Facial recognition is the plutonium of AI

Luke Stark:

By analogizing facial recognition to plutonium, I want to add two broad points to an increasingly lively debate about the risks of facial recognition technologies. First, facial recognition technologies, by virtue of the way they work at a technical level, have insurmountable flaws connected to the way they schematize human faces. These flaws both create and reinforce discredited categorizations around gender and race, with socially toxic effects. The second is, in light of these core flaws, the risks of these technologies vastly outweigh the benefits, in a way that's reminiscent of hazardous nuclear technologies. That is why the metaphor of plutonium is apt. Facial recognition, simply by being designed and built, is intrinsically socially toxic, regardless of the intentions of its makers; it needs controls so strict that it should be banned for almost all practical purposes.

It's a good analogy, and one that I agree with, but if he made a stronger case for it than in that single paragraph, none of us will ever get to know, because he paywalled his paper. Sorry, buddy. The rest of your argument has been consigned to oblivion.

Update: Non-paywalled version here.

In the case of facial recognition, the schematization of human facial features is driven by a conceptual logic that these theorists and others [...] have identified as fundamentally racist because it is concerned with using statistical methods to arbitrarily divide human populations.

This process of biopolitical management is grounded in finding numerical reasons for construing some groups as subordinate, and then reifying that subordination by wielding the “charisma of numbers” to claim subordination is a “natural” fact. As such, racism’s function, as Foucault describes it, is “a way of introducing a break into the domain of life [...] of fragmenting the field of the biological that power controls”. Race and racism are “the preconditions that make killing acceptable” in societies focused on making discriminations based on technical norms and standards -- the justification in turning authority’s custodianship of life and living into that of death and dying.

Transuranic phrenology!

Previously, previously, previously, previously, previously, previously, previously, previously.

Tags: , , , , , , ,

37 Responses:

  1. jtth says:

    "He paywalled his paper" is a weird way to put it. Unless he pays off the journal, he usually doesn't have much of a say in that. Also, in his tweet announcing it he mentioned that he'll send the paper to you if you ask on twitter, which he did immediately when I asked. So c'mon.

    • jwz says:

      It's funny, when I write words, I am able to make decisions that allow other people to see them.

      The decision to not publish his argument in public is 100% his.

      • jtth says:

        Given that he works for Microsoft Research and this is part of that job, it probably isn't.

        • jwz says:

          Decisions. Life's full of them.

        • Anonymous says:

          MSR publishes in public all the time. Like this work that was in the spotlight literally this past week.

          • James says:

            This isn't a Microsoft decision, it's because the article is from an ACM publication, and nothing is as absurdist as the ACM's approach to open access. They have eleven broad categories of OA each with their own details and quirks. Their flagship Communications of the ACM where this article was published is constantly changing from one category to another, so their masthead doesn't tell you anything other than a permissions email address. Their copyright assignment form is mandatory for all their publications, and only gives you the ability to retain copyright on visual artwork.

            It took a two decade campaign to get them to allow prior publication as "informal postings on non-peer reviewed servers," so anyone who doesn't post their ACM article to ArXiv, ResearchGate or the like first deserves the ridicule and legal trouble they're likely to get. And there's still someone in the ACM business office trying to claim that "informal" means personal or organizational, not ArXiv etc., so it's best to just avoid ACM altogether if you want to be sure you won't get a nasty surprise down the road. As a corollary, how much do you want to take advice from people who signed over their copyright without publishing a preprint?

        • David E. Konerding says:

          When I wrote papers at Google, we typically posted a PDF of the pre-journal-copyrighted article on the Research site (that's how I get copies of my own published/copyrighted/paywalled articles). I don't think MSFT really encourages publishing in paywall journals unless it's a prestigious one (and you don't see a lot of Nature papers from MSFT).

          • James says:

            Many publishers have ethical standards giving them pause before publishing the works of repeated antitrust abuse felons.

  2. Tony Finch says:

    Yeah the ACM digital library is a shameful relic of dead tree publishing :-( but I found a copy on the author’s web site https://starkcontrast.co/s/Facial-Recognition-is-Plutonium-Stark.pdf

  3. NT says:

    Are you really agreeing with the author that the core problem with facial recognition is that it reinforces gender and race categories? So if it were something like universal lookup by your fingerprints, that would be better?

    • jwz says:

      I am agreeing that facial recognition databases create an enormous racial and gender problem that is largely unacknowledged, and that the problems created by this technology far outweigh the problems solved by it.

      Your question starts from the assumption that "universal lookup" is something that should even exist.

      I am no fan of any massive, let alone "universal", database of identity or biometrics. And I certainly think that any such database needs to be highly regulated. For example, I think the credit data brokers like EvilCorp Equifax should be legislated out of existence.

      • NT says:

        > Your question starts from the assumption that "universal lookup" is something that should even exist.

        Uh, no.
        My point is that the problem with facial recognition is universal lookup, and that the author's focus on race and gender issues is a dangerous distraction from the turnkey dictatorship problem.

        • jwz says:

          Mechanizing the "othering" of people is a key component in the Turnkey Dictatorship toolkit! I do not think that's a distraction.

          • NT says:

            You are taking a bipartisan issue and highlighting the most partisan possible approach to it. Please stop.

            > Mechanizing the "othering" of people is a key component in the Turnkey Dictatorship toolkit!

            Facial recognition is not about "othering" people, it's about recognizing and tracking them. Put down the Foucault.

            • nooj says:

              I just reread your comments, and I still am unsure what your argument is. Is it simply that the ability to distinguish faces reliably, combined with a massive flood of imagery and location data, is unconstitutional and leads to abuses? If not, can you provide more detail?

