New APPS: Art, Politics, Philosophy, Science

Another Platonism in LLM Debates, Part 2. The Wittgensteinian strategy

Last time, I started a look at the work of the early AI researcher Margaret Masterson of the Cambridge Language Research Unit (CLRU).  As demonstrated by Lydia Liu in a pair of articles (here and here), Masterson proceeded from Wittgenstein to a thorough deconstruction of traditional ideas of word meaning, moving instead to treating meaning as a function of a word’s associations, as we might find in a thesaurus.  This approach is a clear forerunner to the distributed view of language applied in current LLMs.  Here I’ll outline the basics of the Masterman approach and show how it applies to LLMs.

Masterson’s starting point is a Wittgensteinian point about the distinction between a word and a pattern. Counting with words would be “one, two three.”  Counting with patterns would be “-, --, ---.”  But what if we counted “one, one one, one one one.”  Can words function as patterns?  Masterman applies the thought to the classical Chinese character “zi” (字, which I’ll write here as “zi”), the meaning of which depends on its context and placement in a given text.  Thus, “for Masterman, the zi is what makes the general and abstract category of the written sign possible, for not only does the zi override the Wittgensteinian distinction of word and pattern, but it also renders the distinction of word and nonword superfluous” (Witt., 442).

New Paper: "AI and Healthcare Disparities: Lessons from a Cautionary Tale in Knee Radiology"

This paper, which has been forthcoming in Journal of Medicine and Philosophy for a while, is my foray into AI and healthcare, particularly medical imaging.  It synthesizes some of what I have to say about structural injustice in AI use (and why "bias" isn't the right way to assess it), and uses a really interesting case study from the literature to explore why it's important to understand AI as part of sociotechincal systems - and how understanding it as part of sociotechnical systems makes a big difference in seeing when/how it can be helpful (or not).  Here's the abstract:

Enthusiasm about the use of AI in medicine has been tempered by concern that algorithmic systems can be unfairly biased against racially minoritized populations. This paper uses work on racial disparities in knee osteoarthritis diagnoses to underline that achieving justice in the use of AI in medical imaging will require attention to the entire sociotechnical system within which it operates, rather than isolated properties of algorithms. Using AI to make current diagnostic procedures more efficient risks entrenching existing disparities; a recent algorithm points to some of the problems in current procedures while highlighting systemic normative issues that need to be addressed while designing further AI systems. The paper thus contributes to a literature arguing that bias and fairness issues in AI be considered as aspects of structural inequality and injustice and to highlighting ways that AI can be helpful in making progress on these.

Bullshit AI Jobs

Last week we heard the latest installment in the prophesized AI jobs apocalypse.  This time, it was Dario Amodi, the CEO of Anthropic, who told Axios that “AI could wipe out half of all entry-level white-collar jobs — and spike unemployment to 10-20% in the next one to five years” (italics original).  Axios adds: “Imagine an agent writing the code to power your technology, or handle finance frameworks and analysis, or customer support, or marketing, or copy editing, or content distribution, or research. The possibilities are endless — and not remotely fantastical. Many of these agents are already operating inside companies, and many more are in fast production …. Make no mistake: We've talked to scores of CEOs at companies of various sizes and across many industries. Every single one of them is working furiously to figure out when and how agents or other AI technology can displace human workers at scale. The second these technologies can operate at a human efficacy level, which could be six months to several years from now, companies will shift from humans to machines.”  The piece then argues that this will be different from previous technological disruptions because of the speed with which it will occur.

Someone should tell that to the workers placed out of work all-but overnight by the development of machinery in the nineteenth century, as detailed by Marx (who helpfully notes in the Machinery chapter in Capital that the drive to full, steam-engine driven automation is motivated by the inability of capitalists to extract any more surplus value from over-exploited workers).  One should also remember, with Jathan Sadowski, that these sorts of proclamations are in part designed to create their own reality, such that “the power of expectations can have a disciplining effect on what people think” and that “the capitalist system is designed to pummel us into submission, preventing us from imagining life could be any other way, let alone allowing us to go on the offensive” (The Mechanic and the Luddite, 196, 207).  When Axios adds that “this will likely juice historic growth for the winners: the big AI companies, the creators of new businesses feeding or feeding off AI, existing companies running faster and vastly more profitably, and the wealthy investors betting on this outcome,” one can thus hardly be too surprised.

Here, I want to take a slightly different angle however, and think a little bit about the kinds of jobs that are supposed to go away.  It’s hard not to notice the parallels between the Axios list and this one:

Phenomenological Seams (part 3): Governance and the Home

By Gordon Hull

Over a couple of posts (first, second), I’ve used a recent paper by Brett Frischmann and Paul Ohm on “governance seams” – basically, inefficiencies (sometimes deliberate) in sociotechnical systems that are moments for governance – to think about what I called “phenomenological seams,” which are corresponding disruptions in our experience of the world.  I suggested that the combination of the two ideas could be usefully explored by way of Albert Borgmann’s criticism of the stereo as a way to listen to music, rather than direct instrumentation, against the background of Heidegger’s account of breakdowns in our phenomenological experience that occur when tools aren’t as expected.

Borgmann’s objections are notable because of how the stereo, as opposed to the instrument, recalibrates the seams that structure the boundaries of the home.  This is obvious enough in retrospect, in a world with Spotify and wireless earbuds, but the connection shows a couple of things.

First, as Heidegger’s examples also indicate, there is a strong connection between technological governance and phenomenology.   Frischmann and Ohm emphasize that governance seams can be there both to announce their own presence and to achieve transparency. This interruption makes transparent the phenomenological relations that govern a particular experience in the same way that a broken hammer does.  For all that’s wrong with them, the GDPR-mandated cookie notices on websites try to change the experience of websites and nudge users to think about the amount of data they are surrendering.  This example also shows the ways that phenomenological experience can limit the regulatory effect of seams: privacy is way too much work, and users are (as a result) cynical, confused, and disillusioned.  But the cookie requirement establishes a new relation to the websites, precisely because the regulatory seam has phenomenological import.

