New APPS: Art, Politics, Philosophy, Science

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.”

Privacy and “Injury in Fact,” Part 4: Administrative Procedure Act Cases

Here I want to complete my review of federal legal precedents for the Supreme Court’s sudden invocation of “injury in fact” language to understand judicial standing in its 1970 Data Processing decision (recall the earlier installments: first, second, third. The first one explains the issue; if you want to escape my rummaging through the archive, you can skip to this one).  Congress passed the Administrative Procedure Act in 1946 to, well, regulate administrative procedures and provide checks against their being arbitrary (this is one of the Acts that virtually of Trumps recent executive orders violates).  Litigation about agency actions after the APA thus had to route through the APA, which imposed its own standards for judicial review of agency actions.

Here, the language of the lower courts gets very close to the issues in Data Processing.  Consider first Curran v. Laird, in which a maritime union sought enforcement of the Cargo Preference Act, which required that American ships be used for military cargo.  After awarding standing, the DC Cicruit concluded that the decision was a matter for agency discretion under the APA, ruling in favor of the government on the merits.  The Court opens its standing discussion by noting that “plainly [plaintiffs are] aggrieved in fact by the allegedly unlawful action of the Secretary of Defense.” The Court then writes that:

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.

Explainable AI as Disciplinary Power

I want to take a break from judicial standing doctrine to note a recent and helpful paper by Emily Sullivan and Atoosa Kasirzadeh about explainable AI.  Explainable AI is a research agenda – there’s a lot of papers and techniques (for a current lit review, see here) – that is designed to get at a central problem in using AI: we often have no idea why machine learning systems produce the outputs they do.  In a variety of contexts, ranging from safety critical systems to democratic governance, being able to understand why the algorithm made the prediction is did is important. Hence the research agenda. 

First, a little detour. Algorithmic governance can be disciplinary in that it can nudge people inexorably toward conforming with norms, whether social or statistical.  Insurance has been well-studied for its normalizing techniques.  In an early paper on privacy unraveling, Scott Peppett showed how the addition of smart-driving surveillance (where insurers give a discount to people who install these devices that record their speed, when they drive, etc.) generate a downward ratcheting on privacy: users who are good drivers have the incentive to adopt the devices, since they get lower insurance rates.  Those who are in the next tier down (above-average drivers) have an incentive to get the devices because that associates them with the good drivers.  And so it goes, until only the worst drivers are declining the surveillance.  And at some point, not having the surveillance device becomes a stigma that raises your rates.  So pretty soon, surveillance devices can become normal.  In the meantime, once drivers have the surveillance installed, surveillance-enabled insurance can nudge them to drive less at night and to otherwise comply with whatever the insurance company says makes you a good risk.  All of that can be automated - the insurance app can tell you, real time, how your driving is impacting your premium.

Privacy and “Injury in Fact,” part 2: ye olde “injury in fact”

Last time, I set out the question of judicial standing and the abrupt switch by the Supreme Court in 1970 to the requirement that plaintiffs show an “injury in fact” to obtain standing.  Here I want to look at the historical development of that term.

The earliest use of the phrase “injury in fact” in US courts is in Hepburn and Dundas v. Auld, an 1809 Supreme Court case about contracts.  “Specific performance” of a contract means that you can compel the other party to do exactly what the contract says (as opposed to, for example, doing something else of equivalent value).  Hepburn and Dundas (H&D) contracted with Graham for the sale of some 6,000 acres of land for $18,000, payable in installments with interest.  H&D also owed Dunlop & Co. a large sum of money, and agreed with Auld (Dunlop’s agent) to transfer the contract to Auld, including the rights to enforce it, “towards the discharge” of this debt (271).

Casemine.com offers the following helpful summary of what happened:

“Hepburn and Dundas entered into an agreement with Colin Auld, wherein they bound themselves to assign a land contract to Auld in lieu of paying a debt. When they failed to execute this assignment on the stipulated date, they sought specific performance from the court to compel Auld to accept the assignment and release them from all claims by Dunlop Co. The Supreme Court, through the majority opinion delivered by Chief Justice Marshall, ruled against Hepburn and Dundas. The Court found that despite subsequent efforts to cure title defects, the initial omissions rendered the contract unsuitable for specific performance. The bill for specific performance was dismissed, upholding the decision that fundamental contractual and title requirements must be met unequivocally”

In other words, Hepburn and Dundas tried to force Auld to accept assignment of a land contract, per an agreement they had made with him.  The Court ruled that, because the title had problems, it was inappropriate to enforce the contract exactly as written and force Auld to accept the (defective) title. 

