result that fooled my wife. Their system focuses on what they call “micro-expression fidelity” — the involuntary facial movements that happen between words, the way pupils dilate slightly during emphasis, the almost imperceptible head movements that accompany natural speech. The avatar blinks at irregular intervals, not on a timer. Its breathing is visible. When it pauses mid-sentence, it does what I do: a slight compression of the lips and a half-degree head tilt to the right.

Synthesia Express Avatars were more polished in some respects — smoother skin, better lighting consistency — but less uncanny. The result looked like a well-produced video of someone who resembles me closely, not a clone of me specifically. MIT Technology Review noted that Synthesia’s latest avatars are better at matching intonation to sentiment, and I can confirm this — the avatar sounded more emotionally natural than it looked. But a postdoctoral researcher studying deepfake faces described the same phenomenon I noticed: static features and occasional small, abrupt body jerks that reveal the artificial origin.

Microsoft VASA-2 was the most technically impressive in isolation. The lip sync was flawless. The emotional range was broader. But it had a quality I can only describe as “too perfect.” My actual face, on video, has micro-imperfections — a slightly uneven smile, an occasional tic near my left eye. VASA-2 smoothed these away, producing something that looked more like me than I look on camera. It was accurate to my geometry but unfaithful to my imperfections, which are, it turns out, a significant part of what makes a face recognizable.

Why “Almost Perfect” Feels Worse Than “Obviously Fake”

There is a well-documented psychological phenomenon called the uncanny valley, first described by robotics professor Masahiro Mori in 1970. The idea is simple: as a non-human entity becomes more humanlike, our emotional response grows increasingly positive — until a threshold where it becomes almost but not quite human, at which point affinity plummets into revulsion. Then, as it becomes indistinguishable from a real human, affinity climbs back up.

What I experienced with these avatars is something Mori’s original framework did not quite anticipate. The HeyGen avatar was not in the uncanny valley. It had crossed over to the other side. It was so close to real that my brain accepted it as real, and the discomfort came not from artificiality but from the existential implications of a perfect copy.

Think about what your face represents. It is the most personal identifier you possess. More than your name, more than your voice, your face is how the world recognizes you and how you recognize yourself. When someone else can generate a pixel-perfect replica of that face — one that moves, speaks, and expresses in your patterns — something fundamental about personal identity shifts.

I spent an afternoon watching my avatar deliver a script I had written. The words were mine. The face was mine. The voice was a synthesis trained on my recordings. But the person on screen had never existed. He was assembled from statistical patterns, a weighted average of my features optimized to minimize perceptual distance from reality. Watching him was like looking into a mirror that reflects not what you are, but what you could be reduced to.

The Practical Uses (and the Ones That Keep Me Up at Night)

To be clear, I am not against this technology. The practical applications are genuinely valuable.

Corporate training videos that previously required a studio, crew, and a full day of an executive’s time can now be produced in minutes. Localization that once meant hiring voice actors for each language now means typing a script and selecting a target language — HeyGen supports 175+ languages, Synthesia over 160. Accessibility benefits are real: AI avatars can deliver sign language, maintain eye contact with viewers, and present information in formats that accommodate different learning needs.

Content creators face a genuine efficiency revolution. A YouTuber who normally spends hours filming can generate consistent, well-lit talking-head segments from text. Educators can produce course material at a pace that was previously impossible. Customer service videos can be personalized at scale.

But then there are the applications that erode something important.

A family member could receive a video call from someone who looks and sounds exactly like me, asking for money. A political candidate’s face could be attached to words they never said, released 48 hours before an election when there is no time for debunking. A deceased person’s likeness could be used indefinitely, speaking new lines as a corporate spokesperson years after death. These are not theoretical risks. The EU AI Act and multiple U.S. state laws enacted in 2025 now regulate deepfake use, but legislation consistently lags capability.

The platforms themselves are implementing consent frameworks. Synthesia requires explicit consent verification before creating a custom avatar and prohibits impersonation. HeyGen’s terms of service include similar restrictions. But terms of service are speed bumps, not walls. The open-source model ecosystem means anyone with a GPU and determination can generate unconsented avatars without platform guardrails.

What This Means for the Rest of Us

I have been thinking about a specific question since I started this experiment: at what point does a copy of your face stop being a representation and start being a form of identity?

We already accept that photographs capture something real about us. Video, more so. An AI avatar that is visually indistinguishable from a real video occupies a strange ontological position. It is not a recording of something that happened. It is a fabrication that is perceptually indistinguishable from a recording. The philosophical distance between “this is a video of me” and “this looks exactly like a video of me but is not” is enormous, but the perceptual distance is zero.

This asymmetry — huge philosophical difference, zero perceptual difference — is what makes the technology genuinely unsettling. It is not that the avatar is bad. It is that the avatar is too good for our existing frameworks of trust, identity, and authenticity to handle.

I deleted my HeyGen avatar after finishing this article. Not because I think deleting it accomplishes much — the model that generated it still exists, and anyone with my reference photos could regenerate it. I deleted it because keeping a digital twin of myself felt like leaving a loaded weapon on the table. Probably harmless. Possibly not. Not worth finding out.

My advice, for whatever it is worth: try generating an AI avatar of yourself. Not because the technology is fun (though it is), but because seeing your own face rendered by a machine changes how you think about faces, identity, and the visual trust that holds social interactions together. That education is worth the $24 monthly subscription. What you do with the avatar afterward is a question each person will answer differently, and I suspect the range of answers will tell us a great deal about where we are headed.

Frequently Asked Questions