Microsoft says it has improved its facial recognition tech to work better with all skin tones, reducing error rates by up to 20 times for people of darker skin
Microsoft still has a contract with ICE to use its facial recognition technology — Microsoft says its facial recognition tools …
Context & Ripple Effects
The improvement lands five months after the benchmark study found Microsoft, IBM, and Megvii systems misidentified dark-skinned women at 21%-35% rates versus under 1% for light-skinned men — a gap that put Microsoft's accuracy claims, and its ongoing ICE contract, under direct scrutiny. It also follows Microsoft's own published playbook on fighting bias through diversified training data and per-group accuracy disclosure (Wired's survey of industry fixes).
The timing matters politically as much as technically: weeks later, Brad Smith would call for government regulation of facial recognition while defending the ICE deal (Smith's regulation push), making this error-rate reduction the technical proof point Microsoft can cite when arguing the technology can be fixed rather than banned.
First-order effects
- Microsoft's dark-skin error rate drops by up to 20 times, directly answering the failure mode documented in the February study and giving government customers on existing contracts — including ICE — a more accurate tool without renegotiation.
- The claim hands Microsoft a concrete rebuttal to critics like those behind the California law-enforcement request it later denied over bias concerns (the denied request and prison sale).
Second-order effects
- IBM and Megvii, named alongside Microsoft in the same study, face pressure to publish comparable per-group error reductions or cede the 'responsible vendor' position in government procurement.
- Per-group accuracy disclosure shifts from voluntary transparency to a de facto sales requirement, since buyers can now compare vendors on subgroup performance rather than headline accuracy.
Third-order effects
- If bias proves fixable through training data rather than architecture, the regulatory debate pivots from whether facial recognition should exist to what accuracy floors it must meet — the path that ends four years later with Microsoft removing Azure's emotion- and attribute-guessing tools entirely under its Responsible AI Standard (Azure facial analysis removal).
The trend: Facial recognition is moving from contested deployment toward regulated deployment, where vendors compete on demonstrated per-group accuracy and governments set the thresholds.