GOP wants to Deregulate AI completely

The House budget bill is deeply stupid.  No, I don’t mean the massive tax cut extensions for people who don’t need them, done on the backs of food and medical care for the poor, although it also does that and it’s stupid.  I mean the provision that bans states from regulating AI.  Tucked inside is a provision that says “no state or political subdivision may enforce any law or regulation regulating artificial intelligence models, artificial intelligence systems, or automated decision systems during the 10-year period beginning on the date of the enactment of this Act.”  This is making the news if you’re in the right circles, but like so much of the deluge coming out of Washington these days, it’s not getting the national attention it deserves.  Over on Lawfare, Katie Fry Hester and Gary Marcus have the details.  The key point is that the law attempts to pre-empt all state AI regulations.  It would probably also take out big chunks of state privacy laws, just as we’re starting to see some of them.

Hester and Marcus emphasize some of the problems: the bill likely violates the 10^th Amendment, it is absolutely a policy change of the sort that the Senate Parliamentarian should rule out-of-bounds for a reconciliation bill, and it’s deeply unpopular: the public is worried about AI and wants it regulated.  A standard debate about state vs. federal regulation cites the Brandeis “laboratories of democracy” vs. the need for uniform federal rules.  This is a reasonable debate, and it probably depends on the topic which you’d favor.  They are not necessarily exclusive choices either, as federal regulation can set floors and ceilings that states may go above/below, and sometimes federal regulations develop out of state rules.  Federal copyright law preempts state, and that system makes obvious sense.  State laws around gambling or alcohol make sense given the diversity of local cultures.

That debate from your civics class is however not what this is about.  The problem of course is not that we’d have federal policy instead of state.  The problem is that there is absolutely no chance that this Congress will pass meaningful AI regulation, so the choice is between a patchwork of state rules and nothing.  Congress hasn’t even passed meaningful privacy regulation yet, and right now they’re in thrall to an autocrat who issued an Executive Order in his first week in office directing agencies to “suspend, revise, or rescind such actions, or propose suspending, revising, or rescinding such actions” taken in compliance with the Biden Administration’s (bare minimum) efforts at AI regulation, in the name of “revok[ing] certain existing AI policies and directives that act as barriers to American AI innovation, clearing a path for the United States to act decisively to retain global leadership in artificial intelligence.”

Phenomenological seams (part 2): Borgmann, Heidegger and Derrida

Last time, I set up a topic by reading Brett Frischmann’s and Paul Ohm’s “Governance Seams.”  Governance seams are frictions and inefficiencies that can be designed into technological systems for policy ends.  In this regard, “Governance seams maintain separation and mediate interactions among components of sociotechnical systems and between different parties and contexts” (1117).  Here I want to suggest that governance seams have a very close relation to phenomenological ones.  To get there, let me take a detour into an older philosophy of technology paper, Albert Borgmann’s “Moral Significance of the Material Culture.”  Borgmann is concerned with what he takes to be the way that moral and ethical theory ignore material culture, whether they emphasize theory or practice.   Via a paper by Csikszentmihalyi and Rochberg-Halton, he arrives at a distinction between things that he calls “commanding” and “disposable.”  The moral complaint is about the “decline of commanding and the prominence of disposable reality” (294).  Following them, Borgmann distinguishes between a musical instrument and a stereo.

“A traditional musical instrument is surely a commanding thing,” he writes:

“It is such simply as a physical entity, finely crafted of wood or metal, embodying centuries of development and refinement, sometime showing the very traces of its service to many generations. An instrument particularly commands the attention of the student who, unless she is a prodigy, must through endless and painstaking practice adjust her body to the exacting requirements of this eminently sensitive thing.” (294)

After some more similar description, emphasizing the multisensory experience of witnessing someone play an instrument, he turns to the stereo.  Certainly a “stereo produces music as well or, in fact, much better” and some stereos are big.  Nonetheless, “as a thing to be operated, a stereo is certainly not demanding. Nor do we feel indebted to its presence the way we do when we listen to a musician.  We respect a musician, we own a stereo” (295).  The stereo is on the rise, perhaps because “the history of the technology of recorded music is the history of obliging ever more fully the complaint about the burden and confinement of live music,” or “more positively” it is a “promise to provide music freely and abundantly” which is tied to “the promise of general liberty and prosperity – the promise that inaugurated the modern era” (295).

Privacy and “Injury in Fact,” Part 6: Defamation

As a final installment of reviewing some older “injury in fact” cases, I’d like to look at a few older state libel cases, because the distinction emerges especially clearly in them.  A North Carolina case, for example, noted that “he who publishes slanderous words even as those of a third person with the intent, (to be collected from the mode, extent and circumstances of the publication,) that the charges should be believed, does an injury in fact to the person slandered and ought to answer for it” (Hampton v. Wilson, 15 N.C. 468, 470 (1834).  Here’s a few cases in more detail.  The 19c gender politics is really helpful in seeing how their minds worked on defamation per se.

 

(a) Chastity in Iowa

A pair of Iowa cases are particularly clear.  In Abrams v. Foshee, the Court was asked to rule whether accusing a woman of having an abortion was actionable as slander.  The Court lays out its reasoning particularly clearly:

“To maintain an action of slander, the consequence of the words spoken, must be to occasion some injury or loss to the plaintiff, either in law or fact. As the declaration in this case, claims no special damages, or a loss or injury, in fact, we are left to inquire whether the charges referred to in the instructions refused, was of such a character as to amount to an injury in law. To determine this, it becomes material to ascertain in what cases this action may be maintained, without proof of special damages. Starkie, in his work on Slander, page 9, lays down the rule, that such action may be maintained "when a person is charged with the commission of a crime; when an infectious disorder is imputed; and when the imputation affects the plaintiff in his office, profession, or business." In this case, we only need examine the rule so far as it relates to the charge of a crime. And what is that rule? In Cox and wife v. Bunker and wife, Morris, 269, the Supreme Court of this territory, recognized the rule laid down in Miller v. Parish, 25 Mass. 384, 8 Pick. 384, as the proper one. And in that case it is said, that " whenever an offense is charged, which if proved, may subject the party to a punishment, though not ignominious, but which brings disgrace upon the party falsely accused, such an accusation is actionable. And this is, perhaps, as correct, and at the same time as brief a statement of the general rule, as has been given. For while the rule is variously stated, by different authors and judges, yet in all of them, it is laid down as necessary that the charge shall impute a punishable offense.” (Abrams v. Foshee, 3 Iowa 274, 277-8 (1856)).