            • Zach says:

              It's possible, even likely, probably inevitable, for a technology to have more than one problem. One of these problems is the way facial recognition interacts with race and gender, which is not some "only-in-SF complaint"—it's being used in China to "identify Uighur/non-Uighur attributes." Facial recognition, right now, is being used to identify members of an ethnic majority and alert police to pay attention to their actions.

              I don't care what is and isn't "partisan"; race and gender are pretty darn important to how the world works as it is today, and we already have facial recognition systems being used for racism at scale right this minute. We shouldn't pretend race doesn't exist because we don't want to look "partisan," whatever that's supposed to mean.

              If you insist all complaints must be kept non-partisan, that just amounts to giving Fox News a veto over literally everything: "oh nope, can't talk about that now. Sean Hannity ranted about it for 10 minutes on his show."

      • M.E. says:

        So what's your solution to how to establish trust between parties who don't know each other, such as a lender and a loan applicant? It seems like especially given that more and more transactions are performed electronically it is impossible to rely on the mechanisms of trust we evolved with. And it is an open question to me whether those mechanisms are actually any more benign than the state-sponsored ones. The tyranny of tribes and family groups is no nicer than the tyranny of the state for those who fall victim to them. And I believe you are not a fan of cryptocurrency either, which is how at least some people propose to solve this problem.

        • jwz says:

          The fuck does that have to do with anything? Do you think society and finance did not exist before 2002?

          • M.E. says:

            Equifax has existed since 1899. In some form I'm sure the concept of recordkeeping for insurance and credit rating purposes goes back far longer than that.

    • NT says:

      It's not fine, but it's not "facial recognition" either. Worrying about whether a robot might be misgendering you is one of those only-in-SF complaints.

      Facial recognition is the keystone technology that collects all the security cameras whose paths you've crossed, all the photos of you publicly shared by your "friends", and all the dashcams. This is a universal danger, an incoming nightmare for rednecks, queers, and anyone who has learned that power corrupts. The easiest way to sabotage any alliance like that is to turn every goddamn conversation toward race and gender.
      We're roughly on the same side here, which is why it bothers me that you are playing to lose.

      • nooj says:

        Accepting facial recognition in society begins with accepting facial reco-like technology.

        You can't just accept something like this taxi and then ten years from now complain that facial reco is ubiquitous. You have to get out of the taxi and take your money somewhere else. You have to argue against even these dumb encroachments.

        • NT says:

          The seat-back screens are one of the reasons I finally gave up on taxis, so I've never seen this thing.

          The point is, real facial recognition is already here and spreading. Arguing against the dumb taxi encroachment is like shuffling the proverbial deck chairs. Obscure critiques from the MLA crowd are not going to help. If you want to help, you could start by shunning friends who post their photos to social media and everybody with a networked security camera. Why is it so hard to find a non-networked security camera, anyway?

          I mean, go ahead and complain that the camera guesses your pronoun instead of asking, but don't call that facial recognition because that already means something and people need to learn what it means ASAP.

          • jwz says:

            The point is, real facial recognition is already here and spreading

            "I have already surrendered, and you should too."

            • NT says:

              That's not even remotely what I said. Has Twitter destroyed your ability to have a decent debate?

            • rozzin says:

              > > > Accepting facial recognition in society begins with accepting facial reco-like technology.
              > > The point is, real facial recognition is already here and spreading. Arguing against the dumb taxi encroachment is like shuffling the proverbial deck chairs.
              > "I have already surrendered, and you should too."

              If you're saying that "I have already surrended, and you should too" is how you read NT's statement, I find that really... confusing. I read it as "The threat isn't `coming to town', it's on our doorstep right now--it's much more urgent than you think that we all band together and repel it RIGHT NOW before it becomes insurmountable!".

              And I read the "don't call that facial recognition because that already means something and people need to learn what it means ASAP" bit similarly to "I think we really fucked up when we decided to use the same word for both cute little quadcopters and also long-range remote-control planes that drop bombs" (or whatever the actual quotation was).

      • jwz says:

        You think this taxi ad camera (which is in Japan, BTW) is limiting its data ingestion and retention to the gender of the person sitting in front of it? Because I'm sure whichever advertising company committed this con said to themselves, "Hey, whoa, it would be immoral of us to collect a massive database of everyone who ever sat in a cab that had our cameras, correlate, mine, and sell that. We'd have to really think about the implications first." That surely is a thing that would have happened.

        The easiest way to sabotage any alliance like that is to turn every goddamn conversation toward race and gender.

        Just noting here that this the same argument made by the "mainstream" Democratic Party centrists who say, "Hey now, let's stop talking about 'identity politics' and focus on what 'normal people' care about."

        • NT says:

          > Just noting here that this the same argument made by the "mainstream" Democratic Party centrists

          Compromise is part of grown-up politics, yes.

          • James says:

            "compromise" ... I don't think that word means what you think it means. Your usage is more like:

            First they compromised against the socialists, and I did not speak out—
            Because I was not a socialist.

        • margaret says:

          jwz, relax. soon you'll come to find you enjoy the warm embrace of the boot on your neck. you'll feel much better once you stop concerning yourself about how it got there and who's applying the pressure.

          • NT says:

            Why are you so obsessed with warm jackboots?
            Is this one of those goth things? Too many Laibach shows, not enough history classes?

  4. bob says:

    I guess the real issue is the lack of commitment of individuals to paint ball camera lenses, or smear vasoline on them. The more hardcore can just aim laser pointers at any spycams or license cams in public places.

    You have your orders citizen.

  5. Mark Crane says:

    I hate columns in PDF documents. Here is a plain text version.

    https://pastebin.com/2VuvL72Z

  6. Joshka says:

    How long until we see a tattoo artist sue the various facial recognition software products for copyright infringement?

  • Previously