How does this have anything to do with injury in fact?  As part of the proceedings - basically a subplot that doesn't bear on the final resolution - questions had arisen as to whether H&D actually held title to the land, Auld having averred that “it was apparent that their title was bad, or at all events doubtful” (264).  In response:

“Hepburn and Dundas filed a supplemental bill which states their title. It avers possession ever since 1773, and refers to certain title papers; they say that they verily believe their title to be good, and never heard a doubt till long after the tender of the assignment; that as soon as the objections were made known, they took pains to remove them, and have lately obtained deeds of confirmation from the surviving patentees. That the title of Sarah, one of the co-devisees of John West, after her death in 1795, descended upon her brothers Thomas, John, and Hugh and her sister Catharine, and that John, Hugh, and Catharine have lately confirmed their title, and refer to the deeds, and they suppose that Thomas had passed all his title to Sarah's part by a deed executed before her death.” (264).

The case then recites the details of these efforts and their documentation. The question, then arose as to whether H&D’s efforts to clarify and secure the title to the land constituted interference with the contract. The Court concludes that it did not because it did not cause Auld an “injury in fact:”

“The interference of Hepburn and Dundas, in accommodating the suit with Graham, is also urged as an objection to their conduct. They had certainly no right to interfere without the consent of Colin Auld. But when the correspondence is inspected, and it is perceived that they interfered only to effect the object he had himself desired, and which he had avowed his own inability to effect without their consent, the interference must be considered as innocent in point of intention, and unproductive of injury in fact.  The court, then, perceive[s] nothing in the conduct of the plaintiffs, up to the decision of the suit with Graham, which ought to defeat their right to demand a specific performance of this contract. Could they at that time have conveyed a good title, Colin Auld ought to have accepted it.” (Hepburn & Dundas v. Auld, 9 U.S. 262, 275 (1809)

That is, H&D were trying to do what Auld wanted them to do by honoring the contract, and that effort didn’t hurt Auld.

The Court then invokes a counterfactual.  Would things have been different had H&D not tried to remedy the defects in the title?

“These omissions, then, to record the deeds of Thomas and Hugh West, and the total want of title as to Mrs. Bronaugh's part, have produced no real inconvenience to Colin Auld. Had the title been unexceptionable, it would still have been refused, and this contest would still have been carried on with the same determined perseverance which marks the conduct of the parties. Under these circumstances, it is the opinion of the majority of the Court that this case ought to be governed by those general principles which regulate the conduct of a court of chancery in decreeing a specific performance, if the defect of title, which existed at the time of contract, be cured before the decree” (275).

In other words, equitable principles dictate that if the defects in the title can be remedied, that’s ok.  H&D shouldn’t be penalized for interfering with the contract by trying to do so.  Unfortunately for H&D, they didn’t fix the problems with the title: “The omission to record the deed from Thomas West is not cured, and this Court is now to decide whether, under these circumstances, Hepburn and Dundas are entitled to claim a specific performance.”

Should H&D be entitled to force Auld to accept the land transfer, given the problems in recordkeeping?  Equity says no, because Auld would likely incur expenses and take on risk in dealing with the missing deed:

“Had there been simply a deficiency of 208 acres, the majority of the Court would have considered it as a case for compensation; or had the parties entitled to this land been before the court, a division might possibly have been directed, and compensation for that quantity ordered; but however this might be, as persons not before the Court hold this interest, no order can be made respecting it, and it may very much embarrass those acts for asserting the title which may possibly be necessary. The part actually conveyed by Thomas West, too, never having been confirmed by a deed from himself or his heirs, properly recorded, might impose on Colin Auld the necessity of bringing a suit in chancery to perfect his title, or of being subjected to the inconveniences constantly attending the establishment of a deed not recorded, and the risks inseparable from such a deed.” (278).

The Court concludes that “this, therefore, is thought by a majority of the Court to be a case not proper for a specific performance, and the bill is to be dismissed.” (278).

The case is byzantine, and noted for its establishment of principles of contract and equity in real estate (tl;dr: make sure you have clean title to land you want to transfer the rights to!).  But the “injury in fact” language does read like a proto-version of what the Court came up with in 1970: did anything materially bad happen to Auld?  Significantly, the question is whether behavior by H&D which was technically not proper actually caused any such “inconvenience” (“inconvenience” can mean serious harm: Locke uses it, for example, to refer to life in the state of nature, e.g., at Second Treatise secs. 13 and 90).  Concluding it did not, the Court decided that it did not warrant legal intervention.  That said, it’s worth underlining that all of this is framed in traditional juridical terms: the question is one of fulfillment of contractual obligations, and the parties are understood to be bound by the contractual terms and the questions are about whether their behavior hinders the ordinary performance of those terms.

