Why Do AI‑Generated People All Look the Same?

When we ask any AI image‑generation tool to create a "pretty girl"—high nose bridge, big eyes, small V‑shaped face—different prompts and scenes still yield the same facial features, revealing that AI struggles to imagine diverse human beauty.

1. The Birth of an “Average Face”

In 1878 Francis Galton exposed a batch of criminal portraits on a single plate, hoping to derive a typical criminal visage. Instead, the composite was unexpectedly symmetrical and more attractive than any original. He repeated the experiment with various groups and concluded that averaging many faces usually produces a more pleasing result.

2. Is Average Equal to Beauty?

Psychologists Langlois and Roggman (1990) proposed the “attractiveness‑average hypothesis”: the closer a face is to the population mean, the higher its perceived attractiveness. Later studies (Rhodes & Tremewan, 1996) showed the relationship follows an inverted‑U curve—moderate deviation from the mean yields the peak of attraction, while extreme deviation (distortion) or perfect averageness (blandness) are less appealing.

3. Why AI Converges on the Average

Modern AI generators have seen millions of “good‑looking” faces. Their objective—"as many people as possible find the result beautiful"—leads the model to compute a statistical average of the "beauty" dimension. The result is a highly symmetrical face where individual quirks cancel out.

4. The Practical Upsides of an Average Face

Copyright & portrait risk avoidance: A face that resembles no real person sidesteps celebrity likeness claims and deep‑fake liability.

Content safety & compliance: Neutral expressions and non‑offensive features pass advertising and platform audits with minimal friction.

Maximum aesthetic safety: The average face scores high on general acceptance, delivering stable conversion rates across diverse audiences.

5. The Hidden Costs

Aesthetic fatigue: Repeated exposure to the same bland visage leads viewers from curiosity to disgust, as documented by online comments.

Uncanny valley effect: The face looks almost human but lacks pores, micro‑expressions, and subtle asymmetries, triggering a sense of eeriness described as “broken doll” or “creepy.”

6. Steering Away from the Average

To produce distinctive characters, prompt engineers can deliberately inject “imperfection” cues: freckles, moles, skin texture, pores, uneven skin tone, asymmetrical features, tired eyes, or crooked teeth. Even a single freckle can shift the generated face from a generic template to a seemingly real individual.

7. Conclusion

The “average face” solves legal and safety constraints but sacrifices story, personality, and the subtle charm that makes a face memorable. Future AI should aim not merely for a statistically perfect average but for controllable, nuanced representations that capture genuine human appeal.