Facial recognition privacy: what the regulators have ruled

This page is part of the face recognition section. It gathers the rules in date order; the decisions against one company are set out case by case on the Clearview AI rulings page, and the research datasets that disappeared are on the face datasets and consent page.

01What do privacy laws say about face data?

The General Data Protection Regulation defines biometric data as "personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person, which allow or confirm the unique identification of that natural person, such as facial images or dactyloscopic data" (Article 4(14)). Article 9 then prohibits processing "biometric data for the purpose of uniquely identifying a natural person" unless an exception applies, the first being explicit consent. The regulation has applied since 25 May 2018, and Article 83 sets fines for breaching Article 9 at up to 20,000,000 euros or 4 percent of worldwide annual turnover, whichever is higher (Regulation (EU) 2016/679).

The Illinois Biometric Information Privacy Act, in effect since 3 October 2008, takes a different route: a private right of action. Section 15 bars a private entity from collecting a biometric identifier unless it first informs the subject in writing and "receives a written release executed by the subject", and bars selling or otherwise profiting from it (740 ILCS 14/15). Section 20 lets "any person aggrieved by a violation" sue for liquidated damages of $1,000 per negligent violation or $5,000 per intentional or reckless one, and an amendment effective 2 August 2024 provides that repeated collection of the same identifier from the same person by the same method counts as a single violation (740 ILCS 14/20).

02What have regulators ruled on scraped face databases?

Five privacy regulators reached the same conclusion about one scraped database between 2021 and 2024, and the decisions are the most-cited facial recognition privacy rulings in existence. The table lists them with the primary record for each.

The UK decision then went through two tribunals. The First-tier Tribunal allowed the company's appeal on 17 October 2023, holding that the Commissioner lacked jurisdiction; the Upper Tribunal set that decision aside on 6 October 2025, held that the ICO did have jurisdiction, and sent the case back for a fresh hearing on the substance ([2025] UKUT 319 (AAC)). The penalty itself therefore remains contested as of this page's read date, which is why the ICO figure is listed as imposed rather than as final.

03What did the FTC order about face recognition in stores?

The United States has no federal biometric statute, so the one national decision is a consumer protection order. On 19 December 2023 the Federal Trade Commission announced a settlement under which the retailer Rite Aid "will be prohibited from using facial recognition technology for surveillance purposes for five years". The complaint states that from 2012 to 2020 the company used the technology in hundreds of stores to flag suspected shoplifters, that it "failed to take reasonable measures to prevent harm to consumers", and that its system "generated thousands of false-positive matches" (FTC press release).

The order's findings are a privacy case built from accuracy failures. The FTC says the technology "was more likely to generate false positives in stores located in plurality-Black and Asian communities than in plurality-White communities", that the company never tested or measured its accuracy before or after deployment, and that it enrolled low-quality images from security cameras and phone cameras. The remedies require deletion of the images and of any models built from them, notice to consumers when they are enrolled, and deletion of any biometric information within five years.

04Which research datasets were withdrawn?

Regulators acted on commercial databases, but the research datasets that trained a generation of recognition models were withdrawn by their own hosts. MS-Celeb-1M, released in 2016 with a training set of "10M images in version 1", is no longer served from its original site. MegaFace's project page states that "MegaFace is being decommissioned and MegaFace data are no longer being distributed", after its 1 million distractor photos of 690,572 users had carried the benchmark past "over 99%" performance. NIST "discontinued distribution of all three IJB challenges (IJB-A, IJB-B, and IJB-C) on March 14th 2023" (NIST face challenges).

None of those three notices gives a legal reason on the page itself, so this site does not attribute them to any one law. What they share is the collection method the regulators ruled on: faces gathered from the web without the subjects' knowledge. The WebFace260M licence model and the dataset card fields a consented set should carry are described on the face datasets and consent page.

05What does the EU AI Act add?

Regulation (EU) 2024/1689, published in the Official Journal on 12 July 2024, prohibits in Article 5(1)(e) "the placing on the market, the putting into service for this specific purpose, or the use of AI systems that create or expand facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage" (Official Journal text). Recital 43 gives the reason: the practice "adds to the feeling of mass surveillance and can lead to gross violations of fundamental rights, including the right to privacy".

Article 113 makes Chapters I and II, which hold the Article 5 prohibitions, apply from 2 February 2025, eighteen months before the regulation's general application date of 2 August 2026. Article 99(3) sets the fine for breaching Article 5 at up to 35,000,000 euros or 7 percent of worldwide annual turnover, whichever is higher, the highest tier in the regulation. The scraping ban codifies what four of the five regulators above had reached case by case under the GDPR, and what the OAIC reached under Australia's Privacy Act.

06Why accuracy differentials are a privacy question

Every decision above turns on a face being matched to a person who did not agree to it, and the harm scales with how often the match is wrong. NISTIR 8280, NIST's 2019 study of 189 algorithms on 18.27 million images of 8.49 million people, found that false positive rates "often vary by factors of 10 to beyond 100 times" across demographic groups, while false negatives vary by factors below 3 (NISTIR 8280). The demographic differentials in face recognition page sets out those results; the FTC's finding about plurality-Black and Asian neighbourhoods is the same effect observed in one deployment. The police use of face recognition page covers the public-sector side, where the GDPR and the AI Act carve out separate rules, and the video call deepfakes page shows the newer privacy risk: a face taken from the web and worn by someone else in a live call.