New APPS: Art, Politics, Philosophy, Science

Bing also hallucinates, even with footnotes

Large Language Models (LLMs) like ChatGPT are well-known to hallucinate – to make up answers that sound pretty plausible, but have no relation to reality.  That of course is because they’re designed to produce text that sounds about right given a prompt.  What sounds kind of right may or may not be right, however.   ChatGPT-3 made up a hilariously bad answer to a Kierkagaard prompt I gave it and put a bunch of words into Sartre’s mouth.  It also fabricated a medical journal article to support a fabricated risk to oral contraception. ChatGPT-4 kept right on making up cites for me.  It has also defamed an Australian mayor and an American law professor.  Let’s call this a known problem.  You might even suggest, following Harry Frankfurt, that it’s not so much hallucinating as it is bullshitting.

Microsoft’s Bing chatbot-assisted search puts footnotes in its answers.  So it makes sense to wonder if it also hallucinates, or if it does better.  I started with ChatGPT today and asked it to name some articles by “Gordon Hull the philosopher.”  I’ll spare you the details, but suffice it to say it produced a list of six things that I did not write.  When I asked it where I might read one of them, it gave me a reference to an issue of TCS that included neither an article by me nor an article of that title.

So Bing doesn’t have to be spectacular to do better!  I asked Bing the same question and got the following:

ChatGPT Reread Kierkegaard but still makes up cites

Recall that ChatGPT a couple of months ago did a total face plant on the topic of Kierkegaard's knight of faith from the knight of infinite resignation.  Well, with the fullness of time and an upgrade, it's a lot better now: (screen grabs below the fold)

Some Reasons to be Skeptical of AI Authorship, Part 3: Capitalism and Social Justice

By Gordon Hull

In the previous two posts (here and here) I’ve developed a political account of authorship (according to which whether we should treat an AI as an author for journal articles and the like is a political question, not one about what the AI is, or whether its output resembles human output), and argued that AIs can’t be property held accountable.  Here I want to argue that AI authorship risks social justice concerns.

That is, there are social justice reasons to expand human authorship that are not present in AI.  As I mentioned in the original post, researchers like Liboiron are trying to make sure that the humans who put effort into papers, in the sense that they make it possible, get credit.  In a comment to that post, Michael Muller underlines that authorship interacts with precarity in complex ways.  For example, “some academic papers have been written by collectives. Some academic papers have been written by anonymous authors, who fear retribution for what they have said.”  Many authors have precarious employment or political circumstances, and sometimes works are sufficiently communal that entire communities are listed as authors. There are thus very good reasons to use authorship strategically when there are minoritized individuals or people in question.  My reference to Liboiron is meant only to indicate the sort of issue in the strategic use of authorship to protect minoritized or precarious individuals, and to gesture to the more complex versions of the problem that Muller points to.  The claim I want to make here is that , as a general matter, AI authorship isn’t going to help those minoritized people, and might well make matters worse.

If anything, therre’s a plausible case that elevating an AI to author status will make social justice issues worse.  There’s at least two ways to get to that result, one specific to AI and one more generally applicable to cognitive labor.

ChatGPT Putting Words in Sartre's Mouth

As if Sartre didn't produce enough words all by himself!

ChatGPT's response to the following prompt is instructive for those of us who are concerned about ChatGPT being used to cheat.  Read past the content of the answer to notice the made-up citations.  The "consciousness is a question..." line is in fact in the Barnes translation of Being and Nothingness, but is actually a term in the glossary provided by the translator (so it's not on p. 60 - it's on p. 629).  Where did the AI find this?  I'm guessing on the Wikipedia page for the book, which has a "special terms" section that includes the quote (and attributes it to Barnes.  I should add as an aside that Barnes puts it in quote marks, but doesn't reference any source).  The "separation" quote is, as far as I can tell, made up whole cloth.  It does sound vaguely Sartrean, but it doesn't appear to be in the Barnes translation, and I can't find it on Google.  It's also worth pointing out that neither quote is from the section about the cafe - both page numbers are from the bad faith discussion. 

I don't doubt that LLMs will get better (etc etc etc) but for now, bogus citations are a well-known hallmark of ChatGPT.  Watch it make-up quotes from Foucault (and generally cause him to turn over in his grave) here.

Some Reasons to be Skeptical of AI Authorship, Part 2: Accountability

By Gordon Hull

As I argued last time, authorship is a political function, and we should be applying that construction of it to understand whether AI should be considered an author.  Here is a first reason for doing so: AI can’t really be “accountable.”

