Regulators kept ruling Clearview illegal as its face database grew past 30 billion photos

A face index scraped from the public web

Clearview AI operated quietly until January 2020, when Kashmir Hill reported in The New York Times that a little-known company had scraped billions of photographs from the public internet and was selling searches against them to police. By late 2021 Australia's privacy regulator, investigating on its own, put the number at more than three billion images, collected from social media and other public sites without the knowledge or consent of the people in them.¹ From each photo the company derives a biometric template: a string of numbers computed from the geometry of a face, precise enough to match one picture of a person to another.¹

The product works like a reverse image search for faces. A police client uploads a photo, which the tribunal that later examined the system called a “Probe Image,” and Clearview returns other photos of that same face from its database, each tagged with the web address where it was scraped.² The UK First-tier Tribunal recorded that every one of Clearview's clients “carry out criminal law enforcement and/or national security functions,” in the United States and in countries including Panama, Brazil, Mexico and the Dominican Republic.²

The three-billion figure was already old by the time regulators reached it. Italy's authority counted more than ten billion faces in early 2022,³ the UK tribunal recorded over twenty billion that October,² and by 2024 the Dutch regulator put the collection at more than thirty billion.⁴ Clearview has since claimed larger numbers still, in the forties, but those are the company's own assertions with no regulator to confirm them. Set only the figures a court or an authority actually counted side by side, and the scraping never paused for any of the rulings against it.

The same finding in five countries

Between 2021 and 2024, privacy regulators in five countries examined Clearview and arrived at the same conclusion. Australia went first. On 3 November 2021 Commissioner Angelene Falk determined that Clearview had breached the Privacy Act 1988 by collecting Australians' sensitive information without consent and by unfair means, and she ordered the company to stop collecting facial images from people in Australia and to destroy what it already held.¹

Europe's four fines rest on two findings the reader needs in order. The first is what the photos legally are. Under the European Union's General Data Protection Regulation, a picture used to recognize someone by their face is “biometric data,” which sits in a protected class the law calls Article 9 data and forbids anyone to process at all unless a narrow exception applies.⁴ The second is that the company had no right to process it. Every controller, meaning the organization that decides why and how personal data is used, needs a “lawful basis,” one of a short list of justifications such as the person's consent, before it may touch personal data.⁵ Clearview had neither. It never obtained consent, and France's CNIL found it could not fall back on “legitimate interests” given “the intrusive character of the data collection and data subjects' lack of awareness.”⁵ Greece's authority reached the same place, holding that the company had “violated the principles of lawfulness and transparency.”⁶

Clearview does not dispute the central fact. Its answer is that the photos were “sourced from public-only web sources” and that its customers are “government agencies.”⁷ No one, Clearview included, claims anyone in the database agreed to be there. The disagreement is not about what happened. It is about whether a photo being public makes it free to take, and five regulators answered no.

Found unlawful, rarely enforced

A regulator finding a practice unlawful is not the same as a penalty being paid, and Clearview is where the two come apart. Take the United Kingdom. The Information Commissioner fined Clearview just over £7.5 million in May 2022 and ordered it to delete UK residents' data.⁸ Clearview appealed, and in October 2023 the First-tier Tribunal set the fine aside. It did so not because the collection was lawful but because it held the regulator had no jurisdiction: the searches were run for foreign governments' law-enforcement work, which the tribunal placed outside the reach of UK data-protection law.² Two years later the Upper Tribunal disagreed, ruling in October 2025 that the lower tribunal had “erred materially in law,” restoring the Commissioner's authority, and sending the case back to be decided on its merits.⁹ Clearview was then granted permission to appeal once more, to the Court of Appeal, on 19 December 2025.¹⁰ So the £7.5 million has never been paid, and no court has yet ruled on whether the privacy findings themselves are right, only on who holds the power to make them.

The European fines face a plainer obstacle. Clearview is a US company that appointed no representative in Europe, itself one of the breaches Italy cited, which leaves the authorities holding judgments with little to collect against.³ France wrote the problem into its own order: it attached a penalty of €100,000 for every day Clearview ignored the demand to delete.⁵ The rulings are real findings of law. Whether the money is ever collected is a separate question, and the record here shows little of it changing hands.

The one order that reshaped Clearview's business came from an American courtroom under a state statute. In 2020 the American Civil Liberties Union sued Clearview in Illinois under the Biometric Information Privacy Act, a law that requires a company to get written permission before capturing a person's faceprint.¹¹ The 2022 settlement did not delete the database. It bars Clearview, nationwide, from selling access to its faceprint database to private companies, and for five years from giving access to any Illinois police department or state agency.¹² It limited who Clearview may sell to, not what it may collect.

A verification score sold as identification

Clearview's public defense leans on accuracy. It reports that in testing by the US National Institute of Standards and Technology its algorithm scored above 99%, including 99.81% on one set of visa photos and 99.76% on mugshots, and its chief executive called the results “an unmistakable validation.”¹³ The number is real. It measures the wrong task.

NIST tested Clearview on what it calls 1:1 verification: given two photographs, are they the same person? That is the question a phone answers when it unlocks to a face it already has on file. Clearview's product does the opposite kind of search, 1:many identification: given one photo, find that person among billions of scraped images. The second task is harder, and it gets harder as the database grows, because every extra face is another chance to return a confident match for the wrong person. A strong verification score says nothing about how often the 1:many search is right.

On that question the record is thin, and what exists is not reassuring. The only regulator to examine the identifications Clearview actually delivered, Australia's, found that the company took no reasonable steps to ensure the matches it handed to clients were accurate.¹ Earlier, in 2020, before it had submitted to any NIST test, Clearview told police it “finds matches 75% of the time” and gave no false-positive rate. It also claimed to have cracked “a case of alleged terrorism in a New York City subway station” in seconds, a claim the New York Police Department denied, saying it “did not use Clearview technology to identify the suspect.”¹⁴

Both things can hold at once. Clearview's algorithm is genuinely good at verification, and NIST measured it. That is not evidence that a search against thirty billion scraped faces returns the right person, and no independent test of that search, on that database, exists in the public record. The marketing puts a verification score where an identification score would go, and the identification score does not exist.