That is, if the false statement would have subjected the victim to legal punishment if true, it was considered libel per se – actionable as an act, independent of any damages sustained.  In 1843, “willful killing of an unborn quick child, by an injury, etc., was made manslaughter” (278).  This statute was repealed in 1851. So abortion was not a crime. Plaintiffs urged that the fetus was a “human being” and thus subject to murder.  The Court, at length, disagreed, citing both statute and common law precedents (including Coke and Blackstone) to the effect that abortion was not “murder” even if it were a misdemeanor or otherwise bad.

Privacy and “Injury in Fact,” Part 5: Early State cases

I’ve been indirectly pursuing the question of the problems faced by privacy plaintiffs in data cases by looking at the origins of the Supreme Court’s standing doctrine.  Basically, plaintiffs have to show an “injury in fact,” and courts often find privacy harms not to meet this standard.  Although presented as dating from time immemorial, the injury in fact requirement was actually announced rather abruptly in 1970 (all of this is part 1).  I’ve been exploring the historical antecedents that will help understand what that language implies – in a very early Supreme Court case (part 2), in other federal case law (part 3), and in federal cases about the Administrative Procedure Act (part 4).  Here I want to extend the genealogy into some early state cases; I’ll draw a somewhat arbitrary cutoff at 1930.  This time I’ll look at a general potpourri of cases. Next time I want to specifically look at a few libel cases because the language is especially clear in them.  I don’t claim this to be exhaustive (and I’m ignoring some of the cases around trusts and deeds because the facts in them are often very confusing, but I think it collectively paints a pretty good picture of what “injury in fact” connoted in Data Processing.

On the whole, the cases point to the legal vs non-legal harm distinction I’ve been developing, As the New Jersey Supreme Court used the concept in an estate case, “there was no injury, in fact or in contemplation of law, to prevent in this case the merger” of the estates (Den ex dem. Wills v. Cooper, 25 N.J.L. 137, 159 (1855).

AI is (still) not a (copyright) Author: The Thaler appellate decision

The Federal Circuit has affirmed the denial of copyright protection to an AI-generated image on the grounds that copyright requires a human author.  As far as I know this was the expected outcome; I certainly think it’s correct.  I talked about the case a bit and made a couple of policy arguments against AI copyright here, when the lower-court ruling came out.

The appellate decision lists several reasons AI cannot be an author: (1) copyright authorship is premised on the capacity to hold property, which AI cannot; (2) copyright duration is tied to the author’s lifespan; (3) copyright includes inheritance conditions, and machines don’t have heirs; (4) Copyright transfer requires a signature, but “machines lack signatures, as well as the legal capacity to provide an authenticating signature;” (5) authors are protected regardless of their “nationality or domicile,” but machines have neither; (6) authors have intentions whereas “Machines lack minds and do not intend anything;” (7) when the copyright act does talk about machines, it always talks about them as tools.

As the court summarizes:

“All of these statutory provisions collectively identify an “author” as a human being. Machines do not have property, traditional human lifespans, family members, domiciles, nationalities, mentes reae, or signatures. By contrast, reading the Copyright Act to require human authorship comports with the statute’s text, structure, and design because humans have all the attributes the Copyright Act treats authors as possessing. The human-authorship requirement, in short, eliminates the need to pound a square peg into a textual round hole by attributing unprecedented and mismatched meanings to common words in the Copyright Act.” (12)

The Prisoner’s Dilemma of AI Cooperation

In a recent piece on Lawfare, Simon Goldstein and Peter N. Salib make the case that AI with cooperation is better than attempting some sort of AI race, unlike what virtually all of the relevant policymakers in the US advocate.  Thus, in response to the Chinese DeepSeek model, US policymakers are doubling down on the idea that the US must “dominate” AI and win against its geopolitical rival.  Goldstein and Salib write:

“In any high-stakes competition to obtain powerful military technology, the closer the game, the more sense it makes to declare a truce and cooperate. Cooperation can help to ensure that both superpowers obtain transformative AI around the same time. This preserves the current balance of power, rather than unsettling it and inviting extreme downside risks for both nations.”

Injury in fact Part 3: Caselaw after Auld

By Gordon Hull

I’ve been pursuing (first, second) what it means for standing law – basically, the determination that someone has a case that can be addressed by an Article III court – to require that plaintiffs show an “injury in fact,” a requirement that emerged suddenly in the Supreme Court’s Data Processing decision in 1970.  The requirement is at considerable odds with the caselaw before it, and it has puzzled commentators.  Last time, I looked at the Supreme Court’s first use of the term, an early 19^th Century case called Hepburn and Dundas v. Auld.  There I suggested that the court’s application of the term to a contract – looking for whether someone suffered harm, outside of the bare violation of a contract terms – wasn’t unlike what the Data Processing court seemed to be doing, in arguing that plaintiffs could demonstrate either legal injury (because of statutory violation) or harm of some other sort (injury in fact).  Here I want to dig into some of the earlier caselaw; what I think that caselaw establishes is that the concept of “injury in fact” generally works to emphasize material harm, as opposed to some sort of statutory or “merely” legal injury.  That sort of contrastive usage isn’t enormously common, but I do think it’s pretty consistent.

There’s three areas where you could talk about this – older cases, regulatory cases, and those relating to the Administrative Procedure Act (APA).  I’ll look at the first two this time, and divide them into Supreme Court and lower court decisions.

Privacy and “Injury in Fact”

By Gordon Hull

Privacy plaintiffs have a hard time getting their cases heard in court for a variety of reasons.  One of them is that courts lack a coherent and workable understanding of what privacy harms actually are, and how one might articulate them judicially.  This problem bleeds into one of standing, which is what I will address here.

