Clearview AI rulings: OAIC, ICO, CNIL, Garante and Dutch DPA decisions

This page sits in the face recognition section as the case record behind the face datasets and consent page, which explains how web-scraped face sets are built; the deepfake side of the same question is on the consent in deepfake datasets page. Every date and amount below is read from the regulator's or tribunal's own published decision.

01The decisions in order

The company's database is described differently in each document: "more than three billion images" by the OAIC in 2021, "tens of billions" of mapped images by the Upper Tribunal in 2025, and "more than 30 billion photos" by the Dutch DPA. The figures are not in conflict; they are statements made years apart about a database that kept growing.

02What did the Australian OAIC decide in 2021?

The Office of the Australian Information Commissioner issued its determination on 3 November 2021 after a joint investigation with the UK ICO. It found that Clearview AI breached the Privacy Act 1988 by "collecting Australians' sensitive information without consent", by collecting personal information by unfair means, and by failing to notify people or to ensure accuracy (OAIC media release, 3 November 2021). The determination ordered the company to "cease collecting facial images and biometric templates from individuals in Australia, and to destroy existing images and templates collected from Australia". The release mentions no fine; the determination worked through orders.

The release records that the tool held "a database of more than three billion images taken from social media platforms and other publicly available websites", and that the company argued the data was not personal information and that, as a US company, it fell outside the Act. Both arguments were rejected, and both were argued again in the United Kingdom.

03What did the ICO fine, and what happened on appeal?

The UK Information Commissioner's Office announced on 26 May 2022 that it had fined "facial recognition database company Clearview AI Inc £7.5m" and ordered UK data to be deleted (ICO enforcement listing). The underlying enforcement notice and monetary penalty notice are dated 18 May 2022. The company appealed, and the record then runs through two tribunals.

The First-tier Tribunal heard the appeal on 21 to 23 November 2022 and gave its decision on 17 October 2023: "The appeal is Allowed". It held that the Commissioner "did not have jurisdiction to issue the Enforcement Notice and the Monetary Penalty Notice", because the company's processing fell outside the material scope of the GDPR given that its clients were foreign law enforcement and national security bodies (Clearview AI Inc v The Information Commissioner, [2023] UKFTT 819 (GRC)). The penalty fell with the notices.

The Information Commissioner appealed, and the Upper Tribunal heard the case on 9 to 11 June 2025 with Privacy International intervening. Its decision, authorised for issue on 6 October 2025, allowed the appeal: the First-tier decision "was materially in error of law", and "On a proper construction of the GDPRs the ICO had jurisdiction to issue the Notices" (The Information Commissioner v Clearview AI Incorporated, [2025] UKUT 319 (AAC)). The Upper Tribunal did not rule on whether the fine was justified; it remitted "this matter to a new FTT for consideration of the substantive appeal", so the GBP 7.5 million penalty awaits a fresh hearing on its merits.

04What did the CNIL and the Garante decide?

France's CNIL acted through its restricted committee in deliberation SAN-2022-019 of 17 October 2022. It found breaches of articles 6, 12, 15 and 17 of the GDPR, that is, no legal basis for the processing and failures on transparency, access and erasure, and imposed "une amende administrative d'un montant de 20 000 000 (vingt millions) euros" (Legifrance, CNILTEXT000046444859). It also ordered the company to stop collecting and processing the data of people in France without a legal basis and to delete what it held, backed by a penalty of EUR 100,000 per day of delay after two months; a formal notice of 26 November 2021 had gone unanswered.

Italy's Garante had reached the same amount eight months earlier. Its ordinanza ingiunzione of 10 February 2022, register of measures n. 50, ordered the company "di pagare la somma di euro venti milioni a titolo di sanzione amministrativa pecuniaria" (Garante, docweb 9751362). It banned any further web scraping of images and metadata of people in Italy and any further processing of their data through the facial recognition system, ordered the data erased, and required the company to designate a representative in the European Union within thirty days. The Garante's decision treats the face vectors as biometric data under article 9 of the GDPR, a category whose processing is prohibited unless an exception applies, and the Dutch decision below reasons the same way.

05What did the Dutch DPA add in 2024?

The Dutch Data Protection Authority imposed the largest of the fines. Its press release states that it "imposes a fine of 30.5 million euro and orders subject to a penalty for non-compliance up to more than 5 million euro on Clearview AI", with the penalty payments capped at "a total maximum amount of 5.1 million euro on top of the fine" (Dutch DPA press release). It describes a database of "more than 30 billion photos of people" and finds two violations: the database should never have been built, because the biometric codes are prohibited data, and the company does not answer access requests.

The authority also warned that using the service is unlawful for Dutch organisations and said it was examining whether the company's directors could be held personally liable. The release records that "Clearview has not objected to this decision and is therefore unable to appeal against the fine", so unlike the UK penalty this one is final.

06What do the rulings mean for face datasets?

The decisions concern a commercial product, not a research dataset, and each applies its own national law. What they share is a finding that faces taken from public web pages are personal data, that face vectors derived from them are biometric data, and that a public source image makes neither lawful to process. The same collection method built most large face recognition training sets, as the face datasets and consent page documents, and two of the sets listed among the list of face recognition benchmarks were withdrawn before any of these rulings.

A search of a scraped database is a 1:N identification, the task NIST scores by false positive and false negative identification rates, and the OAIC release warns that "Individuals featured in the database may also be at risk of misidentification". How that risk is measured, and how it differs by group, is the subject of the demographic differentials page; the two tasks themselves are set out under what face recognition is. This page states decisions and amounts only; it does not assess the company's accuracy, which none of the decisions measured.

What official audits found about police use more broadly is collected on the law enforcement face recognition page.