(a) Research accountability: The various journal editors all emphasize accountability.  This seems fundamentally correct to me.  First, it is unclear what it would mean to hold AI accountable.  Suppose the AI fabricates some evidence, or cites a non-existent study, or otherwise commits something that, were a human to do it, would count as egregious research misconduct.  For the human, we have some remedies that ought, at least in principle, to discourage such behavior.  A person’s reputation can be ruined, their position at a lab or employer terminated, and so on.  None of those incentives would make the slightest difference to the AI.  The only remedy that seems obviously available is retracting the study.  But there’s at least two reasons that’s not enough.  First, as is frequently mentioned, retracted studies still get cited.  A lot.  Retraction Watch even keeps a list of the top-10 most cited papers that have been retracted.  The top one right now is a NEJM paper published in 2013 and retracted in 2018; it had 1905 cites before retraction and 950 after.  The second place paper is a little older, published in 1998 and retracted in 2010, and has been cited more times since its retraction than before.  In other words, papers that are bad enough to be actually retracted cause ongoing harm; a retraction is not a sufficient remedy for research misconduct.  If nothing else, whatever AI is going to find and cite it.  And all of this is assuming something we know to be false, which is that all papers with false data (etc) get retracted.  Second, it’s not clear how retraction disincentivizes an AI any more than any other penalty.  In the meantime, there is at least one good argument in favor making humans accountable for the output of an AI: it incentivizes them to check its work.

ChatGPT Didn't Do the Kierkegaard Reading

You know how sometimes your students don't do the reading?  And then how, when you give them a writing prompt based on it, they try to guess their way to a good answer from the everyday meaning of the words in the prompt?  And how, sometimes, the outcome is spectacularly, wonderfully wrong?

Well, I don't know what else ChatGPT can do, but it can do an uncannily good imitation of such a student!

(oh, and like that student, it blew through the wordcount, apparently on the theory that a lot of words would makeup for a lack of reading)

This was a prompt from my existentialism class (the instructions also tell them they have to quote the text, but I omitted that here, because we already know ChatGPT can't do that).  It's two images because I am technically incompetent to capture the longer-than-a-screen answer into one image:

Come Study Philosophy in Charlotte!

The MA Program at UNC Charlotte has a number of funded lines for our two-year MA program in philosophy.  We're an eclectic, practically-oriented department that emphasizes working across disciplines and philosophical traditions.  If that sounds like you, or a student you know - get in touch!  You can email me ([email protected]), though for a lot of questions I'll pass you along to our grad director, Andrea Pitts ([email protected]).  Or, there's a QR code in the flyer below

 

Some Reasons to be Skeptical of AI Authorship, Part 1: What is an (AI) Author?

By Gordon Hull

Large Language Models (LLMs) like Chat-GPT burst into public consciousness sometime in the second half of last year, and Chat-GPT’s impressive results have led to a wave of concern about the future viability of any profession that depends on writing, or on teaching writing in education.  A lot of this is hype, but one issue that is emerging is the role of AI authorship in academic and other publications; there’s already a handful of submissions that list AI co-authors.  An editorial in Nature published on Feb. 3 outlines the scope of the issues at hand:

“This technology has far-reaching consequences for science and society. Researchers and others have already used ChatGPT and other large language models to write essays and talks, summarize literature, draft and improve papers, as well as identify research gaps and write computer code, including statistical analyses. Soon this technology will evolve to the point that it can design experiments, write and complete manuscripts, conduct peer review and support editorial decisions to accept or reject manuscripts”

As a result:

“Conversational AI is likely to revolutionize research practices and publishing, creating both opportunities and concerns. It might accelerate the innovation process, shorten time-to-publication and, by helping people to write fluently, make science more equitable and increase the diversity of scientific perspectives. However, it could also degrade the quality and transparency of research and fundamentally alter our autonomy as human researchers. ChatGPT and other LLMs produce text that is convincing, but often wrong, so their use can distort scientific facts and spread misinformation.”

The editorial then gives examples of LLM-based problems with incomplete results, bad generalizations, inaccurate summaries, and other easily-generated problems.  It emphasizes accountability (for the content of material: the use of AI should be clearly documented) and the need for the development of truly open AI products as part of a push toward transparency.