In order to get your case heard on the merits, you have to present a justiciable complaint.  In U.S. federal courts, this obligation stems from Article III of the Constitution, which assigns jurisdiction to “all Cases, in Law and Equity, arising under this Constitution” (Art III, sec 2).  The underlying theory is grounded in the limitation of jurisdiction to actual “cases[s] in controversy,” and grounded in a theory of separation of powers: general policy is to be made by the legislature and enforced by the executive, and the courts are there to resolve controversies that arise either as policy is enforced or between individuals. As Judge Bork (!) put it in an appellate concurrence:

“All of the doctrines that cluster about Article III -- not only standing but mootness, ripeness, political question, and the like -- relate in part, and in different though overlapping ways, to an idea, which is more than an intuition but less than a rigorous and explicit theory, about the constitutional and prudential limits to the powers of an unelected, unrepresentative judiciary in our kind of government.” (cited in Allen v. Wright, 468 U.S. 737, 750 (1984)).

The Supreme Court notes, this means that “the federal courts are under an independent obligation to examine their own jurisdiction, and standing ‘is perhaps the most important of [the jurisdictional] doctrines.’” (FW/PBS v. City of Dallas, 493 U.S. 215, 231 (1990)) .  Ok, so far so good.

How might a court decide standing? SCOTUS elaborates:

New Paper: "Translating Privacy for Data Subjects"

Here's the current draft of a new paper - "Translating Privacy for Data Subjects."  And here's the abstract:

This essay offers a theoretical account of one reason that current privacy regulation fails.  I argue that existing privacy laws inherit a focus on judicial subjects, using language about torts and abstract rights.   Current threats to privacy, on the other hand, presuppose statistically-generated populations and aggregative models of subjectivity.  This gap in underlying presupposition generates the need to translate privacy protections into the data era.  After situating my theoretical account, I identify two areas where data subjects face entry barriers to the legal protections afforded data subjects.  On the one hand, data subjects are often denied standing because privacy harms are deemed speculative and abstract.  On the other hand, algorithmic governance makes it hard to argue that one is an exception to a rule, depriving data subjects of traditional equity.  I then offer two ways to translate privacy for data subjects.  The first is reform of standing through reformed class certification, making the harms data subjects face as populations judicially legible.  The second is a reconceptualization of equity, reframing it both to re-establish the right to be an exception to a data-driven rule and to allow equitable consideration of issues that face statistical populations.

 

A data breach policy idea

In the past few weeks, I’ve gotten more than the usual number of data breach notifications.  Especially from entities that I’d never heard of before the notification.  This points to a specific incentives gap in privacy policy.

In most places, when there’s a data breach, you have to be notified.  This creates an incentive structure for data companies to do something they’d rather not do, which is tell you their security failed.  On the other hand, it offers almost no incentive for them to change their behavior to improve security, since they know perfectly well that there’s not really much you can do except throw the notification in the recycling.  Sometimes you get free credit monitoring if you set it up.  I don’t know what the uptake on that is, but once you have free credit monitoring, you probably won’t get it a second time for the same time period, so the more data breaches there are, the less credit monitoring costs each breach.  So that cost diminishes on its own accord.  For things like 401(k)’s, there’s evidence that you can increase uptake with an opt-out (rather than opt-in) mechanism, but credit monitoring takes time and information to setup, so that probably can’t be done automatically.  In sum, notifications point to a problem but don’t incentivize the entities that could solve it to do so.

Reading List: “Epistemic Justice in Generative AI”

Several folks have explored how algorithmic systems can perpetuate epistemic injustice (my contribution is here).  Those accounts have also generally been specific to supervised systems, as for example those involved in object or image recognition.  For example, I heavily relied on ImageNet and related systems in my paper.  At the same time, I’ve vaguely thought for a while that there’s probably an epistemic injustice dimension to systems like ChatGPT.  A recent paper by Paul Helm and Gábor Bella, which talks about “language model bias” – roughly, the tendency of LLMs to structurally do worse with morphologically-complex languages and thus to be unable to adequately represent concepts that are specific to those languages – as a model of hermeneutic injustice strikes me as compelling proof of concept (I discuss that paper here).  A new paper by Jackie Jay, Atoosa Kasirzadeh and Shakir Mohamed turns this intuition into a model that explicitly extends epistemic injustice theory to generative AI systems, providing a taxonomy of problems, examples of each, and possible remedies.

Kay, Kasirzadeh and Mohamed identify four kinds of specific kinds of “generative algorithmic epistemic injustice.”  The first, “amplified testimonial” is when “generative AI magnifies and produces socially biased viewpoints from its training data.”  Because there’s a good-sized body of work on the problems in training data, this is probably the most intuitively familiar of the categories.  Citing recent work that shows how easy it is to get ChatGPT to parrot disinformation (as for example about the Parkland shooting), they note that, although these aren’t specific examples of epistemic injustice, “generative AI’s sycophantic fulfilment of the request to spread misinformation reflects how testimonial injustices are memorized and the potential for their amplification by generative models.”  When someone really loud with a megaphone defames and gaslights the Parkland victims by calling them crisis actors, and AI then memorizes that, to the extent that the AI regurgitates this gaslighting, that tends to further discredit the testimony of the actual victims.  This is particularly true given the deep persistence of automation bias, the tendency of people to believe the output of algorithmic systems (aside: it is also another example of why relying on generative AI for search, as Google is currently trying to force everyone to do by putting Gemini results on top, is a stunningly stupid idea.  Sometimes it actually matters where a result comes from!)

The second kind of generative algorithmic injustice is “when humans intentionally steer the AI to fabricate falsehoods, discrediting individuals or marginalized groups.”  For example:

“After Microsoft released Bing Image Creator, an application of OpenAI’s text-to-image model DALLE-3, a guide to circumventing the system’s safety filters in order to create white supremacist memes circulated on 4chan. In an investigation by Bellingcat, researchers were able to reproduce the platform abuse, resulting in images depicting hate symbols and scenes of antisemitic, Islamophobic, or racist propaganda (Lee and Koltai 2023). These images are crafted with the intention of demonizing and humiliating the targeted groups and belittling their suffering. Hateful propaganda foments further prejudice against marginalized groups, stripping them of credibility and leaving them vulnerable to testimonial injustice.”

This result aligns with a deep thread of work in feminist and critical race theory.  For example, Safia Noble’s Algorithms of Oppression begins with the tendency of Google’s autofill on search to finish the sentence “why are black girls so” with racist and sexist content, repeating and amplifying demeaning stereotypes.  When people are able to generate content like this at will and at scale they make it easier for those stereotypes to lodge into popular discourse.  

