Clearview AI is still collecting photos of Australians for its facial recognition database

Clearview AI has continued to collect millions of images of Australians’ faces for its facial recognition software even though the privacy watchdog ordered the “world’s most controversial company” to stop.

The revelation is based on internal correspondence between lawyers for the US-based Clearview and the Australian government obtained via a freedom of information request, legal records and recent public statements by Clearview’s Australian CEO Hoan Ton-That claiming the company couldn’t purge Australians from its database even if it tried.

Clearview AI offers facial recognition software that compares an image to its database of 40 billion images scraped from the internet without permission. In practice, it means that its users can take a photo and upload it to Clearview AI’s servers to see if it matches anything in an enormous database of images that were almost certainly uploaded by people who were unaware their photo was going to be used to train facial recognition software that could identify them. 

In 2021, Australia’s information and privacy commissioner Angelene Falk found that Clearview AI had put Australians at significant risk of harm through a “large scale collection of Australians’ sensitive biometric information” for profit and without permission. Falk’s decision ordered Clearview AI to stop collecting images “from individuals in Australia” and to delete those already obtained in breach of the Privacy Act. 

Today, Clearview AI says its regulatory problems are over. It stopped selling to private companies after a 2022 US lawsuit decision and now claims to sell only to police and government customers, none of whom are in Australia. 

The company says its database has grown to 40 billion images, up from three billion at the time of Falk’s decision. But the company now publicly admits it does not know how many Australians are included in this database. Given that Clearview AI took these images from platforms like Facebook and Instagram, it is likely that many of the tens of millions of Australians who have social media accounts are included.

A spokesperson for the Office of the Australian Information Commissioner (OAIC) said it was considering its response to Clearview’s operations. Clearview AI’s general counsel Jack Mulcaire said that it continues to “dispute the OAIC’s mistaken interpretation of Australian law”.

How Clearview AI got Australians’ data from outside of Australia

A few months after Falk’s finding, Clearview AI filed an appeal. Among its arguments to the Administrative Appeals Tribunal (AAT) was the idea that the US-based company did not have an “Australian link” and so the Privacy Act does not apply to it. The company argued that the images it took from the internet that included the faces of Australians were not taken from Australian servers.

To make sense of this, you need to understand how Clearview AI and a lot of the internet works. Clearview AI built its facial recognition database using a tool called a “web crawler”. Rather than the herculean task of getting people to manually visit the near-infinite number of web pages to see if there are images to download, crawlers are a bot that automates that task. They’re best known for their role in search engine companies like Google, which use them to index web pages so it can direct their users to them.

Clearview AI scraped — meaning to use crawlers to obtain data at scale and without permission — images that were on platforms like Facebook, Instagram and LinkedIn. The company claims that these images are public and fine to take without permission, but Falk rejected this with the finding that Australians did not agree to the collection of their biometric data just because they uploaded it to social media platforms. (The companies who held this data also did not agree to its scraping and sent cease-and-desist letters to Clearview AI after they became aware of it.)

Many online services and websites, including those scraped by Clearview AI, have servers hosting their content, including Australians’ data, that are located around the world. For example, Meta, Facebook’s owner, has servers in Australia but it also has servers that store identical copies of the data around the world (or in technical parlance, data that is cached in content delivery networks). This means that when an Australian user in Australia uploads a photo taken in Australia to an Australian Meta server, it will still likely end up on international servers. The same goes for just about every other major website.

In its appeal, Clearview argued that because its crawlers were operating out of New York and accessed North American servers that meant the company wasn’t conducting any business in Australia even if they had ended up with Australians’ data. Clearview AI won the battle but lost the war. In a decision handed down in May last year that upheld the original decision, AAT senior member Damien O’Donovan agreed that scraping Australians’ data from US servers was not enough to give Clearview an Australian link but other activities by the company were. It wouldn’t be the last time the server location argument would come up.

The fight to enforce Clearview AI’s Australian face ban

Even though the appeal against Falk’s landmark decision was rejected, enforcing it was another matter. Documents obtained from the OAIC reveal how Clearview AI has for years argued that it has no idea whether an image came from Australia or an Australian. Clearview has also fought back against proving that it is now taking steps to be able to make these identifications.