Next time I’ll continue with the historical rabbit hole.

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 3)

By Gordon Hull

I’ve been using (part 1, part 2) a new paper by Fabian Offert, Paul Kim, and Qiaoyu Cai to think more about Derrida’s use of iterability as a way in to thinking about transformer-based large language models (LLMs) like ChatGPT.  Here I want to wind that up with some thoughts on Derrida and Searle.

Near the end of the paper, Offert, Kim and Cai summarize that:

“The transformer does exactly more and less than language. It removes almost all non-operationalizable sense-making dimensions (think, for instance of interior paratextuality, of performativity and contingency, or of anything metaphorical that is more complex than a simple analogy) – but it also adds new sense-making dimensions through subword tokenization, word embedding, and positional encoding. Importantly, these new sense-making dimensions are exactly not replacing missing information, but they are adding new, continuous information” (15).

This returns us to the Derridean concerns I’ve been articulating.  Recall that in his polemic against Searle, Derrida accuses Searle’s version of speech act theory of too closely modeling phenomenology, both in assuming an intentional agent behind speech acts and in taking “typical” speech situations as central, as opposed to “parasitic” ones like humor.  I suggested that there is good evidence that various efforts to impose normative structures on language models – RLHF, detoxification, etc. – push them to perform in ways that call to mind Derrida’s critique of Searle.  By taking certain language situations as normal, Searle is making the account of speech acts normative before it even gets started.  In his own defense, Searle argues that the reduction to typical speech acts is for convenience only, and for keeping the model tractable.  Techniques like detoxification and RLHF similarly reduce the range of the models’ output.  Evidence of this is that LLMs lack the contextual richness to have a sense of humor, no matter how otherwise sophisticated their output.

Offert, Kim and Cai’s paper lets one add that this reduction runs much deeper.  The very processes of tokenization, for example, are designed to reduce the number of possible tokens to a tractable number.  In this respect, the move is analogous to Searle’s.  It is defensible for the same reason: it lets you get to a generalizable model.  But it’s vulnerable to the Derridean critique, also for the same reason.  The model makes a number of assumptions that aren’t the same as what language does.  So there is a certain sloppiness in talking as though it’s an accurate representation of language.  All models abstract; that’s not the point.  For subword encoding, the point is that the abstraction isn’t choosing to ignore certain aspects of reality in order to produce a model, it’s that the abstraction changes the nature of what it is modeling.  That’s fine – but that also means that the although the transformer model is producing something that looks like language, the process by which it gets there is definitively not linguistic.

Transformer models, Iterability, and language (part 2)

By Gordon Hull

Last time, I looked a new paper by Fabian Offert, Paul Kim, and Qiaoyu Cai and applied it to a reworking of some of my earlier remarks on Derrida’s use of iterability in transformer-based large language models (LLMs) like ChatGPT.  In particular, I tried to draw out some of the implications of subword tokenization for iterability.  Here I want to continue that process with other aspects of the transformer model.

If subword tokenization deals with the complexity of semantics by limiting the number of tokens, another attempt to preserve semantics while reducing complexity is through word embeddings.  Here, the model constructs a vector mapping that puts words that tend to occur near to one another in text in locations proximate to one another.  We might imagine a two-dimensional space, which locates words by age and gender.  In such a space, “man” and “grandfather” would likely be closer to one another than “grandfather” and “girl,” since man and grandfather are closer to one another in terms of both age and gender than grandfather is to girl. As Offert, Kim and Cai explain, “Either learned during the training process or sourced from another model, embedding matrices position each word in relation to other words in the input sequence” (11).  The sentence comes with a footnote about efforts to draw Derridean implications:

“This relational aspect of word embedding has often been compared to poststructuralist notions of meaning, particularly Derrida’s notion of différance. It should be noted, however, that the relational saliency of embedded words is a product only of their operationalization: only words that are, in fact, numbers, gain relational saliency”

In other words, if it is true that the meaning of words is only in relation to other words, it is also true that a language model’s available catalog of words is arbitrarily limited by the training data it ingests and any other limitations on which words make it onto vectors. 

These mappings can be incredibly complex, but they are finite – which generates another limit on iterability.  As Offert, Kim and Cai put it:

“Though this proximity is not arbitrary, it is still constrained by the continuous but finite dimensions of the embedding vector space (typically in the hundreds or even thousands) that may not capture all possible nuances in the relation between words. As such, relationships between words are to be understood as contextually determined and constrained by the fundamental limitations inherent in translating words to numbers. Nevertheless, as the famous analogy test shows, some semantic aspects are indeed preserved” (11).

