AI faces: 300-portrait study measures why they look "doll-like"

The study's starting point is a familiar observation: current image models no longer produce obvious breakage like mangled ears or odd pupils, yet their portraits still read as "AI" at a glance. The authors, working with an AI agent named Muse, call this quality "doll-like" and set out to turn it into measurable geometry rather than impression. The sample was small by design — 100 portraits per group, all Japanese women seen from the front — which keeps the comparison controlled but limits how far the findings stretch.

The most interesting structural finding concerns how the two headline properties relate. Symmetry and averageness turned out not to be independent traits. Averaging the real photos cut their asymmetry to 28% of the individual average, and within every group, faces closer to the average shape were also more symmetric (r ≈ 0.57). The authors read this as evidence that generative models, by landing near the center of the learned distribution, inherit symmetry as a side effect — not as a separately optimized feature.

That reframes the residual unease. Real faces carry small, randomly directed left-right differences from development, which the authors describe as a signature of being a specific individual. The study's numbers suggest AI faces are thin on exactly that signature, which would make the uncanny quality a kind of absence rather than a defect. The caveats are real, though: the real-photo group was mostly red-carpet event photography with different lighting, makeup and expressions, and the authors note that swapping the control group already flipped the texture conclusion between two rounds of measurement.