Amazon's Rekognition mistakes women as men 19% of the time, and darker-skinned women as men 31% of the time, more than similar services from IBM and Microsoft
Commercial AI System for Analyzing Faces. aies-conference.com : Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products Abrar Al-Heeti / CNET : Amazon's facial tech shows gender, racial bias, MIT study says Tali Arbel / ABC News : Researchers say Amazon face-detection technology shows bias Cat Zakrzewski / Washington Post : The Technology 202: Here's how 2020 hopeful Pete Buttigieg wants to help workers who could lose jobs to AI Tajha Chappellet-Lanier / FedScoop : Study finds biases in Amazon Rekognition's facial analysis tool Kris Holt / Engadget : Amazon's facial-analysis tool showed gender and race bias, says study James Vincent / The Verge : Gender and racial bias found in Amazon's facial recognition technology (again) Tweets: Flavia Dzodan / @redlightvoices : hmmm tell me again how Linnaeus taxonomies are not shaping our tech... 300 years after the fact? http://twitter.com/... Natasha Singer / @natashanyt : What do companies do when M.I.T. researchers find their facial recognition systems have race and gender biases? a) agree and improve the technology (Microsoft and IBM) b) argue the researchers' findings are incorrect and irrelevant (Amazon) http://www.nytimes.com/... Joy Buolamwini / @jovialjoy : New @MIT study shows gender and racial bias in @amazon Rekognition AI product -100% accuracy on pale males vs 69% accuracy on women of color http://www.nytimes.com/... - Study link http://www.aies-conference.com/ ... @ACLU @Data4BlackLives @black_in_ai @medialab @AIESConf @AINowInstitute @AOC http://twitter.com/... @nytimesbusiness : Our @natashanyt writes that as Microsoft has called on Congress to regulate facial recognition technology, Amazon has aggressively marketed its facial recognition technology to police departments and federal agencies. Here's why it's important: http://www.nytimes.com/... Dare Obasanjo / @carnage4life : Algorithms can't be biased because they are math?Amazon's Rekognition facial recognition system misclassified women as men 19% of the time & mistook darker-skinned women for men 31% of the time.BUT...it never misclassified gender of lighter skinned men http://www.nytimes.com/... Matt Wood / @mza : Some quick thoughts on the paper and article that published today on Rekognition: http://www.nytimes.com/... http://twitter.com/...
Context & Ripple Effects
This MIT audit extends the line of research Joy Buolamwini began with the [[a:926589|2018 facial-recognition study that found 21%-35% error rates for dark-skinned women versus under 1% for light-skinned men]] across Microsoft, IBM, and Megvii — but this time it names Amazon's Rekognition as the worst performer among the major US vendors, misclassifying women as men 19% of the time overall and 31% of the time for darker-skinned women.
The finding lands on a product Amazon sells commercially, including to government users, and it compounds earlier scrutiny: months later an investigation found Rekognition frequently misgenders trans, queer, and nonbinary individuals by design. The through-line is that independent researchers keep documenting demographic error gaps in shipped products, not lab demos.
First-order effects
- Amazon faces a direct competitive and reputational problem: IBM and Microsoft offer similar face-analysis services with measurably lower error rates on the same benchmark, giving enterprise and government buyers a documented reason to switch or avoid Rekognition.
Second-order effects
- Rivals can now compete on audited fairness rather than raw capability, making third-party benchmark results like this one a de facto procurement criterion for agencies and enterprises evaluating face-analysis vendors.
Third-order effects
- With bias findings accumulating across modalities — the same research community later showed [[a:951855|speech systems from Amazon, Apple, Google, IBM, and Microsoft erring far more often on black speakers than white ones]] — the pattern points toward mandatory independent bias auditing and regulation of commercial AI before sale, rather than voluntary vendor testing.
The trend: Commercial AI vendors are being held to independently published demographic error benchmarks, shifting accountability from internal claims to external audits.