There at least two aspects of this to look at.  First, the caution about différance.  This seems correct at least in part because of the way Derrida, well, embeds his usage of différance.  In Limited, Inc, for example, he writes that “the parasitic structure is what I have tried to analyze everywhere, under the names of writing, mark, step, margin, différance, graft, undecidable, supplement, pharmakon, hymen, parergon, etc.” (103).  The final indeterminacy of is of course marked by the indefinite “etc.” with which the list closes, but that Derrida associates it with the parasitic structure suggests that the target here is the sort of intentional phenomenology that Derrida says underlies speech act theory.  Early in the text, for example, he proposes that différance is “the irreducible absence of intention or attendance to the performative utterance” (18-19).  The meaning of the words depends on their proximity to other words because it does not depend on authorial intent in the sense that iterability means that authorial intent can never be dispositive of the meaning of an utterance, and that the meaning of the utterance can never be made fully present.  It is not clear what it means to take that argument and apply it to a vector representation.

Second, the analogy test. The analogy test refers to the model’s ability to fill in analogy statements: “x is to y as z is to…”  The brilliance of this capacity and its limitations are best shown through an example.  Here is ChatGPT:

Both cases show that the model has preserved a lot of semantics, and sees the logical structure of the analogy.  But it’s also limited: when you read “computer programming is to employment as philosophizing is to what,” I’m guessing your mind immediately jumped to “unemployment.”  That’s both a longstanding joke, and a real fear.  But both of those reasons why you might have said “unemployment” require broadening the context quite a bit or seeing contexts other than the typical ones.  ChatGPT was capable of the joke – but it had to be explicitly told to make it:

It seems to me, at least, that this shows the connection between word embeddings and the sort of complaint Derrida makes against Searle’s demotion of parasitic forms of language: both of them rely on an implicit normativity to make their project tractable in the first place.  But there are contexts where the parasitic meaning is in fact the first meaning, and precisely that sense of unusual context is what disappears in the language model.

More next time…

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 2)

By Gordon Hull

Last time, I followed a paper by Maya Indira Ganesh that thinks about the construction of autonomy in autonomous vehicles. Ganesh points out that what happens isn’t so much the arrival of an autonomous vehicle but the displacement of human labor into a sociotechnical system and the production of the driver as a supervisor of this system. The result is a system of (in Karen Levy’s terms) “baton passing” where the machine gets the human to help it navigate tricky situations.

So why expect it to or act like it will?  It lets people get blamed when they fail to do the impossible; as Ganesh puts it, “no doubt this surveillance data will protect car and ride-sharing companies against future liability if drivers are found to be distracted. This monitoring is literal bodily control because it is used to make determinations about people” (7).  In other words, the imaginary of an autonomous vehicle that passes control to a driver when needed allows everyone to forget that we know that the visual recognition systems in AVs aren’t reliable.  Instead of a failure on the part of whoever made the AV, the failure is on the driver who should have known that the vehicle would need assistance, and who failed to promptly act when it did.  As I indicated last time, Levy cites quite a bit of research to the effect that humans cannot plausibly perform this function.  But that doesn’t mean that there won’t be efforts to stick it on us; Levy outlines how various forms of biosurveillance are coming for truckers, and we can imagine similar efforts coming for drivers of AVs.

Levy argues that one effect of all of this is to deflect our attention away from actual social problems.  Her book is structured around the adoption of electronic logging devices (ELDs) to monitor truck drivers.  The core social problem is that truckers are only compensated for the time they are actually moving their vehicle – not the time they spend waiting to unload it, finding a place to sleep, eating meals on the road, etc.  Here is Levy:

“By using digital surveillance to enforce rules, we focus our attention on an apparent order that allows us to ignore the real problems in the industry, as well as the deeper economic, social, and political causes. Under the apparent order envisioned by the ELD, the fundamental problem in trucking is that truckers cannot be trusted to reliably report how much they work, and the solution to that problem is to make it more difficult for them to fudge the numbers. But under the actual order, the problem in trucking is that drivers are incentivized to work themselves well beyond healthy limits – sometimes to death. The ELD doesn’t solve this problem or even attempt to do so. It doesn’t change the fundamentals of the industry – its pay structure, its uncompensated time, its danger, its lack of worker protections …. More meaningful reform in trucking would require a ground-up rethinking of how the industry is structured economically, in order to make trucking a decent job once more. So long as trucking is treated as a job that churns through workers, these problems won’t be solved …. If we paid truckers for their work, rather than only for the miles they drive – for the actual number of hours they work, including time they are waiting to be loaded and unloaded at terminals, time they are inspecting their trucks and freight, time they are taking the rest breaks that their bodies need to drive safely – drivers would be far less incentivized to drive unsafely or when fatigued” (153-4, emphases original).

