Microsoft has pulled its facial recognition database, MS Celeb, which contained 10M+ images of ~100K individuals scraped under the Creative Commons license
Madhumita Murgia / Financial Times :
Context & Ripple Effects
The pull is the sharpest move yet in a story the Financial Times has been building all spring: an [[a:940864|April survey of the face datasets available on request from universities and the US government]] showed how much of commercial facial recognition rests on images collected for other purposes. MS Celeb was the biggest example — 10M+ images of ~100K people swept up under Creative Commons licenses that were never meant to authorize biometric reuse.
First-order effects
- Researchers and companies that trained models on MS Celeb lose their primary training corpus overnight, and Microsoft removes its own brand from the most-cited example of license-scraped biometric data.
- The move lands weeks after Microsoft publicly refused a California law enforcement request over bias concerns while still selling the tech to a US prison — so the company is retreating from the data layer while keeping the product layer.
Second-order effects
- Every university and government lab holding similar request-based face datasets now faces the same question Microsoft just answered for itself: whether a Creative Commons tag constitutes consent, and whether their holdings are a reputational liability.
- Creative Commons itself is pushed into clarifying that its licenses govern attribution and redistribution of works, not the harvesting of identifiable faces — a boundary the licensing framework was never designed to draw.
Third-order effects
- If the pattern holds, the face-recognition data supply chain gets rebuilt around explicit collection rather than scraping: Microsoft went on to strip gender, age, and emotion inference from Azure under its Responsible AI Standard (the 2022 removal), and Meta deleted over a billion face scans outright — suggesting dataset withdrawal is the first step in a broader corporate unwind of facial analysis.
- For regulators, each voluntary pull becomes evidence that the permission gap between copyright law and biometric privacy cannot be left to corporate discretion, strengthening the case for rules that treat a face as something no existing license can grant.
The trend: The companies that built facial recognition on repurposed public images are dismantling that data supply chain themselves, converting license-scraped datasets from an asset into a liability.