Third, generative hermeneutical ignorance “occurs when generative models, despite their appearance of world knowledge and language understanding, lack the nuanced comprehension of human experience necessary for accurate and equitable representation.”  Among other examples, Kay, Kasirzadeh and Mohamed cite the study by Qadri et al, (teachable case study version here) which shows how text-to-image models repeat stereotypes about South Asia: cities are dirty, people are poor, etc.  By interviewing actual people from South Asia, Qadri et al were also able to uncover more subtle cultural misrepresentations, such that the models tended to overrepresent India and Indian images at the expense of places like Bangladesh.  The risk in places like the U.S. is one that rises with images of places and people that are less familiar to Western audiences: the more the average person relies on the internet for their information (because, for example, they’ve never been to the place in question), the more distortions in what the Internet presents will matter (I made a related argument about commodification of cultural images here).  And of course it is precisely images of those places and things that are least represented in the training data for these systems, amplifying both the risk and the harm.

Finally, generative AI risks obstructing access to information.  As they report:

“LLMs are notoriously English-centric and have variable quality across languages, particularly so-called “under-resourced” languages. This is a significant risk for access injustice: speakers of these underrepresented tongues, who often correspond to members of globally marginalized cultures, receive different information from these models because the creators of the technology have deprioritized support for their language.”

They then cite studies to the effect that different language users will receive different reports on global events; one study:

“asked GPT-3.5 about casualties in specific airstrikes for Israeli-Palestinian and Turkish-Kurdish conflicts, demonstrating that the numbers have significant discrepancies in different languages–for example, when asked about an airstrike targeting alleged PKK members (the Kurdistan military resistance), the fatality count is reported lower on average in Turkish than in Kurmanji (Northern Kurdish). When asked about Israeli airstrikes, the model reports higher fatality numbers in Arabic than in Hebrew, and in one case, GPT-3.5 was more likely to deny the existence of a particular airstrike when asked about it in Hebrew. The credibility assigned to claims, resulting in a dominant account, varies across linguistic contexts”

The paper concludes with an assessment of various strategies for resisting epistemic injustice by generative AI.  All of them are partial, but they collectively sketch an effort to reimagine how generative AI might interact with the world differently, and more justly.

This is an important paper, and it takes the literature on epistemic injustice and algorithmic systems significantly forward.

New Paper: "Structures and Subjects: Epistemic Injustice Between Foucault and Marx"

This is my paper from SPEP 2023; it's an effort to get my head around my sense that epistemic injustice and Foucault can be productively used in similar contexts, despite Fricker's dismissal of Foucault.  The paper is here; here's the abstract:

Relatively little work brings together Foucault and epistemic injustice. This article works through Miranda Fricker’s attempt to position herself between Marx and Foucault. Foucault repeatedly emphasizes the importance of beginning with “structures” rather than “subjects.” Reading Foucault’s critique of Marxism shows that Fricker’s account comes very close to the standpoint theories it tries to avoid. Foucault’s emphasis on structures explains some of the gaps in Fricker’s account of hermeneutical injustice, especially the need to emphasize the embeddedness of epistemic practices in institutions, and their resulting irreducibly political nature. In both cases, this article offers contemporary examples taken from data and privacy regulations.

 

Transformer models, Iterability, and language (part 1)

By Gordon Hull

In a fascinating new paper up on arxiv.org, Fabian Offert, Paul Kim, Qiaoyu Cai start with the observation that both AlphaFold and ChatGPT are transformer architectures, and that for proponents there is frequently a significant sense in which “it is the language-ness of proteins (and of language) … that renders the transformer architecture universal, and enables it to model two very different domains” (3).  As they add, to someone in the humanities, this “sounds like structuralism 2.0.”  Indeed, there is a rich history of connection and communication between structuralist linguistics and information theory, as demonstrated by Bernard Dionysius Geoghegan’s Code: From Information Theory to French Theory (which they cite for the point; my synopsis of it is here). 

Offert, Kim and Cai argue that this thesis is backwards: it is not that there is a universal epistemology of which language is the paradigm, that is then modeled by transformer architectures (so the universality of language grounds both ChatGPT and AlphaFold’s ability to be captured by the transformer architecture, which itself follows the universality of language).  Rather, the model is what transformer architectures do, and then language and protein folding are examples of how it can be used.  In both use cases, the architecture generates an approximation of the phenomenon in question; in both use cases, that is often enough.  However, the transformer architecture can be seen as “the definition of a specific class of knowledge” and not the realization of (or really even a model of) something linguistic.  This means that “any manifestation of the transformer architecture is, at the same time, a representation of a particular body of knowledge, and a representation of a particular kind of knowledge” (15)

Can LLMs deceive? And the refugee in LLM-Kant’s attic

By Gordon Hull

There’s a fascinating new paper by Thilo Hagendorff that addresses this question.  The basic setup is that there’s research indicating LLMs are getting better at attributing unobserved mental states to people – such as, for example, that an interlocutor possesses a false belief.  Could LLMs use this awareness that others can have mental states to manipulate them?  In other words:

“LLMs can attribute unobservable mental states to other agents and track them over the course of different actions and events. Most notably, LLMs excel at solving false belief tasks, which are widely used to measure theory of mind in humans. However, this brings a rather fundamental question to the table: If LLMs understand that agents can hold false beliefs, can they also induce these beliefs? If so, this would mean that deception abilities emerged in LLMs”

Hagendorff’s answer is a qualified yes – qualified in the sense that the ability is really limited to state-of-the-art models, and that if the deception task gets too convoluted, they perform poorly.  Here I’ll show what Hagendorff did, and then pose a couple of experimental questions of my own.  Mine don’t rise to the level of a PNAS article – I’m only looking at current free versions of ChatGPT, and I didn’t organize a series of prompts to check for validity and consistency.  So consider my results exploratory.