But of course paying truckers for all that work would be bad for the profit margins of trucking companies and might well require slightly more expensive consumer goods.  Also, most of the unpaid work is also work that cannot be automated or quantified in the same way that miles can.  Paying truckers only for the miles they drive is convenient both for capitalism and quantification.

Ganesh mentions the story of Rafaela Vasquez, an Uber test driver who failed to stop her AV from striking and killing a cyclist crossing the road.  Ganesh’s article is from 2020; in July 2023, Vasquez pleaded guilty to a reduced charge of endangerment and will serve four years of supervised probation.  Vasquez had apparently been looking at her phone up until the last second before the crash.  But recall Levy: Vasquez might have been unable to stop the vehicle even if she’d been staring at the road with both hands on the wheel, because the failure of the image recognition system to recognize a cyclist surely counts as unpredictable and because she’d likely be suffering from passive fatigue and vigilance decrement.  Worse, lots of research tells us an ugly truth about the criminal justice system: most cases don’t go to trial and accused folks take plea bargains because of the risk of going to trial on a charge like negligent homicide, which is what Vasquez was originally charged with, is too great.  This is especially true for defendants who don’t have lots of money and therefore can’t afford really good representation.  So the criminal record and probation and guilty plea may very well be the result of a maximin calculation on Vasquez’s part, not any sense of guilt in the way most of us use the term.

In any case, Vasquez just took all the blame for a cascade of failures:

“The investigation found that the probable cause of the crash was Vasquez’s failure to monitor the self-driving car’s environment while she was distracted by her phone. But it also accused Uber of contributing to the crash by operating with an “inadequate safety culture” and failing to properly oversee its safety drivers. Uber’s automated driving system failed to classify Herzberg as a pedestrian because she was crossing in an area without a crosswalk, according to the NTSB. Uber’s modifications to the Volvo also gutted some of the vehicle’s safety features, including an automatic emergency braking feature that might have been able to save Herzberg’s life, investigators wrote.”

That’s right: Uber somehow failed to train its system on the idea that a pedestrian might cross the street somewhere other than the sidewalk, and they also disabled safety features on the car!  And the NTSB cited this contribution to the crash!  And yet only Vasquez faced criminal liability; “prosecutors declined to criminally charge Uber in the crash in 2019 but recommended that investigators further examine Vasquez’s actions as the safety driver.”  Uber got away with paying a settlement to the victim’s family.  

The decision to prosecute Vasquez and not Uber resulted from a one-sided application of a “but-for” causality standard:

“Vasquez's eyes were focused on the phone screen instead of the road for approximately 32 percent of the 22-minute period, the report said. Tempe investigators later determined the crash would not have occurred if Vasquez had been ‘monitoring the vehicle and roadway conditions and was not distracted.’ But [County Attorney Sheila Sullivan] Polk said that wasn't enough to prosecute Uber. ‘After a very thorough review of all the evidence presented, this Office has determined that there is no basis for criminal liability for the Uber corporation arising from this matter,’ Polk wrote.”

We need to separate the question of how much Vasquez could/should have done from the disappearance of liability for Uber.  Their conflation, and the urge to focus only on Vasquez is is the work done by the baton-passing view.

AV advocates will invariably lead by telling you that most vehicle accidents are caused by people.  This somehow then legitimates the idea that we should adopt their product because it isn't a people.  That move ignores the extent to which AVs are systems that displace human labor and reconfigure the biological agent into a distributed, sociomorphic system.  Because this system isn't very reliable, they call up on a human to supervise it.  But that then obscures that the “baton passing” model they are promoting is provably very dangerous, and it makes humans even less reliable.  It also hides that the image recognition systems can and will fail in predictable and catastrophic ways.  Beyond that, it hides all of the massive cost to a country where cars are the only viable mode of transportation for the vast majority of people, and perpetuates a system where other kinds of transit are unimaginable.  And it buries all of that with rhetoric of responsibility and autonomy that underscores that hapless humans – see, we told you! – are bad drivers (and bad pedestrians).  In this regard, the moral language around humans supervising (and failing to adequately supervise) AVs is analogous to the use of moral language to blame poor mothers for their condition in order to justify intrusive surveillance of them, rather than thinking about the structural features that make so many mothers poor. If only they would work harder!  If only drivers would work harder! As Ganesh says, “’Autonomous’ driving is less about the promise of freedom from the car, and is perhaps for the car” (5).  And the people that get rich selling you on the imaginary of the car.

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).