New APPS

I want to take a break from Derrida and language models this week to explore an emerging policy issue.  As is impossible to miss, “AI” is everywhere.  Not everything that claims to be “AI” really is, but it’s getting hard to avoid things that call themselves “AI” as the AI companies look to make the technology profitable.  This is happening despite the decidedly lukewarm public attitude toward AI.  Current Pew research, for example, shows that AI experts are very enthusiastic about it, while the public isn’t: only 17% of all the adults surveyed thought AI was going to have a positive effect on the US over the next 20 years.  Concern is growing.

This has generated at least three industry responses.  One is to push for deregulation of AI at the federal level.  Industry advocates nearly snuck in a total ban on state regulation of AI into Trump’s spending bill; it was excised at the last minute by the Senate on a 99-1 vote.  Industry has simultaneously tried to get the executive branch to push (mostly unregulated) AI as vital to national economic competitiveness and security.  Trump has obliged repeatedly, starting with an executive order all the way back in January.  Trump is all about this AI narrative, but it has been a consistent U.S. approach to and story about AI for quite a while.

The second and third approaches are to try to (for lack of a better term) engineer stronger public support.  This second is in the form of PR campaigns about the inevitability and magnificence of AI and the need of it to be shepherded by the incumbent AI companies.  Those who aren’t as fully on board the train – women, for example -  are chastised and presented as doing damage to their careers; their concerns are frequently ignored.  The third is related, and that is the all-out push to get AI into education at every level.  Ohio State and Florida have mandated that AI be in all the curriculum (what does this mean, other than as a branding exercise? Nobody knows).  OpenAI is doing everything it can to make itself ubiquitous on college campuses.  Microsoft is dropping a cool $4 billion on AI education in K12, and OpenAI and Microsoft are sponsoring teacher training.  A couple of weeks ago, Trump dropped an executive order promoting AI in education.

(more…)

By Gordon Hull

As part of thinking through the implications of Lydia Liu’s papers (here and here) demonstrating a Wittgensteinian influence on the development of large language models, I’ve made a detour into Derrida’s critique of writing (my earlier parts: one, two, three).  My initial suggestion last time was that Derrida’s discussion is designed to show that “Platonism” is a political move (not a metaphysical one).  For Derrida the Platonic priority of voice over writing disguises the fact that both are (in his own terms) repetitions of the eidos, and so the claim that writing is bad is the claim that it’s the wrong kind of repetition.  I suggested that for Platonism as read by Derrida, one could easily imagine a hierarchy of writing systems, based on their proximity to voice/speech.  Chinese ideography – which Liu argues is central to Masterman’s breakthroughs in computer language modeling – would be at the very bottom of a Platonic hierarchy.  But because this is Derrida, we can’t either proceed quickly or proceed without talking about Hegel.  So I closed last time with a long passage from Hegel in which he denigrates Chinese for being insufficiently spiritual and too hard to learn.  Today I’ll start with why it’s relevant.

(1) First, Derrida takes up Hegel’s understanding of language in “The Pit and the Pyramid,” first delivered in 1968 and thus almost exactly contemporaneous with “Plato’s Pharmacy.”  There, working from the other end of metaphysics (Hegel, not Plato), Derrida describes such a hierarchy:

(more…)

By Gordon Hull

I’ve been looking (part 1, part 2) at a couple of articles by Lydia Liu (here and here) demonstrating a Wittgensteinian influence on the development of large language models.  Specifically, Wittgenstein’s emphasis on the meaning of words as determined by their contexts and placement relative to other rules gets picked up by Margaret Masterman’s lab at Cambridge and then becomes integrated into the vector semantics models that underlie current LLMs.  Along the way, Liu argues that the Masterman approach to language, which learns a lot from Chinese ideographs, in this sense goes farther than the Derridean critique of logocentrism.  Here I want to transition to Derrida’s critique, not to criticize Liu’s account, but to see an additional point in Wittgenstein, one that Derrida takes further.

To recall, Liu notes that one effect of the Wittgenstein-Masterman approach is to overturn the logocentrism in Western writing, but that Masterman is doing something different from Derrida, who remains in the space of alphabetic writing:

“For Masterman, to overcome Western logocentrism means opening up the ideographic imagination beyond what is possible by the measure of alphabetical writing. This is important, as it follows that the scientist’s and philosopher’s reliance on conceptual categories derived from alphabetical writing in their commitment to logical precision and systematization as well as their deconstruction must likewise be subjected to post-Wittgensteinian critique” (Witt., 437)

As she adds a few pages later, “I am fully convinced that Masterman is the first modern philosopher to push the critique of Western metaphysics beyond what is possible by the measure of alphabetical writing, and, unlike deconstruction, her translingual philosophical innovation refuses to stay within the bounds of self-critique” (Witt., 444).

(more…)

I’ve been loosely tracking the AI and copyright cases, most notably the Thaler litigation, where Thaler keeps losing the argument that work solely by an AI should get copyright protection.  To summarize: everybody who has ruled on that said that only work involving humans can get copyright protection.  As I said at the time, I think a good policy reason in support of this argument is that if pure AI could get copyright, it could produce millions of copyrighted images in almost zero time.  That’s got nothing to do with incentivizing human creation.  It was easy to miss given the deluge of atrocious Supreme Court decisions, but last week, a pair of district court judges ruled on a different (but not unrelated, in terms of markets) AI copyright question – whether scraping online text for training data is fair use.  Both cases are in the Northern District of California, so we can expect the 9^th Circuit to have the first appellate decision on this topic.