Gatecrashers, Blue Buses and Speeding Drivers: Philosophy of Law as a way into AI Accountability (part 2)

By Gordon Hull

Last time, I introduced a number of philosophy of law examples in the context of ML systems and suggested that they might be helpful in thinking differently, and more productively, about holding ML systems accountable.  Here I want to make the application specific.

So: how do these examples translate to ML and AI?  I think one lesson is that we need to specify what exactly we are holding the algorithm accountable for.  For example, if we suspect an algorithm of unfairness or bias, it is necessary to specify precisely what the nature of that bias or unfairness is – for example, that it is more likely to assign high-risk status to Black defendants (for pretrial detention purposes) than it is white ones.  Even specifying fairness in this sense can be hard, because there are conflicting accounts of fairness at play.  But assuming that one can settle that question, we don’t need to specify tokens or individual acts of unfairness (or demand that each of them rise to the level where they would individually create liability) to demand accountability of the algorithm or the system that deploys it – we know that the system will have treated defendants unfairly, even if we don’t know which ones (this is basically a disparate impact standard; recall that one of the original and most cited pieces on how data can be unfair  was framed precisely in terms of disparate impact).

Further, given the difficulties of individual actions (litigation costs, as well as getting access to the algorithms, which defendants will claim as trade secrets) in such cases, it seems wrong to channel accountability through tort liability and demand that individuals prove the algorithm discriminated against him (how could they?  The situation is like the blue bus: if a group of people is 80% likely to reoffend or skip bail, we know that 20% of that group will not, and there is no “error” for which the system can be held accountable).  Policymakers need to conduct regular audits or other supervisory activity designed to ferret out this sort of problem, and demand accountability at the systemic level.

Gatecrashers, Blue Buses and Speeding Drivers: Philosophy of Law as a way into AI Accountability (part 1)

By Gordon Hull

AI systems are notoriously opaque black boxes.  In a now standard paper, Jenna Burrell dissects this notion of opacity into three versions.  The first is when companies deliberately hide information about their algorithms, to avoid competition, maintain trade secrets, and to guard against gaming their algorithms, as happens with Search Engine Optimization techniques.  The second is when reading and understanding code is an esoteric skill, so the systems will remain opaque to all but a very small number of specially-trained individuals.  The third form is unique to ML systems, and boils down to the argument that ML systems generate internal networks of connections that don’t reason like people.  Looking into the mechanics of a system for recognizing handwritten numbers or even a spam detection filter wouldn’t produce anything that a human could understand.  This form of opacity is also the least tractable, and there is a lot of work trying to establish how ML decisions could be made either more transparent or at least more explicable.

Joshua Kroll argues instead that the quest for potentially impossible transparency distracts from what we might more plausibly expect from our ML systems: accountability.  After all, they are designed to do something, and we could begin to assess them according to the internal processes by which they are developed to achieve their design goals, as well as by empirical evidence of what happens when they are employed.  In other words, we don’t need to know exactly how the system can tell a ‘2’ from a ‘3’ as long as we can assess whether it does, and whether that objective is serving nefarious purposes.

I’ve thought for a while that there’s potential help for understanding what accountability means in the philosophy of law literature.  For example, a famous thought experiment features a traffic accident caused by a bus.  We have two sources of information about this accident.  One is an eyewitness who is 70% reliable and says that the bus was blue.  The other is the knowledge that 70% of the buses that were in the area at the time were blue.  Epistemically, these ought to be equal – in both cases, you can say with 70% confidence that the blue bus company is liable for the accident.  But we don’t treat them as the same: as David Enoch and Talia Fisher elaborate, most people prefer the witness to the statistical number.  This is presumably because when the witness is wrong, we can inquire what went wrong.  When the statistic is wrong, it’s not clear that anything like a mistake even happened: the statistics operate at a population level; when applied to individuals, the use of statistical probability will be wrong 30% of the time, and so we have to expect that.  It seems to me that our desire for what amounts to an auditable result is the sort of thing that Kroll is pointing to.

Foucault and Epistemic Injustice (Part 3)

In the previous two posts (first, second), I took up the invitation provided by a recent paper by Daniele Lorenzini to develop some thoughts on the relationship between Foucault’s thought and theorizing around epistemic injustice.  In particular, Miranda Fricker’s account both draws heavily from Foucault and pushes back against his historicism to advocate for a more a-historical normative ground for the theory: testimonial injustice “distorts” who someone is.  Last time, I looked at some of Foucault’s own work in the lectures leading up to Discipline and Punish to develop a sense of how both “truth” and “power” are relevant – and distinguishable – in that work, even as they both are historical constructs. In particular, following Lorenzini, we can distinguish between “x is your diagnosis” and “therefore, you ought to do y.” Here I begin with the complexity introduced in Foucault’s work by his addition of an embryonic genealogy of truth practices. 

