Investigation finds that Amazon's Rekognition frequently misgenders trans, queer, and nonbinary individuals by design, further alienating a marginalized group
In 2017, the Washington County Sheriff's Office, just outside Portland, Oregon, wanted to find a man covered in dollar bills. Tweets: @jezebel , @annamerlan , @esjayxx , @fittsofalexis , @jezebel , @jawnita , and @mercedes_allen Tweets: @jezebel : “What happens when you can't use a bathroom because an AI lock thinks that you shouldn't be there?” A look at how Amazon's powerful new facial recognition and analysis system, and others like it, pose a great danger to already vulnerable populations https://jezebel.com/... https://twitter.com/... Anna Merlan / @annamerlan : So @dmehro built a working model of Rekognition, Amazon's facial recognition and analysis software, and discovered it does some pretty disturbing shit with images of trans and nonbinary people. https://jezebel.com/... @esjayxx : This is hysterical. This article complains that a facial recognition system cannot identify ‘non binary’ or trans people. OF COURSE NOT. It it looking at faces. It can't mind read. It can only identify what you are, not what you THINK you are. https://jezebel.com/... Alexis Sobel Fitts / @fittsofalexis : “What happens when a cop looks at your license and your machine predicted gender doesn't match what they see?” A crucial investigation into Amazon's erasure of vulnerable populations, by @annamerlan and @dmehro https://jezebel.com/... @jezebel : Amazon's facial analysis program is building a dystopic future for trans and nonbinary people https://jezebel.com/... https://twitter.com/... J. Escobedo Shepherd / @jawnita : This feature by @annamerlan and @dmehro is important, terrifying, and speaks volumes about the limitations of the tech that affects the way we live our day to day lives, particularly for people who are marginalized https://jezebel.com/... @mercedes_allen : Building trans accommodation into facial recognition wouldn't seem a priority. Until you realize the consequences of not: “A world governed by these tools is one that erases entire populations. It's a world where individuals have to conform to be seen...” https://jezebel.com/...
Context & Ripple Effects
This investigation lands alongside a companion deep dive into Amazon's quiet build-out of surveillance infrastructure — together they sketch a company selling facial analysis to police (the Washington County Sheriff's Office outside Portland is the on-record customer) while the system's binary gender classifications fail on trans and nonbinary faces by design. Jezebel's framing goes beyond accuracy stats: a bathroom door or checkpoint keyed to a misgendering classifier becomes a physical barrier, not just a wrong label.
The piece also feeds a documented pattern rather than opening one — later coverage shows [[a:956870|researchers warning that AI race and ethnicity detection in market research risks encoding discrimination]], and Google quietly fixing transphobic autocomplete suggestions for 50+ trans celebrities — making Rekognition an early instance of automated systems structurally misreading marginalized users.
First-order effects
- Trans, queer, and nonbinary individuals face immediate material risk where agencies like the Washington County Sheriff's Office run Rekognition, since a classifier that assigns gender 'by design' turns identity mismatch into false matches or denied access.
Second-order effects
- Police customers of Rekognition inherit an accuracy-and-liability problem: every misgendering finding undermines confidence in gender-based classification outputs used for identification, pressuring Amazon to document or constrain those features.
Third-order effects
- If the pattern holds — Rekognition's misgendering, ethnicity-detection worries, and search-engine transphobia being the same failure class — biometric and classification vendors move from shipping defaults to defending them, pulling public-safety AI toward explicit bias auditing and governance requirements.
The trend: Facial recognition and automated classification are being deployed into policing and commerce faster than their accuracy for gender-nonconforming faces can be established, making bias findings like this one the forcing function for oversight.