By way of background: fair use is an affirmative defense against copyright infringement.  That means that if you accuse me of infringement, I can defend myself as having engaged in “fair use,” which basically means “use that the copyright owner doesn’t like, but that we as a society think should be allowed for policy reasons.”  It could also mean “use that everybody thinks is ok, but for which licensing would be so inefficient that a licensing market would never emerge.”  Fair use is supposed to be decided case-by-case.  It depends on four factors: the “nature and purpose” of the (allegedly infringing) use, the nature of the copyrighted work, the amount of it used, and the market effects of the infringing use.  The middle two factors tend not to matter much.  The first factor is usually decided by a deciding whether the use in question is “transformative.”  For example, consider parody; the Supreme Court ruled back in 1994 that a 2 Live Crew parody of Roy Orbison’s “Pretty Woman.” was fair use.  The most closely analogous case I know to the training data was an appellate decision about Google thumbnails.

(more…)

By Gordon Hull

There’s an emerging literature on Large Language Models (LLMs, like ChatGPT) that basically argues that they undermine a bunch of our existing assumptions about how language works.  As I argued in a paper a year and a half ago, there’s an underlying Cartesianism in a lot of our reflections on AI, which relies on a mind/body distinction (people have minds, other things don’t), and then takes language use as sufficient evidence that one’s interlocutor has a mind.  As I argued there, part of what makes LLMs so alarming is that they clearly do not possess a mind, but they do use language.  So they’re the first examples we have of an artifact that can use language; language-use is no longer sufficient to indicate mindedness.  In that paper, drew the implication that we need to abandon our Cartesianism about AI (caring whether it “has a mind”) and become more Hobbesian (thinking about the sociopolitical and regulatory implications of language-producing artifacts).  Treating LLMs as the origin points of speech has real risks, including making the human labor that produces them invisible, and making it harder to impose liability since machines can’t meet a standard scienter requirement for assigning tort liability.

Here I want to take up a somewhat different thread, one that I started exploring a while ago under the general topic of iterability in language models.  This thread takes the literature on language models seriously; where I want to go with it is to talk about an under-discussed latent Platonism in how we tend to approach language(models).   I’ll start with the literature, which divides into a couple of sections, a Wittgensteinian and a Derridean.

1. The Wittgensteinian Rejection of Cartesian AI

Lydia Liu makes the case for a direct Wittgensteinian influence on the development of ML, via the Cambridge Researcher Margaret Masterson.  I only ran into this work recently, so on the somewhat hubristic assumption that other folks in philosophy also don’t know it, I’ll offer a basic summary here (in my defense: Liu says that “the news of AI researchers’ longtime engagement with Wittgenstein has been slow to arrive.”  She then adds that “the truth is that Wittgenstein’s philosophy of language is so closely bound up with the semantic networks of the computer from the mid-1950s down to the present that we can no longer turn a blind eye to its embodiment in the AI machine” (Witt., 427)).

(more…)

The NSF had attempted to reduce indirect costs (F&A) on all future grants to 15%, in a somewhat more coherent version of the NIH's effort to do so for all ongoing and future grants.   A federal court today enjoined the rate cut, vacating the new rule, finding that "National Science Foundation’s 15% Indirect Cost Rate and the Policy Notice: Implementation of Standard 15% Indirect Cost Rate, NSF 25-034 are invalid, arbitrary and capricious, and contrary to law"

I wish I’d come up with that title, but it actually belongs to a new study led by Natalia Kosmyna of the MIT Media Labs.  The study integrates brain imaging with questions and behavioral data to explore what happens when people write essays using large language models (LLMs) like ChatGPT.  I haven’t absorbed it all yet – and some of the parts on brain imaging are well beyond my capacity to assess – but the gist of it is to confirm what one might suspect, that writing essays with ChatGPT isn’t really very good exercise for your brain.  The study assigned participants to one of three groups and had each write essays. One got to use an LLM, one used Google search, and one wasn’t allowed to use either.

The results weren’t a total surprise:

“Taken together, the behavioral data revealed that higher levels of neural connectivity and internal content generation in the Brain-only group correlated with stronger memory, greater semantic accuracy, and firmer ownership of written work. Brain-only group, though under greater cognitive load, demonstrated deeper learning outcomes and stronger identity with their output. The Search Engine group displayed moderate internalization, likely balancing effort with outcome. The LLM group, while benefiting from tool efficiency, showed weaker memory traces, reduced self-monitoring, and fragmented authorship. This trade-off highlights an important educational concern: AI tools, while valuable for supporting performance, may unintentionally hinder deep cognitive processing, retention, and authentic engagement with written material. If users rely heavily on AI tools, they may achieve superficial fluency but fail to internalize the knowledge or feel a sense of ownership over it” (138)

These results were corroborated by the brain imaging, and “brain connectivity systematically scaled down with the amount of external support: the Brain‑only group exhibited the strongest, widest‑ranging networks, Search Engine group showed intermediate engagement, and LLM assistance elicited the weakest overall coupling” (2 (from the abstract)).   That is:

(more…)