Let’s begin with the Psychiatric Power lectures, where Foucault had been talking about the strange role of science (and its personification in the doctor) in the governance of asylums.  There, when speaking of the historical contingency of the modern scientific enterprise, Foucault writes:

New Paper: How Foucault got rid of (Bossy) Marxism

Now published in Critical Review.  Here's the abstract:

Foucault distanced himself from Marxism even though he worked in an environment—left French theory of the 1960s and 1970s—where Marxism was the dominant frame of reference. By viewing Foucault in the context of French Marxist theoretical debates of his day, we can connect his criticisms of Marxism to his discussions of the status of intellectuals. Foucault viewed standard Marxist approaches to the role of intellectuals as a problem of power and knowledge applicable to the Communist party. Marxist party intellectuals, in his view, had developed rigid and universal theories and had used them to prescribe action, which prevented work on the sorts of problems that he uncovered—even though these problems were central to the development of capitalism.

The paper is an attempt to cut a path through some (mostly 1970s) texts to get a handle on what Foucault is doing with his inconsistent references to Marx and Marxism.  There's a complex tangle of issues here, many related to the vicissitudes of the reception of Marx, and I hope that others will be able to add to our understanding of them and the period.

A huge thanks to Shterna Friedman, whose editorial work resulted in a much better article.  Also, my paper is going to be part of a special issue of Critical Review on Foucault - the other papers should be appearing relatively soon.

Foucault and Epistemic Injustice (Part 2)

By Gordon Hull

Last time, I took the opportunity provided by a recent paper by Daniele Lorenzini to develop some thought on the relationship between Foucault’s thought and theorizing around epistemic injustice.  Lorenzini’s initial point, with which I agree fully, is that Fricker’s development of epistemic injustice is, on her own terms, incompatible with Foucault because she wants to maintain a less historicized normative standpoint than Foucauldian genealogy allows.  Epistemic injustice, on Fricker’s reading, involves a distortion of someone’s true identity.  Lorenzini also suggests that Foucault’s late work, which distinguishes between an epistemic “game of truth” and a normative/political “regime of truth” offers the distinction Fricker’s theory needs, by allowing one to critique the regime of truth dependent on a game of truth.  In terms of Foucault’s earlier writings, he does not fully reduce knowledge to power, in the sense that is can be useful to analytically separate them. Here I want to look at a couple of examples of how that plays out in the context of disciplinary power.

Consider the case of delinquency, and what Foucault calls the double mode of disciplinary power (Discipline and Punish, 199): a binary division into two categories (sane/mad, etc.) and then the coercive assignment of individuals into one group or the other.  The core modern division is between normal and abnormal, and we have a whole “set of techniques and institutions for measuring, supervising and correcting the abnormal” (199).  The delinquent, then, is defined epistemically or juridically (in other words, as a matter of science or law; as I will suggest below, Foucault thinks that one of the ways that psychology instituted itself as a science was by successfully blurring the science/law distinction), and then things are done to her.  This is the sort of gap that epistemic injustice theory, at least in its testimonial version, needs: in Fricker’s trial example, there is the epistemic apparatus of “scientific” racism, and then there is the set of techniques that work during the trial.  Both of these can be targets of critique, but testimonial injustice most obviously works within the second of the two.

Foucault and Epistemic Injustice (part 1)

By Gordon Hull

Those of us who have both made extensive use of Foucault and made a foray into questions of epistemic injustice have tended to sweep the question of the relation between the two theoretical approaches under the rug.  Miranda Fricker’s book, which has basically set the agenda for work on epistemic injustice, acknowledges a substantial debt to Foucault, but in later work she backs away from the ultimate implications of his account of power on the grounds that his historicism undermines the ability to make normative claims.  In this her argument makes a fairly standard criticism of Foucault (whose “refusal to separate power and truth” she aligns with Lyotard’s critique of metanarratives (Routledge Handbook of Epistemic Injustice, 55).  As she describes her own project:

“What I hoped for from the concept of epistemic injustice and its cognates was to mark out a delimited space in which to observe some key intersections of knowledge and power at one remove from the long shadows of both Marx and Foucault, by forging an on-the-ground tool of critical understanding that was called for in everyday lived experience of injustice … and which would rely neither on any metaphysically burdened theoretical narrative of an epistemically well-placed sex-class, nor on any risky flirtation with a reduction of truth or knowledge to de facto social power” (Routledge Handbook, 56).

