Clearview AI, the facial recognition company that claims to have a database of more than 3 billion photos scraped from websites and social media, has been telling prospective law enforcement clients that a review of its software based on “methodology used by the American Civil Liberties Union” is stunningly accurate.
“The Independent Review Panel determined that Clearview rated 100% accurate, producing instant and accurate matches for every photo image in the test,” read an October 2019 report that was included as part of the company’s pitch to the North Miami Police Department. “Accuracy was consistent across all racial & demographic groups.”
But the ACLU said that claim is highly misleading and noted that Clearview’s effort to mimic the methodology of its 2018 facial recognition study was a misguided attempt in “manufacturing endorsements.”
“The report is absurd on many levels and further demonstrates that Clearview simply does not understand the harms of its technology in law enforcement hands,” ACLU Northern California attorney Jacob Snow told BuzzFeed News, which obtained the document through a public records request.
Clearview’s announcement that its technology has been vetted using ACLU guidelines is the latest questionable marketing claim made by the Manhattan-based startup, which has amassed a vast repository of biometric data by scraping photos from social media platforms, including Facebook, Instagram, Twitter, YouTube, and LinkedIn. Among those claims, Clearview AI has told prospective clients that its technology was instrumental in several arrests in New York, including one of an individual involved in a Brooklyn bar beating and another of a suspect who allegedly planted fake bombs in a New York City subway station. The NYPD denied using Clearview’s technology in both of these cases.
Got a tip? Email one of the reporters of this story at [email protected] or [email protected], or contact us here.
Clearview, which claims to be working with more than 600 law enforcement agencies, has also been sued and publicly denounced by critics including New Jersey Attorney General Gurbir Grewal, who ordered a moratorium on the state’s use of the technology after the company included his image without permission in a promotional video. As of last week, Facebook, YouTube, LinkedIn, and PayPal had all sent cease-and-desist letters to Clearview in an attempt to stop it from using images taken from their sites.
Clearview CEO Hoan Ton-That, however, has remained defiant, arguing in a CBS interview last Wednesday that his company has a First Amendment right to scrape public photos from social media. The company also defended the test in statements to BuzzFeed News.
“The ACLU is a highly-respected institution that conducted its own widely distributed test of facial recognition for accuracy across demographic groups,” Ton-That told BuzzFeed News. “We appreciated the ACLU’s efforts to highlight the potential for demographic bias in AI, which is why we applied their test and methodology to our own technology.”
The ACLU’s July 2018 examination of Amazon’s facial recognition tool, Rekognition, is a highly referenced report that illustrates how facial recognition technology misidentifies or falsely matches people of color more often than white people. For the report, ACLU researchers fed Rekognition photos of all sitting members of Congress and asked it to find matches from a database of 25,000 publicly available arrest photos. The test returned dozens of false positives, mostly for elected officials of color.
In October, Clearview said that it “used the same basic methodology used by the American Civil Liberties Union” to determine that its technology made instant and accurate matches “across all racial & demographic groups.” By citing the ACLU’s name, Clearview AI implied that its technology has been vetted with the same standards that had been employed in the civil rights group’s analysis.
But the ACLU told BuzzFeed News its methodology was quite different from the one used in Clearview AI’s test.
The key difference is between the image databases used in the two studies. While the ACLU employed a database of tens of thousands of suspect mugshots, Clearview’s test was run against what was then a set of 2.8 billion photos pulled from social networks and public websites. Given that Clearview has claimed to have scraped millions of websites, it’s entirely possible that the images of lawmakers that were used in Clearview’s test were already in the database, making the prospect of creating matches much easier.
“But what happens when police search for a person whose photo isn’t in the database? How often will the system return a false match?”
“Rather than searching for lawmakers against a database of arrest photos, Clearview apparently searched its own shadily-assembled database of photos,” Snow said. “Clearview claim[ed] that images of the lawmakers were present in the company’s massive repository of face scans. But what happens when police search for a person whose photo isn’t in the database? How often will the system return a false match? Are the rates of error worse for people of color?”
It’s also atypical for an officer investigating a case to have a clear headshot of a suspect, like the ones Clearview presumably used as inputs for the lawmakers. Clearview’s tool is intended to be used in real-world situations, where photo quality, lighting, and other factors can skew the process, and should be tested as such, said Snow.
“If Clearview is so confident about its technology, it should subject its product to rigorous independent testing in real-life settings,” he said. “And it should give the public the right to decide whether the government is permitted to use its product at all.”
Ton-That said that the independent panel’s test was already diligent and thorough. “The Clearview test ran the same photos as the ACLU did, but against a database that was over 100,000 times larger,” he said, noting that in addition to searching for 535 US congressional leaders, his company’s technology was tried on state legislators from Texas and California. “With that higher level of difficulty, Clearview scored 100% following the ACLU standard.”
Facial recognition researchers expressed serious doubts about Clearview’s report. While the ACLU study was effective in demonstrating a facial recognition software’s deficiencies, it is by no means a sufficient methodology for definitively assessing the accuracy of a commercial tool like Clearview AI, Liz O’Sullivan, the technology director at the Surveillance Technology Oversight Project, told BuzzFeed News. O’Sullivan also questioned the panel’s claim that Clearview’s tech is accurate for “all demographic groups,” given the study’s test group of 834 state and federal lawmakers is not representative of all people or ethnicities.
Clearview CEO Ton-That disagreed with this assessment. “The rigors of the test have covered every demographic group that is represented in the general population and have shown Clearview’s accuracy when searching out of billions of photos,” he said.
Facial recognition and privacy researcher Adam Harvey told BuzzFeed News that it is impossible to evaluate the accuracy of the Clearview study without more insight into how it was conducted. “This document does not provide sufficient information to validate their claim,” he said. “It appears that no one in the panel has any prior experience with face recognition.”
Clearview’s report was signed off by a three-person panel, which included Jonathan Lippman, chief judge of the New York Court of Appeals from 2009 to 2015; Dr. Nicholas Cassimatis, an artificial intelligence academic and entrepreneur; and Aaron Renn, an urban analyst and former senior fellow at the Manhattan Institute, a conservative think tank. The panel determined whether the two top-ranked matches from Clearview search results showed the same person in the original search.
“In October 2019, the undersigned Panel conducted an independent accuracy test of Clearview AI, a new image-matching technology that functions as an Internet search engine for faces,” the report read.
None of the panelists appear to have any expertise in facial recognition. Lippman told BuzzFeed News that he “was introduced to Clearview by Richard Schwartz,” one of the company’s cofounders whom he’s known since Schwartz’s time as editorial page editor of the New York Daily News. He said he was not paid for his work on the study.
“I assume I was approached because of my experience as a judge in government and criminal justice, and in looking at and weighing empirical evidence,” Lippman said.
Cassimatis told BuzzFeed News that he had worked in artificial intelligence for 20 years, and pointed to his work as a professor at Rensselaer Polytechnic Institute and as the former “head of Samsung’s North American AI research.” He said he met Ton-That through a mutual friend and was selected “for my expertise in this field.” Cassimatis said he was not paid to work on the study.
Renn did not reply to a request for comment. Previously, Ton-That has said he and Schwartz had met during an event at the Manhattan Institute, where Renn served as a senior fellow until last year.
Clearview’s “ACLU” study isn’t the first time the company has touted the accuracy of its technology without much in the way of supporting materials or peer review. Last summer, the company told the Atlanta Police Department in marketing materials that its technology was 98.6% accurate in a test of 1 million faces — an accuracy rate higher than that of tools created by Google and Chinese tech giant Tencent. However, that claim, which was made using the University of Washington’s MegaFace facial recognition benchmark, was never independently verified by the university or a third party, the company later told BuzzFeed News. Clearview declined a request to make the results of this test available for review.
Since October, however, the company seems to have moved away from marketing the MegaFace number and gone forward with the 100% accuracy rating from the ACLU-based test done by its three-person panel. After the first news stories were published about the company last month, the company added a new section to its once sparse website called “Clearview Facts,” where it said that “an independent panel of experts rated Clearview 100% accurate across all demographic groups according to the ACLU’s facial recognition accuracy methodology.”
That did not sit well with the ACLU, which eventually lodged a complaint. On Jan. 28, Clearview removed the civil rights group’s name from its site, though it still claims that “an independent panel of experts reviewed and certified Clearview for accuracy and reliability.”
That may not be enough for the ACLU and Snow, who said that any proof of accuracy was “beside the point.”
“Clearview’s technology gives government the unprecedented power to spy on us wherever we go — tracking our faces at protests, [Alcoholics Anonymous] meetings, church, and more,” he said. “Accurate or not, Clearview’s technology in law enforcement hands will end privacy as we know it.”