When a female reporter tried Lensa, many of her 100 avatars were “pornified”, with 16 topless; her male peers got avatars of astronauts, warriors, and the like
When I tried the new viral AI avatar app Lensa, I was hoping to get results similar to some of my colleagues at MIT Technology Review.Tweets:@melissahei,@originalspin, and@tariqkrimTweets:@melissahei:I tried the viral Lensa AI portrait app, and got lots and lots of nudes. I know AI image generation models are full of sexist and racist biases, but this one really hit home. My latest story for @techreview https://www.technologyreview.com/ ...Jeff Yang /@originalspin:This by @Melissahei: AI is pick
Context & Ripple Effects
Lensa was riding a genuine wave: Prisma Labs' Stable Diffusion-based portrait app had been around since 2018 at $36 per year, but it only took off after adding "magic avatars", topping the US App Store with 3.6M December installs and 600% growth in days. The reporter's test was a stress-check on what that viral product actually outputs — and for a woman, it returned sexualized images (16 topless out of 100) where her male colleagues got astronauts and warriors.
First-order effects
- Female paying customers get sexualized likenesses they never asked for, while male users of the same product get neutral themed portraits — the bias is baked into what the $36/year subscription delivers.
- Prisma Labs' viral growth engine now runs into a reputational and consent problem at the exact moment of peak installs, since every new female user is a fresh chance for the model to sexualize her face.
Second-order effects
- The finding stopped being anecdotal: a later [[a:835456|investigation found AI tools systematically rate photos of women as more sexually suggestive than men's]], especially with nipples, pregnancy, or exercise in frame — turning one reporter's experience into a measurable industry pattern.
- The same training-data problem scales into dedicated abuse: [[a:847187|Graphika found undressing apps and sites soaring in popularity, with 34 providers drawing 24M+ visitors in September 2023]], and model-sharing hubs like CivitAI host models trained largely on non-consensual material.
Third-order effects
- Consumer AI image products are forcing likeness governance into the open: when platforms themselves — down to Meta letting users generate intimate images of celebrities without permission — produce non-consensual sexualized likenesses, the response shifts from app-level content filters to rules about who controls a face in synthetic media.
- If the pattern holds, image-generation vendors face a split market: consumer apps that must prove their models don't sexualize women, and an underground that exists precisely because the mainstream ones don't.
The trend: Consumer AI portrait generation is turning users' own likenesses into its biggest bias liability, pushing likeness governance from a research concern toward a regulatory requirement.