On this reading, then, Marxism relies too much on ideology-critique, on the one hand, and on privileging the position of women/the proletariat (or other, singular subject position).  Foucault goes too far and reduces the normative dimension altogether. 

In a new paper, Daniele Lorenzini addresses the Foucault/Fricker question head-on, centrally focusing on the critique of Foucault’s supposed excessive historicism.  Lorenzini’s contribution, to which I will return later, is to suggest that Foucault’s later writings (1980 and forward) distinguish between “games” of truth and “regimes” of truth. The distinction is basically illustrated in the following sentence: “I accept that x and y are true, therefore I ought to do z.”  The game of truth is the epistemic first half of the sentence, and the “regime” of truth – the part that governs human behavior – is the second half, the “therefore I ought…”  On this reading, genealogy is about unpacking and bringing to light the tendency of the “therefore” to disappear as we are governed by its regime, and to unpacking the power structures that make it operate.  In other words genealogy doesn’t collapse questions of truth and power; rather, it allows us to separate them by showing that a given game of truth does not entail the regime of truth that goes with it.

More on Covid Patents and global inequality

I wrote a piece a little more than a year ago about how intellectual property rights are getting in the way of global vaccine equity, condemning a lot of people to die in lower-income countries.  There have been various initiatives to address the situation.  A piece in Nature by Amy Maxmen today highlights somewhat encouraging news on this front in the form of a consortium of sites across Africa, Asia and South America working to develop mRNA vaccines and manufacturing capacity.  Part of the goal is to deal with Covid – but it’s also to redress the fact that many places are suffering from not just Covid, but tuberculosis, HIV, malaria and other problems.  Developing local capacity will help get vaccines to those who need them, as well as preparing researchers in these countries to respond to new pandemics.

Of course, this is a tremendous challenge, as Maxmen details  It should surprise no one that patents are getting in the way:

“Despite the hub’s efforts, next-generation mRNA vaccines might still be entangled in patent thickets if some components of the technology have been claimed by others. An astounding number of patents — estimated at more than 80 — surround the mRNA vaccines, according to one analysis. A thorny IP landscape isn’t as daunting for big companies with the capital to litigate, explains Tahir Amin, a lawyer and co-founder of the Initiative for Medicines, Access & Knowledge (I-MAK), a non-profit group based in New York City. Amin says that the hub could boldly move forward, too, and harness public condemnation if Moderna or other companies file a lawsuit. But this option is off the table because the Medicines Patent Pool vows not to infringe on patents. Indeed, the agency’s model relies on persuading pharmaceutical companies to voluntarily license their technologies to alternative manufacturers, often in exchange for royalty fees”

Moderna appears to be a particularly bad offender:

“Moderna did not respond to requests from Nature for comment. But in an interview with The Wall Street Journal, Moderna’s chief executive Stéphane Bancel said that the company won’t impede Afrigen’s work in South Africa; he made no mention of the 15 larger companies working with the hub. He added, “I don’t understand why, once we’re in an endemic setting when there’s plenty of vaccine and there’s no issue to supply vaccines, why we should not get rewarded for the things we invented.””

Maybe it’s time to remind everyone of the amount of public money that supported Moderna’s vaccine development, or that the company is on track to record $19 billion in revenue this year (Maxmen does both in the article).  Bancel’s language invokes tired and slippery uses of “endemic” to minimize Covid.  All “endemic” means is that a disease is spread at a more-or-less constant (and deemed acceptable) rate, without big surges.  Since the world is still suffering from big surges of Covid, Bance’s claim is straightforwardly false.  It’s also ignoring the moral and medical issue; if large numbers of people are dying from something (Covid, tuberculosis) it’s not acceptable, even if it comes at a predictably steady rate.

But even if the claim about endemicity were true, the rest of the sentence is still false.  Very few people in low-income countries have been vaccinated, which means that the claim that there’s “plenty of vaccine” is wrong.  It might be the case that there is plenty of vaccine here in the U.S., but if it’s not getting into the arms of Africans, then there by definition not “plenty” of vaccines there, and it is also an issue to supply them: the measure of success is vaccinations, not vaccines in a warehouse.  Finally, Bancel pulls the oldest trick in the IP maximalist’s arsenal: to equate the ability to extract monopoly rents with the ability to profit at all.  If you want to achieve social welfare, then pharma should get just enough IP protection to incentivize product development and not a penny more.

