Investigation: AI tools rate photos of women as more sexually suggestive than photos of men, especially if nipples, pregnant bellies, or exercise are involved
Guardian exclusive: AI tools rate photos of women as more sexually suggestive than those of men, especially if nipples, pregnant bellies or exercise is involved Tweets: @womenreadwomen , @nyuniversity , @random_walker , @msmarymcgill , and @mmitchell_ai Tweets: @womenreadwomen : “Objectification of women seems deeply embedded in the system.” AI algorithms have been found to rate pictures of women as more “racy” or sexually suggestive than comparable pictures of men. https://www.theguardian.com/ ... @nyuniversity : .@nyu_journalism's @HilkeSchellmann and @gianlucahmd used AI tools to analyze hundreds of photos of people in underwear, working out, and using medical tests, and found that AI flags images of women in everyday situations as sexually suggestive: https://www.theguardian.com/ ... Arvind Narayanan / @random_walker : Fascinating audit of social media “raciness” classifiers that don't understand context and are massively biased toward labeling images of women's bodies as sexual. Posts classified racy get shadowbanned. By @gianlucahmd and @HilkeSchellmann. https://www.theguardian.com/ ... Dr Mary McGill / @msmarymcgill : In the 70s, Laura Mulvey theorised that classical Hollywood cinema's ‘look’ objectified women on screen. In the 2020s, AI algorithms are doing the work of the male gaze all over again. A timely read and a reminder that technology is never neutral. https://www.theguardian.com/ ... @mmitchell_ai : Advised on this. Demonstrates effect of women's bodies being objectified, laundered through “AI”, back into what we see(/don't see) online. “It's because of [Meg] [that] on the Guardian there's a gif of me bare-chested wearing a bra🙂” Great praise! 😍 https://www.theguardian.com/ ...
Context & Ripple Effects
The Guardian's test extends a documented pattern rather than opening a new front: back in November 2020, research already flagged gender bias in the image recognition services sold by Google, Microsoft, and Amazon, and the intervening years produced consumer-facing versions of the same skew — Lensa generating pornified avatars for its female users while men got astronauts, and Graphika counting millions of monthly visitors to apps built to undress women in photos.
First-order effects
- Women whose photos pass through these classifiers — fitness, pregnancy, breastfeeding content especially — face disproportionate flagging, removal, or reduced reach under platform rules keyed to 'racy' scores.
Second-order effects
- Vendors of image-classification APIs face pressure to audit and retrain, since the same skewed labels propagate when classifiers are used to curate training data for generative models — the pathway Bloomberg already traced in Stable Diffusion amplifying job-title and crime stereotypes.
Third-order effects
- If classification bias keeps compounding across moderation, dataset curation, and generation, algorithmic discrimination law written for hiring and credit will be pulled toward covering content systems — a gap the untested deepfake legislation tracked by the FT shows regulators have not yet closed.
The trend: Gendered judgments embedded in vision models are propagating through every layer that consumes them — moderation, dataset curation, and generation — making classifier bias a systemic input to how women are represented online.