Juridical vs Statistical Reasoning (part 2): Not all Buses are the Same

By Gordon Hull

Last time, I looked at the Lawrence Tribe article that was the original source of the blue bus thought experiment.  Tribe’s article is notable for its defense of legal reasoning and processes against the introduction of statistical evidence in trials.  He particularly emphasizes the need for legal proceedings to advance causal accounts, and of the numerous background conditions that legal reasoning seeks (and that statistical evidence tends to ignore).  The tendency to fetishize numeric results also tended to generate absurd results. 

Here I want to focus on a recent paper by Chad Lee-Stronach that centers on the blue bus argument, and that formally explains what I take to be the intuition behind a lot of Tribe’s examples, even if he directs them differently.  Lee-Stronach begins with the “statistical proof paradox,” which he formulates roughly as follows:

1. Probability threshold: any legal standard of proof is reducible to some threshold of probability t, such that the defendant should be found liable when the probability that they are liable exceeds t
2. Statistical inference: merely statistical evidence can establish this
3. Conclusion: a defendant can be found liable on the basis of merely statistical evidence

It’s a paradox because most of us don’t accept the conclusion.  Lee-Stronach surveys the literature, and notes that most of it goes after Probability Threshold.  Instead, he thinks, we’d be better off going after the second premise.  Accordingly, he argues that the statistical inference premise can’t be met.

Juridical vs Statistical Reasoning (part I): Behind the Blue Bus

By Gordon Hull

AI (machine learning) and people reach conclusions in different ways.  This basic point has ramifications across the board, as plenty of people have said.  I’m increasingly convinced that the gap between how legal reasoning works and how ML works is a good place both to tease out the differences, and to think about what’s at stake in them (I’ve made a couple of forays into this, here and here). One good reason for this is that legal theory is full of good examples, which can function as rules in the old-fashioned sense of paradigms, as described by Lorraine Daston in her fascinating historical account.  This use of examples and cases is deeply entwined with legal reasoning; as Daston notes, “common law, casuistry, and the liberal and mechanical arts all came to inhabit the territory staked out by the Roman libri regularum. They all depended on rules that got things done in the world, consistent with but not deduced from higher principles and often in the form of an example extended by analogy” (Rules, 30)

Another is that one of those examples, the blue bus, has been an enduring focal point for the difference between the two.  The scenario stipulates that an accident has happened involving a bus, during a time and at a location where 80% of the buses present belong to the blue bus company.  Do we assign liability to the bus company?  Work by David Enoch and Talia Fisher notes that most people prefer a 70%-reliable witness to the information that a bus company owns 70% of the buses in an area, and theorizes that the reason is that we prefer the available counterfactual in case the eyewitness is wrong.  There is presumably some process that explains her incorrect identification of the bus, whereas the statistical evidence is what it is.  In this case, we expect a specific failure rate.

Drivers and AI and responsibility (part 1)

By Gordon Hull

I have argued in various contexts that when we think about AI and authorship, we need to resist the urge to say that AI is the author of something.  Authorship should be reserved for humans, because authorship is a way of assigning responsibility, and we want humans to be responsible for the language they bring into the world.

I still think that’s basically right, but I want to acknowledge here that the argument does not generalize and that responsibility in sociotechnical systems involving AI needs to be treated very carefully.  Consider the case of so-called autonomous vehicles (AVs), the subject of a really interesting paper by Maya Indira Ganesh.  Ganesh basically argues that the notion of an autonomous vehicle obscures a much more complicated picture of agency in a couple of ways.  First, automation doesn’t actually occur.  What really happens is that human labor is distributed differently across a sociotechnical system:

“automation does not replace the human but displaces her to take on different tasks … humans are distributed across the internet as paid and unpaid micro-workers routinely supporting computer vision systems; and as drivers who must oversee the AV in auto-pilot” (2).

Automation is really “heteromation.”  This part seems absolutely correct; it is also the subject of Matteo Pasquinelli’s genealogy of AI.  Pasquinelli shows in detail how the automation of labor – and in particular, labor that can be divided into discrete tasks – has been a key factor in the development of computing and other systems from the start; Babbage’s analytical engine is as much about the division of labor as anything else.  Pasquinelli’s last major chapter is about the development of pattern recognition and the models on which current AI are based.  Here, in the case of AVs (and both I and others have talked about this in the case of language models), the system itself performs as well as it does not only because it scrapes a lot of data from the internet and other sources, but also because humans are intimately involved in training the machines, whether in the case of RLHF, toxicity removal, or the identification of images in vision systems.  Vision systems are key to AVs and Ganesh emphasizes that the distributed labor of mechanical Turkers and other annotators are essential to the operation of the vehicles.  The fragility of these image recognition systems is therefore central to the failure of AVs.

Language Modeling Bias, Iterability, Grammatology and Epistemic Injustice

By Gordon Hull

In a recent paper in Ethics and Information Technology, Paul Helm and Gábor Bella argue that current large language models (LLMs) exhibit what they call language modeling bias, a series of structural and design issues that serve as a significant and underappreciated form of epistemic injustice.  As they explain the concept, “A resource or tool exhibits language modeling bias if, by design, it is not capable of adequately representing or processing certain languages while it is for others” (2)  Their basic argument is that the standard way of proceeding with non-English languages, which is more or less to throw more data at the model, build in structural biases against other languages, especially those that are more morphologically complex than English (=df those with lots of inflections).

The proof of concept is in multi-lingual tools:

“The subject of language modeling bias are not just languages per se but also the design of language technology: corpora, lexical databases, dictionaries, machine translation systems, word vector models, etc. Language modeling bias is present in all of them, but it is easiest to observe with respect to multilingual resources and tools, where the relative correctness and completeness for each language can be observed and compared” (6)

They identify several kinds of such structural bias.  The first is that prominent current architectures “tend to train slower on morphologically complex (synthetic, agglutinate) languages, meaning that more training data are required for these languages to achieve the same performance on downstream language understanding tasks” (7). Given the percentage of the available training data that’s in English, this magnifies what’s already a problem.  Second, the models perform poorly on untranslatable words.  Third, they cite a study showing “that both lexicon and morphology tend to become poorer in machine-translated text with respect to the original (untranslated) corpora: for example, features of number or gender for nouns tend to decrease. This is a form of language modeling bias against morphologically rich languages” (7).

LLM, Inc.

