The more an AI avatar mimics a human, the more unsettling it becomes

In November 2025 Coca-Cola released another AI-generated version of "Holidays Are Coming," the spot with the red trucks rolling through the snow. This time the cast was polar bears, a panda, and a sloth. The Verge noticed what was missing. The ad, its review said, "attempts to sidestep issues around generating realistic humans by instead featuring a cast of various critters."

Coca-Cola's own release has its head of generative AI, Pratik Thakar, boasting that the team "controlled every cinematic detail, from camera angles to physics-driven realism."

None of that rescued the ad. The Verge's headline called it a "sloppy eyesore," with "no consistent style, switching between attempted realism and a bug-eyed toony look," and animals that moved "like flat images that have been sloppily animated rather than rigged 3D models in CG." Coca-Cola had already taken a beating in 2024 for three AI holiday commercials, and it came back anyway.

What interests me is the decision to switch to animals. A year into the backlash, one of the biggest advertisers in the world (or their agency) decided that the best thing to do was stick with AI but render animals instead of humans this time. My guess at how that happened: an art director and a creative director agreed that animals would be easier to get right, the AI would still earn press as innovation, and the client bought the pitch. The miss was in the diagnosis. They read the backlash as "AI can't do people" and swapped the subject. But the valley opens wherever something gets close to real, whatever the subject. The animals were still pushed toward photoreal, and they landed in the same place.

The drop near human

Robotics people have a name for the reaction Coca-Cola was steering around that I'm sure you've heard before: the "uncanny valley." Make an artificial face more humanlike and people warm to it, up to a point. Get too close to human without arriving, and the warmth turns to eeriness. The problem with this ad is that the face doesn't have to be human. Any creature we know well enough to notice when it moves wrong can fall into the valley, polar bears included. Maybe that's anthropomorphism: we read human feelings into an animal's face, so we hold it to a human standard. Or maybe it's simpler, and we just know how a bear is supposed to move.

In 2019 a team led by Astrid Rosenthal-von der Pütten put people in a brain scanner and had them rate artificial agents and real people, then choose between them. The part of the brain that weighs how much we value something tracked likability "as a nonlinear function of humanlikeness, with selective low likability for highly humanlike agents." It's a small lab study. The pattern it found is the uncanny valley, and it showed up in brain activity as well as in what people said they felt.

That's the trap in the brief most agencies get. The client wants the avatar to feel relatable, so the team pushes it toward human: better skin, better eyes, better micro-expressions. Every step of that work costs money, and the last few steps buy the reaction nobody wanted.

What the brand research shows

The best direct test I've found comes with a serious caveat. Faezeh Golparvar and two colleagues showed 750 US adults mock Instagram ads for a fictional skincare brand. Some had a human influencer, some a hyper-realistic AI one with no disclosure, some a stylized, cartoon-like AI labeled as AI. On integrity, the human scored 5.32, the stylized AI 3.67 and the hyper-real one 2.05. Benevolence followed the same order. In the focus groups one participant put it plainly: "Her eyes are dead. It's unsettling."

The unlabeled hyper-real AI scored less than half the human's integrity rating. Style and disclosure changed together in this design, so the chart can't say which one did the damage.

The caveat: the stylized avatar was always the labeled one, and the hyper-real avatar was always the unlabeled one. The study can't say how much of the gap came from the cartoon and how much from the honesty. It's also from a small journal. I'd use it for direction and nothing finer.

A Dutch study points the same way from a different angle. Researchers paired 185 young women with a human or an AI creator for a month. The AI creator's photos ran through Snapchat's "Beauty Cartoon" lens and carried #virtualinfluencer tags, so nobody was fooled. The women bonded with both the same way, and the authors say that challenges "the assumption that the human-likeness of a media character is a decisive factor." I'm reading that one through MediaCat's trade coverage; I haven't seen the paper itself.

An agency survey from Billion Dollar Boy fills in the commercial side. 76% of consumers said they trust virtual influencers for product recommendations, and 57% said AI "digital twins" of real creators erode trust. The firm sells creator marketing, so treat the numbers as rough. They still split along the line you'd expect: people trust the obvious character and distrust the copy of a person.

The strongest evidence against my argument comes from Germany. Jasmin Baake and two colleagues showed each of 491 people a one-minute science video presented by an AI avatar, either highly stylized or highly realistic. The realistic ones were rated more trustworthy, most clearly on integrity and benevolence, and the authors wrote that the result was "contradicting the Uncanny Valley effect." None of the videos were labeled as AI.

That detail matters. Meixian Wang, Jaehwuen Jung and Ravi Bapna ran AI influencer ads on Meta to about 1.8 million people. Without a label, the more realistic influencer drew more engagement. With an AI label, it did significantly worse than the less realistic one, and a human beat them both. It's a working paper, and the field test measured link clicks and three-second video plays. Among Instagram users, whom the authors treat as more used to virtual influencers, the effect flipped. For everyone else, it puts the cost of realism where you'd expect: at the moment people find out.

Pick an end

Taken together, the research says a realistic AI face does fine until people learn it's AI, and then it does worse than one that never pretended. A face can be clearly a person, or clearly a made thing. The one that passes for a person until it's labeled gets judged against a human and loses.

That turns into a short set of design rules:

• If the job needs a human face, hire one. Healthcare, finance, anything where the audience is deciding whether to believe you. Golparvar's own advice is to keep AI out of high-trust contexts.
• If you use AI, stylize on purpose. Push it toward the drawn and the obviously built. A mascot can carry warmth and personality without asking anyone to believe it's a person.
• Label it in plain words. In Golparvar's study a clear "AI-generated virtual influencer" label raised trust, and the vague "Digital Creator" did nothing measurable.
• Stop paying for the last mile of realism. The budget that closes the gap to photoreal is the budget that buys the drop.

This cuts against how most avatar projects get sold. The pitch deck shows the most lifelike render, because lifelike reads as expensive and expensive reads as quality. Audiences read the same render as an attempt to pass.

Coca-Cola changed the subject when it needed to change the style. A polar bear rendered for "physics-driven realism" falls into the same valley a person does. An honest AI face is one nobody could mistake for the real thing.