Allowing mass surveillance at Olympics undermines EU efforts to regulate AI

> That gear, vastly superior to old Soviet equipment, is now being used by Russia against Ukraine.

> Swimmer delivery vehicles, and rebreathers.

How exactly are these underwater gear being used by Russians in their invasion?

Are you saying these were used to blow up their own pipelines?

The original argument was that Russia used it as a pretense to get the equipment to then use it in Ukraine. Did you think Russia didn't have any machine guns before 2014?

[0] says that an organization that literally cages minority children released an app that has racial disparity. That is not convincing evidence that state of the art systems are not colorblind.

[1] is an old article talking about then-obsolete algorithms.

[2] finds that, three years ago, the majority of the algorithms showed racial disparities. That strongly suggests that some did not, which means the technical part of this problem has been solved for years.

The big issue here is what the GP brought up. These tools will be used by abusive (racist or not) law enforcement organizations, and, because of the people using them, will likely be a net negative on society.

At this point, arguing the facial recognition technology is racist is like arguing police-issued handguns are racist. It is ignoring the core issue.

Note I’m not claiming that one Amazon Rekognition test is the only demonstration of inaccuracy on minorities; there is obviously a lot more evidence, as you’ve capably provided. I was talking about the provenance of the belief of inaccuracy, which in our research at the time (term used loosely, we weren’t publishing a paper or anything, just trying to understand the market) overwhelmingly went like so: “I read it in the media. Which article? This one, here’s a link”, and the linked article has a statement that facial recognition fails on minorities, and that statement links to the Rekognition result.

There may well have been many more studies with stronger groundings! As it happened, that wasn’t what the media tended to cite, and so it wasn’t the basis for what people tended to end up believing.

(Disclaimer for the next section: my ego compels me to push back on the claim that I didn’t try very hard to understand, this is probably nitpicking, etc.) I am familiar with the first article, I read it when it was published, and it doesn’t really support your claim of a substantial history of studies - “Unfortunately, the two studies discussed here are among the only works on racial bias published in the past decade—far too little review for a technology with the power to implicate people as suspects in a criminal investigation” is a direct quote (I recall agreeing strongly with it at the time). The examples it gives are also illustrative: “One algorithm, which failed to identify the right person in 1 out of 10 encounters with Caucasian subjects, failed nearly twice as often when the photo was of an African American” and “The 2012 study on mug shots found that an algorithm trained exclusively on either African American or Caucasian faces recognized members of the race in its training set more readily than members of any other race”. The historical context at the time of that article is that NIST really wanted to invest more heavily in surveying and analyzing facial recognition racial bias, and this piece was arguing in support of that.

The 2019 NIST report wasn’t out at the time I did the research, but it was in the works. We had actually discussed submitting to NIST’s call for algorithms, but the broad perception in the industry was that was a silly thing to do with your valuable IP, and I personally suspect that as a result most of the submissions to the NIST were not cutting edge or commercially viable. With that said, the NIST report also agrees with me: “those that are the most equitable also rank among the most accurate”.

As for the CBP One app, it should really use a better facial recognition system! The description in that article made it sound like the app was failing to even register Black faces as faces, which is ridiculously bad for an app with that much responsibility. There are many better systems out there and they should upgrade as a matter of urgency. (I’m not unsympathetic to this issue; a large part of the reason I’m no longer working on face rec is that I was refused entry by US immigration. We actually used to joke* that we were building the tech to keep me out of the country.)

* Note this is a joke, to my knowledge we weren’t actually working on border control applications.