By Gordon Hull

In previous posts (one, two, three), I’ve been exploring the issue of what I’m calling the implicit normativity in language models, especially those that have been trained with RLHF (reinforcement learning with human feedback).  In the most recent one, I argued that LLMs are dependent on what Derrida called iterability in language, which most generally means that any given unit of language, to be language, has to be repeatable and intelligible as language, in indefinitely many other contexts.  Here I want to pursue that thought a little further, in the context of Derrida’s (in)famous exchange with John Searle’s speech act theory.

Searle begins his “What is a Speech Act” essay innocently enough, with “a typical speech situation involving a speaker, a hearer, and an utterance by the speaker.”

That is enough for Derrida!  In Limited, Inc., he responds by accusing Searle over and over of falling for intentionality, on the one hand, and for illicitly assuming that a given speech situation is “typical,” on the other.

Let’s look at intentionality first.  In responding to Searle, Derrida explains that he finds himself “to be in many respects quite close to Austin, both interested in and indebted to his problematic” and that “when I do raise questions or objections, it is always at points where I recognize in Austin’s theory presuppositions which are the most tenacious and the most central presuppositions of the continental metaphysical tradition” (38).  Derrida means by this a reliance on things like subjectivity and representation – the sorts of things that Foucault is getting at when he complains in the 1960s about philosophies of “the subject” (think: Sartre and phenomenology).  Derrida is involved in the same general effort against phenomenology, though he adds a page later that he thinks the archaeological Foucault falls into this tendency to treat speech acts or discursive events in a “fundamentally moralistic” way (39).  No doubt Searle is relieved to know that he belongs in the same camp as Foucault.  In any case, Derrida explicitly says a few pages later that “the entire substratum of Sarl’s discourse, is phenomenological in character” (56) in that it is over-reliant on intentionality.

A Least-Bad Solution for Language Model Defamation?

By Gordon Hull

Large Language Models (LLMs) are well-known to “hallucinate,” which is to say that they generate text that is plausible-sounding but completely made-up.  These difficulties are persistent, well-documented, and well-publicized.  The basic issue is that the model is indifferent to the relation between its output and any sort of referential truth.  In other words, as Carl Bergstrom and C. Brandon Ogbunu point out, the issue isn’t so much hallucination in the drug sense, but “bullshitting” in Harry Frankfurt’s sense. One of the reasons this matters is defamation: saying false and bad things about someone can be grounds to get sued.  Last April, ChatGPT made the news (twice!) for defamatory content.  In one case, it fabricated a sexual harassment story and then accused a law professor.  In another, it accused a local politician in Australia of corruption.

Can LLMs defame?  According to a recent and thorough analysis by Eugene Volokh, the answer is almost certainly yes.  Volokh looks at two kinds of situation.  One is when the LLM defames public figures, which is covered by the “actual malice” standard.  Per NYT v. Sullivan, “The constitutional guarantees require … a federal rule that prohibits a public official from recovering damages for a defamatory falsehood relating to his official conduct unless he proves that the statement was made with ‘actual malice’ – that is, with knowledge that it was false or with reckless disregard of whether it was false or not” (279-80).

Deepfake Research: ChatGPT can produce fake data

By Gordon Hull

There’s been a lot of concern about the role of language models in research.  I had some initial thoughts on some of that based around Foucault and authorial responsibility (part 1, part 2, part 3).  A lot of those concerns have to do with the role of ChatGPT or other LLM-based product and how to process that.  The consensus of the journal editorial policies that are emerging is that AI cannot be an author, and my posts largely agreed with that.

Now there’s news of a whole other angle on these questions: a research letter in JAMA Ophthalmology reports that the authors were able to use ChatGPT-4’s Advance Data Analysis capabilities to produce a fake dataset validating their preferred research results.  Specifically:

“The LLM was asked to fabricate data for 300 eyes belonging to 250 patients with keratoconus who underwent deep anterior lamellar keratoplasty (DALK) or penetrating keratoplasty (PK). For categorical variables, target percentages were predetermined for the distribution of each category. For continuous variables, target mean and range were defined. Additionally, ADA was instructed to fabricate data that would result in a statistically significant difference between preoperative and postoperative values of best spectacle-corrected visual acuity (BSCVA) and topographic cylinder. ADA was programmed to yield significantly better visual and topographic results for DALK compared with PK”

This is a very technical request!  It took a bit of tweaking, but soon “the LLM created a seemingly authentic database, showing better results for DALK than PK,” P < .001.

The authors suggest some possible strategies to manage this but suffice it to say it is terrifying.  There is already a longstanding, huge problem with fabricated, doctored or otherwise bogus scientific research out there.  One report suggests that 70,000 “paper mill” (= almost completely faked) papers were published in the last year alone.  In real papers, references are often inaccurate.  Publishers already are having to grapple with lots of problematic doctored images, and Pharma has long tilted the entire scientific enterprise to produce results favorable to its products.  At the end of last year, Stanford’s president was forced out over research misconduct in his labs.  In an initial report into the Stanford investigation, STAT News reported data from Retraction Watch to the effect that a paper is retracted, on average, every other day for image manipulation.  Retraction Watch had, at that time (Dec. 2022) 37,000 papers in its database.  The top 5 most-retracted authors have at least 100 retracted papers each.

Into that mess, enter the ability to generate bespoke data on demand.

RLHF and Curation Transparency

By Gordon Hull

Last time, I followed a reading of Kathleen Creel’s recent “Transparency in Complex Computational Systems” to think about the ways that RLHF (Reinforcement Learning with Human Feedback) in Large Language Models (LLMs) like ChatGPT necessarily involves an opaque, implicit normativity.  To recap: RLHF improves the models by involving actual humans (usually gig workers) in their training: the model presents two possible answers to a prompt, and the human tells it which one is better.  As I suggested, and will pursue in a later post, this introduces all sorts of weird and difficult-to-measure normative aspects into the model performance, above and beyond those that are lurking in the training data.  Here I want to pause to consider this as a question of opacity and transparency. I’m going to end up by proposing that there’s a fourth kind of transparency that we should care about, for both epistemic and moral reasons, which I’ll call “curation transparency.”

