Matthew Kirschenbaum

Widely reported, just the first hit that comes up for me: www.cnn.com/2026/09/18/p...

cnn.com

Example: transformer-circuits.pub/2026/workspa...

Verbalizable Representations Form a Global Workspace in Language Models transformer-circuits.pub

Nor are we lacking in actual ways to learn about what “some guy” made, or why people reacted to it the way they did. mitpress.mit.edu/978026205248...

mitpress.mit.edu

🌲 arxiv.org/abs/2502.19190

Provocations from the Humanities for Generative AI Research arxiv.org

There’s a literary scandal involving AI on a weekly basis now. Prizes, contracts, allegations, retractions. Individual authors will make their own choices as they always have. But collectively and industry-wide, there is no putting the stopper back in this bottle. www.theguard…

$2m crime novel deal collapses amid questions over AI use theguardian.com

A point I will make next week is that I believe the salient fact of the Hugging Face incident is not any of the discourse about consciousness or agency, but that the internet and all adjacent infrastructure offer a common manipulable/executable substrate. visit.bodleian.ox.ac.…

Textpocalypse Now: The Political Economy of Writing in the Age of AI visit.bodleian.ox.ac.uk

The University of Maryland, where four out of five of the researchers co-authoring this paper are based, has a very fine English Department. Just saying. arxiv.org/pdf/2604.03136

arxiv.org

At some point we’re going to need to square the circle between puffing “Why are you asking AI, go ask a real person instead” and all of us who have told someone “If you are soliciting my expertise then please pay me.” +

I’m not sure where the line is beteeen “of course AI agents can’t end the world, they’re just autocomplete bots” and “look, we almost started WWIII by taking our marching orders from bots.” If the point is that it’s people not machines to blame that’s uncomfortably close to an NRA bumper sticker.

Deep Thoughts on the Current AI Discourse. 🧵: 1. I don’t think robots are on the verge of going rogue and murdering us all in our digital sleep. My p-doom is low. Note that this is separate from the question of ongoing enshittification, where my e-shit number is pretty well maxed out. +

What is your favorite, clever, and non-obvious shorthand descriptor for LLMs? Not looking for stochastic parrots or autocomplete on steroids, also not looking for throwaways like bullshit plagiarism machines. Also not looking for explainers: I’m looking for a smart and memorable phrase or sentence.

I don’t think LLMs “think.” But this is a wildly inaccurate description of what they are doing (to start, there is no “database”). The phenomenon, endemic to all social media, is that you have people adopting the very same veneer of pseudo-authority that the models do when at their most unctuous.

At an online event this morning I defended a claim which others found troublesome: That there are legitimately experienced aspects of computational phenomena, such as large language models, which—for all practical ways of speaking—are not always resolvable by or restorable to material conditions.

I spent the first half of the day at a public facing AI convening in downtown DC. Eclectic speakers, collected audience. The biggest blindspot I saw was this: Culture was repeatedly positioned as something *outside* of tech. We can make better tech if we bring more culture into the process. +

I understand why we’re still explaining to people that LLMs are not conscious but I also don’t understand why we’re still explaining to people that LLMs are not conscious. Like imagine thinking this is the conversation to have.

LLMs (next token predictors) are formal operators that operate within formal, or at least computational, systems. Computational systems are fully capable of supporting touch points to the external, or material, world. This is such an obvious point as to be banal. Example: a digital thermostat.

I understand the impulse to make this point, but I think it also misdirects the agency. The language of “AI” was already all over the social milieu of the research behind large language models, including a company named OpenAI which was working on lots of other things besides. +