Overcompensating Pharma mints billionaires, but it’s still social murder.

New Paper: Epistemic Injustice in Machine Learning Systems

From the Department of Shameless Self-Promotion, here is the abstract for my new paper, "Dirty Data Labeled Dirt Cheap: Epistemic Injustice in Machine Learning Systems:"

"Artificial Intelligence (AI) and Machine Learning (ML) systems increasingly purport to deliver knowledge about people and the world or to assist people in doing so.  Unfortunately, they also seem to frequently present results that repeat or magnify biased treatment of racial and other vulnerable minorities, suggesting that they are “unfair” to members of those groups.  However, critique based on formal concepts of fairness seems increasingly unable to account for these problems, partly because it may well be impossible to simultaneously satisfy intuitively plausible operationalizations of the concept and partly because fairness fails to capture structural power asymmetries underlying the data AI systems learn from.  This paper proposes that at least some of the problems with AI’s treatment of minorities can be captured by the concept of epistemic injustice.  I argue that (1) pretrial detention systems and physiognomic AI systems commit testimonial injustice because their target variables reflect inaccurate and unjust proxies for what they claim to measure; (2) classification systems, such as facial recognition, commit hermeneutic injustice because their classification taxonomies, almost no matter how they are derived, reflect and perpetuate racial and other stereotypes; and (3) epistemic injustice better explains what is going wrong in these types of situations than does (un)fairness."

The path from idea to paper here was slow, but I hope the paper is convincing on the point that the literature on epistemic injustice can offer some needed resources for understanding harms caused by (some kinds of ) AI/algorithmic systems.

Just because the computer talks to you about its feelings doesn’t make it sentient

By Gordon Hull

UPDATE: 6/14: Here's a nice takedown ("Nonsense on Stilts") of the idea that AI can be sentient.

I don’t remember where I read about an early text-based chatbot named JULIA, but it was likely about 20 years ago. JULIA played a flirt, and managed to keep a college student in Florida flirting back for something like three days.  The comment in whatever I read was that it wasn’t clear if JULIA had passed a Turing test, or if the student had failed one.  I suppose this was inevitable, but it appears now that Google engineer Blake Lemoine is failing a Turing test, having convinced himself that the natural language processing (NLP) system LaMDA is “sentient.”

The WaPo article linked above includes discussion with Emily Bender and Margaret Mitchell, which his exactly correct, as they’re two of the lead authors (along with Timnit Gebru) on a paper (recall here) that reminds everyone that NLP is basically a string prediction task: it scrapes a ton of text from the Internet and whatever other sources are readily available, and gets good at predicting what is likely to come next, given a particular input text.  This is why there’s such concern about bias being built into NLP systems: if you get your text from Reddit, then for any given bit of text, what’s likely to come next is racist or sexist (or both).  The system may sound real, but it’s basically a stochastic parrot, as Bender, Gebru and Mitchell put it.

So point one: LaMDA is not sentient, any more than ELIZA and JULIA were sentient, but chatbots are getting pretty good at convincing people they are.  Still, it’s disturbing that the belief is spreading to people like Lemoine who really, really ought to know better.

Base Rate Tracking often can’t fix algorithmic fairness

By Gordon Hull

As a criterion for algorithmic assessment, “fairness” has encountered numerous problems.  Many of these emerged in the wake of ProPublica’s argument that Broward County’s pretrial detention system, COMPAS, was unfair to black suspects.  To recall: In 2016, ProPublica published an investigation piece criticizing Broward County, Florida’s use of a software program called COMPAS in its pretrial detention system.  COMPAS produced a recidivism risk score for each suspect, which could then be used in deciding whether someone should be detained prior to their trial.  ProPublica’s investigation found that, among suspects that did not have a rearrest prior to their trial, black suspects were much more likely to have been rated as “high risk” for rearrest than white suspects.  Conversely, among suspects who were arrested a second time, white suspects were more likely to have been labeled “low risk” than black ones.  The system thus appeared to be discriminating against black suspects.  The story led to an extensive debate (for an accessible summary with cites, see Ben Green’s discussion here) over how fairness should be understood in a machine learning context.