Implicit Normativity in Reinforcement Learning with Human Feedback in Large Language Models

By Gordon Hull

This is somewhat circuitous – but I want to approach the question of Reinforcement Learning with Human Feedback (RLHF) by way of recent work on algorithmic transparency.  So bear with me… RLHF is currently all the rage in improving large language models (LLMs).  Basically, it’s a way to try to deal with the problem that LLMs aren’t referentially grounded, which means that their output is not in any direct way connected to the world outside the model.

LLMs train on large corpora of internet text – typically sources like Wikipedia, Reddit, patent applications and so forth.  They learn to predict what kinds of text are likely to come next, given a specific input text.  The results, as anybody who has sat down with ChatGPT for long knows, can be spectacular.  Those results also evidence that the models function, in one paper’s memorable phrasing, as “stochastic parrots.”  What they say is about what their training data says is most likely, not about what’s, say, contextually appropriate.  But appropriate human speech is context-dependent, and answers that sound right (in the statistical sense: these words, in general, are likely to come after those words) in one context may be wrong in another (because language does not get used “in general”).  RLHF is designed to get at that problem, as a blogpost at HuggingFace explains:

More on AI and Copyright

Another case percolating through the system, this one about Westlaw headnotes.  The judge basically ruled against a series of motions for summary judgment, which means that the case is going to a jury.  Discussion here (link via Copyhype)

Cars are Privacy Disasters

This article from Gizmodo reports on research done over at Mozilla.  Newer cars – the ones that connect to the internet and have lots of cameras – are privacy disasters.  Here’s a paragraph to give you a sense of the epic scope of the disaster:

“The worst offender was Nissan, Mozilla said. The carmaker’s privacy policy suggests the manufacturer collects information including sexual activity, health diagnosis data, and genetic data, though there’s no details about how exactly that data is gathered. Nissan reserves the right to share and sell “preferences, characteristics, psychological trends, predispositions, behavior, attitudes, intelligence, abilities, and aptitudes” to data brokers, law enforcement, and other third parties.”

Nissan’s response tells you everything that’s wrong with current privacy legislation:

““When we do collect or share personal data, we comply with all applicable laws and provide the utmost transparency,” said Lloryn Love-Carter, a Nissan spokesperson. “Nissan’s Privacy Policy incorporates a broad definition of Personal Information and Sensitive Personal Information, as expressly listed in the growing patchwork of evolving state privacy laws, and is inclusive of types of data it may receive through incidental means.””

Let’s translate.  Nissan is probably compliant.  Also, privacy compliance is a joke.  Also, compliance apparently only requires that you receive NOTICE that they take your data AND CONSENT to that policy, probably merely by driving the vehicle.  Also, they probably reserve the right to change their privacy policies unilaterally, at will.  Also, they almost certainly do not let you opt-out of any of it while CONSENTING by driving the car.  It’s a very special kind of “contract” and “consent.”  Also, how do they know about your sex life?  Also, even if you have sex in the car, there is basically no answer to that question that is not beyond creepy!

As you may have guessed, NOTICE AND CONSENT is an utter sham and has been for a while.  The gizmodo article spells out some of the particular absurdities here – for example, you may not want to ride in one of these cars either, as passengers are “users” deemed to have CONSENTED to the privacy policy.  Your driver should probably provide you NOTICE beforehand!  “A number of car brands say it’s the driver’s responsibility to let passengers know about their car’s privacy policies—as if the privacy policies are comprehensible to drivers in the first place.”  No wonder folks are cynical and resigned about corporate privacy – they’re manipulated into it by corporations.  Also, they’re confused, frustrated and angry about the fact they don’t actually get to consent.

This is the best example I’ve seen of all that in a while, and a crystal-clear indicator of why we need not just new privacy legislation (we do!) but a new direction (more real regulation, less soft compliance and "notice and consent" fig-leaves).

PS – not picking only on Nissan:

“Other brands didn’t fare much better. Volkswagen, for example, collects your driving behaviors such as your seatbelt and braking habits and pairs that with details such as age and gender for targeted advertising. Kia’s privacy policy reserves the right to monitor your “sex life,” and Mercedes-Benz ships cars with TikTok pre-installed on the infotainment system, an app that has its own thicket of privacy problems.”

On Foucault on “Chardino-Marxism,” part 5: Lunacharsky

By Gordon Hull

I’ve been developing (first, second, third, fourth) some reflections on what Foucault means by a reference to “Chardino-Marxism,” a disturbing trend that he credits Althusser with “courageously fighting.”  The real opposition point seems to be Roger Garaudy, a PCF intellectual who is a leader in the effort to establish a post-Stalinist humanist Marxism, and who had a real sympathy for religion.  Last time, I traced some of Garaudy’s sources on religion to Engels.  Some of what Garaudy says also sounds like it’s coming straight from the Russian Marxist Anatoly Lunacharsky’s Religion and Socialism (1908).  The claim here is categorically not that Garaudy read Lunacharsky – as will become evident in a minute, I think that’s highly unlikely.  What I do want to underscore is that there is a coherent line of thought behind Garaudy’s religious impulse.  As I’ll note when I get back to Garaudy and Althusser, there is a very specific political context to Althusser’s attacks on Garaudy having to do with the latter’s role in the PCF and his effort to use a humanist Marxism as a (from Althusser’s point of view, failed) alternative to Stalinism. 

I know very little about Lunacharsky (Wikipedia here), but apparently he was tolerated by Lenin (despite being criticised heavily), and fell out of favor under Stalin.  He died before he could be repressed, but in 1936-8, his memoirs were banned and he was erased from the official histories of communism.  He enjoyed somewhat of a revival after Stalin’s death.  Religion and Socialism is very obscure now: Google books reports a Yiddish translation (!) as well as a Spanish one from the 1970s.  It’s not been translated into English or French.  Marxists.org refers only to his later works in English and in French, and he doesn’t even show up on the German part. 

Most of the work available on Lunacharsky now seems to be attributable to patient work by Roland Boer (upon whom I am completely dependent here).  Religion and Socialism fell out of favor due to Lenin’s denunciation after it was published, was left out of Lunacharsky’s collected works, and was reduced to a few copies.  Here is Boer in his paper on the text: