Clearview AI’s controversial business practices seem to have produced a more robust facial recognition system. The company is now boasting about its strong performance in the NIST’s most recent Face Recognition Vendor Test, the results of which were released on October 28.
In the test, Clearview AI’s facial recognition algorithm proved to be more accurate than any other algorithm developed in the United States in a range of different categories. The solution matched VISA photos with 99.81 percent accuracy, and posted similar scores for Mugshots, VISA Border photos, and Border photos, which were matched with 99.76 percent, 99.7 percent accuracy, and 99.42 percent accuracy, respectively.
The company also took the top in the US in the Wild photos category, with a score that was good enough for second place when the rest of the world is taken into account. That trend held for all of the categories in which Clearview did well, resulting in a top five finish internationally in all of the aforementioned categories.
While Clearview AI welcomed the news, the announcement will likely give its critics more reason to be alarmed. The company revealed that its database now includes more than 10 billion images pulled from the web, a number that triples the 3 billion photo database that was detailed when the company’s operations first came to light in early 2020. The new number suggests that Clearview has continued to scrape the web for images, despite being hit with several international injunctions, lawsuits, and cease and desist orders asking the company to stop the practice. Clearview has argued that any images posted to social media sites are freely available to the public, and is currently testing that theory in court. Either way, the expanding database indicates that the company still has the same flagrant disregard for privacy.
Nevertheless, the company seems to have used that massive database to build a more accurate facial recognition system. In doing so, it also managed to address potential bias, insofar as the algorithm had a 99 percent accuracy rate across all demographics examined in the NIST test. That should assuage some concerns about the system’s performance, but it does not have any bearing on the broader questions of police surveillance and consent.
Clearview’s algorithm was one of the 650 that were submitted for the NIST’s one-to-one facial recognition evaluation. The company is now planning to submit its algorithm for one-to-many testing at some point relatively soon.
November 3, 2021 – by Eric Weiss