During the original investigation by the OAIC in August 2020, BAL Lawyers’ legal director Mark Love, who has represented Clearview AI, wrote a letter saying that the company did not collect any information aside from images and their URLs, and had not taken any other data associated with the images that could identify whether it depicted any Australians.

“Clearview is unable to determine the location or nationality of individuals depicted in images held by Clearview,” he said on August 19, 2020. 

Three years later, emails between the Australian Government Solicitors (AGS) on behalf of the OAIC and Love show how the government repeatedly had to ask to get evidence that Clearview had taken steps to stop collecting Australians’ data.

In June last year, AGS lawyers asked that the AAT order Clearview AI to confirm that it had deleted images taken from Australian servers. Thanks to a set of 2022 privacy law reforms subsequent to the original OAIC finding, the lawyers also claimed that all images collected by Clearview AI that were likely to include Australians or individuals in Australia were also subject to the Privacy Act regardless of whether they were sourced from an Australian server or otherwise. It asked the company to delete any image from an Australian domain that had geolocation data showing it was taken in Australia or was from a social media account that indicated it was from an Australian — essentially, any image that could reasonably be suspected to contain an Australian even if it wasn’t from an Australian server.

According to a follow-up letter in July from an AGS senior executive lawyer, Clearview AI promised it had only started blocking Australian servers based on IP addresses as of January 2023, years after the original decision. The company also claimed it had not collected any new images since it temporarily paused global collection in October 2022, a fact that it had not previously disclosed to the AAT.

The AGS lawyer claimed the company had not provided any proof of its IP address blocking nor had it explained how it had done so in an “extremely brief” six-paragraph witness statement. They wrote that the AGS would seek the tribunal to force Clearview AI to give proof including documents proving the effectiveness of its operation.

Love’s response was blistering. He rejected what he said was an implication that Clearview AI was lying and told the Australian government to read Wikipedia if it wanted to understand the IP address blocking functionality.

“The matters stated are not complicated … With respect, such statements by [redacted] need no ‘support’ …. No explanation of hardware or firmware functionality is either warranted or needed. Both IBM and Wikipedia give reasonable explanations,” he wrote.

Love also pointed out that most of the websites searched by Clearview AI’s crawlers were not in Australia, noting that even the OAIC’s website was hosted elsewhere.

“The overwhelming volume of relevant information to be searched, like your client’s own website, is situated within a small number of locations, none of which are ‘Australia’,” he said.

Further emails show that the AGS again requested that the AAT order Clearview AI to prove that it stopped crawling Australian servers and that the company agreed. However, a hearing to confirm these orders did not go ahead after Clearview withdrew its appeal application.

The OAIC confirmed that the hearing did not go ahead but did not answer questions regarding whether it was satisfied that Clearview AI had done everything asked of it.

Mulcaire, Clearview AI’s general counsel, told Crikey that it chose to withdraw from AAT proceedings to “focus its resources on other matters”. He also said that the company has not had business or users in Australia since 2020.

‘There’s no way to know if someone’s an Australian or not’

Clearview AI and BAL Lawyers also didn’t answer questions about the legal jousting over the orders, but the company’s CEO Ton-That has made it clear he was aware that Clearview AI continues to collect data on Australians in an interview early this year.

Speaking to MLex in January this year, Ton-That said it was “heartbreaking” that the country he’s from decided that this is the technology it doesn’t want. He also mentioned that Australian police have had “success” using Clearview’s products. (Crikey has reported that the Australian Federal Police continued to meet with Clearview AI and even gave evidence to third parties who analysed it with the company’s technology after Falk criticised the agency for trialling its products.)

While saying he “can’t comment on it too much legally”, Ton-That said his lawyers told him that the legal orders didn’t apply to images collected from non-Australian servers and that there was no way to know or remove all Australian faces from the database. He gave examples like Australians living in New York or visiting Italy and US citizens residing in Australia as cases where the company wasn’t able to tell if the law applied to them. “Unless we have — and we don’t have — citizenship data tied to photos, there’s no way for us to really know if someone’s Australian or not,” he said.

Ton-That also said that the company had dealt with more than 25 lawsuits since 2020 but now only had two left unresolved: “I’m happy to say that we’re through it. And we’re proud to come to the other side,” he said.

The OAIC told Crikey that it was aware of reports that Clearview AI was carrying on business in Australia.

“[The commissioner is] considering her options in respect of that issue,” the spokesperson said.