The debate basically showed that ProPublica focused on outcomes and demonstrated that the system failed to achieve separation fairness, which is met when all groups subject to the algorithm’s decisions receive the same false negative/positive rate.  The system failed because “high-risk” black suspects were much more likely than white to be false positives.  In response, the software vendor argued that the system made fair predictions because among those classified in the same way (high or low risk), both racial groups exhibited the predicted outcome at the same rate.  In other words, among those classified as “high risk,” there was no racial difference in how likely they were to actually be rearrested.  The algorithm thus satisfied the criterion of sufficiency fairness.  In the ensuing debate, computer scientists arrived at a proof that, except in very limited cases, it was impossible to simultaneously satisfy both separation and sufficiency fairness criteria.

In the meantime, on the philosophy side, Brian Hedden has argued that a provably fair algorithm could nonetheless be shown to potentially violate 11 of 12 possible fairness conditions.  In a response piece, Benjamin Eva showed the limits of the twelfth with a different test and proposed a new criterion:

Clearview settles Facial Recognition Case

Luke Stark argues that Facial recognition should be treated as the “plutonium of AI” – something so dangerous that it’s use should be carefully controlled and limited.  If you follow  the news, you’ll know that we’re currently treating it as the carbon dioxide of AI, a byproduct of profit-making that doesn’t look too awful on its own until you realize its buildup could very well cause something catastrophic to happen.  Activists have worried about this pending catastrophe for a while, but lots of big money supports facial recognition, so they have thrown up a smokescreen of distractions – in one case, Facebook denied that its phototagging software in fact recognized faces (!) – in order to lull everyone into accepting it.

One of the worst offenders is a secretive company called Clearview, whose business model is to scrape the web of all the pictures it can find and then sell the technology to law enforcement.  The company even has an international presence: in one disturbing instance, the Washington Post documents the use of its technology by Ukrainians to identify dead Russian soldiers by way of their Instagram and other social media accounts, and then sometimes to contact their families.  More generally, the Post revealed internal documents showing that the company' database is nearing 100 billion images and that "almost everyone in the world will be identifiable."  They're going all-in; the Post reports that "the company wants to expand beyond scanning faces for the police, saying in the presentation [obtained by the WP] that it could monitor 'gig economy' workers and is researching a number of new technologies that could identify someone based on how they walk, detect their location from a photo or scan their fingerprints from afar."

Clearview is also one of a cohort of companies that has been sued for violating Illinois’ Biometric Information Privacy Act (BIPA).  BIPA, uniquely among American laws, requires opt-in assent for companies to use people’s biometric information (the Facebook case is central to my argument in this paper (preprint here); for some blog-level discussion see here and here).  Of course, BIPA is a state-level law, so its protections do not automatically extend to anyone who lives outside of Illinois.  That’s why yesterday’s news of a settlement with the ACLU is really good news.  The Guardian reports:

Facial recognition startup Clearview AI has agreed to restrict the use of its massive collection of face images to settle allegations that it collected people’s photos without their consent.  The company in a legal filing Monday agreed to permanently stop selling access to its face database to private businesses or individuals around the US, putting a limit on what it can do with its ever-growing trove of billions of images pulled from social media and elsewhere on the internet. The settlement, which must be approved by a federal judge in Chicago, will end a lawsuit brought by the American Civil Liberties Union and other groups in 2020 over alleged violations of an Illinois digital privacy law. Clearview is also agreeing to stop making its database available to Illinois state government and local police departments for five years. The New York-based company will continue offering its services to federal agencies, such as US Immigration and Customs Enforcement, and to other law enforcement agencies and government contractors outside Illinois.

Of course, the company denies the allegations in the lawsuit, and insists that it was just in the process of rolling out a “consent-based” product.  Ok, sure!  This is still a win for privacy and for one of the very few pieces of legislation in the U.S. that has any chance of limiting the use of biometric data.

Yet another terrible AI idea

People make snap judgments about those they see the first time – mentally categorizing someone as friendly, threatening, trustworthy, etc.  Most of us know that those impressions are idiosyncratic, and suffused with cultural biases along race, gender and other lines.  So obviously I know what you’re thinking… we need an AI that do that, right?  At least that’s what this new PNAS paper seems to think (h/t Nico Osaka for the link).  The authors start right in with the significance:

“We quickly and irresistibly form impressions of what other people are like based solely on how their faces look. These impressions have real-life consequences ranging from hiring decisions to sentencing decisions. We model and visualize the perceptual bases of facial impressions in the most comprehensive fashion to date, producing photorealistic models of 34 perceived social and physical attributes (e.g., trustworthiness and age). These models leverage and demonstrate the utility of deep learning in face evaluation, allowing for 1) generation of an infinite number of faces that vary along these perceived attribute dimensions, 2) manipulation of any face photograph along these dimensions, and 3) prediction of the impressions any face image may evoke in the general (mostly White, North American) population”

Let’s maybe think for a minute, yes?  Because we know that people make these impressions on unsound bases!

First, adversarial networks are already able to produce fake faces that are indistinguishable from real ones.  Those fake faces can now be manipulated to appear more or less trustworthy, hostile, friendly, etc.  When you make fake political ads, for example, that’s going to be useful.  Already 6 years ago, one “Melvin Redick of Harrisburg, Pa., a friendly-looking American with a backward baseball cap and a young daughter, posted on Facebook a link to a brand-new website,” saying on June 8, 2016 that “these guys show hidden truth about Hillary Clinton, George Soros and other leaders of the US. Visit #DCLeaks website. It’s really interesting!” Of course, both Melvin Redick and the site he pointed to were complete fabrications by the Russians.  Now we can make Melvin look trustworthy, and Clinton less so.

Second, the ability to manipulate existing face photos is a disaster-in-waiting.  Again, we saw crude efforts with this before – making Obama appear darker than he is, for example.  But here news photos could be altered to make Vladimir Putin appear trustworthy, or Mr. Rogers untrustworthy.  This goes nowhere good, especially when combined with deepfake technology that already takes people out of their contexts and puts them in other ones (disproportionately, so far, women are pasted into porn videos, but the Russians recently tried to produce a deepfake of Zelensky surrendering.  Fortunately that one was done sloppily).

Third, and I think this one is possibly the scariest, what about scanning images to see whether someone will be assessed as trustworthy?  AI-based hiring is already under-regulated!  Now employers will run your photo through the software and make customer-service hiring decisions based on who customers will perceive as trustworthy.  What could go wrong?

All of this of course assumes that this sort of software actually works.  The history of physiognomic AI, which uses all sorts of supposedly objective cues to determine personality and which is basically complete (usually racist) bunk suggests that the science is probably not as good as the article acts like.  So maybe we’re lucky and this algorithm does not actually work as advertised.  Of course, the fact that AI software is garbage doesn't preclude its being used to make people's lives miserable.  Just consider the bizarre case of VibraImage.

But don’t worry.  The PNAS authors are aware of ethics, noting that “the framework developed here adds significantly to the ethical concerns that already enshroud image manipulation software:”

“Our model can induce (perceived) changes within the individual’s face itself and may be difficult to detect when applied subtly enough. We argue that such methods (as well as their implementations and supporting data) should be made transparent from the start, such that the community can develop robust detection and defense protocols to accompany the technology, as they have done, for example, in developing highly accurate image forensics techniques to detect synthetic faces generated by SG2. More generally, to the extent that improper use of the image manipulation techniques described here is not covered by existing defamation law, it is appropriate to consider ways to limit use of these technologies through regulatory frameworks proposed in the broader context of face-recognition technologies.” 

Yes, the very effective American regulation of privacy does inspire confidence!  Also, “There is also potential for our data and models to perpetuate the biases they measure, which are first impressions of the population under study and have no necessary correspondence to the actual identities, attitudes, or competencies of people whom the images resemble or depict.”

Do you think?  As Luke Stark put it, facial recognition is the “plutonium of AI:” very dangerous and with very few legitimate uses.  This algorithm belongs in the same category, and should similarly be regulated like nuclear waste.  For example, as Ari Waldman and Mary Anne Franks have written, one of the problems with deepfakes is that the fake version gets out there on the internet, and it is nearly impossible to make it go away (if you even know about it).  Forensic software gets there too late, and those without resources aren’t going to be able to deploy it anyway.  Lawsuits are even less useful, since they're time-consuming and expensive to pursue, and lots of defendants won't be jurisdictionally available or have pockets deep enough to make the chase worth it.  In other words, not everybody is going to be able to defend themselves like Zelensky, who both warned about deepfakes and was able to produce video of himself not surrendering.  In the meantime, faked and shocking things generally get diffused faster and further than real news.  After all, “engagement” is the business model of social media.  Further, to the extent that people stay inside filter bubbles (like Fox News), they may never see the forensic corrections, and they probably won’t believe the real one is real, even if they do.

And as for reinforcing existing biases, Safiya Noble already wrote a whole book on how algorithms that guess what you’re probably thinking about someone can do just that.