It would be pretty surprising if autoregressive models performed otherwise.
That said, I share your concern and I am in the process of re-running the test with the first gen GPT model (the contemporary of BERT).
I will report results…. on a shorter publication timeline than this article!
Yep. That’s right that the experiments are based on BERT, ie an encoder, rather than a decoder unit.
The short answer is that our findings are likely to agree directionally with similar tests on a decoder unit, like GPT. Indeed, they agree with non-LLM computational research on fiction
New paper w/ @teddyroland.bsky.social on "How fiction powers generative AI systems." We designed a computational experiment to test the impact of the vast amount of fiction in LLM training data on how LLMs communicate, w/ implications for both AI design + literary theory. arxiv.org/abs/2603.01220
Yes! Walking to school is one of the great pleasures of our time in Champaign
Starting any new job -- never mind founding a startup -- after being in academia will have a steep learning curve. They will require new skills and compel connections with other people that open new scenarios, many unforeseeable, at lightning speed.
Book revisions -- even if they don't lead to a TT job -- eventually give exposure to the publishing industry (with book pitches, editorial meetings, marketing, etc). It has the potential to branch into different scenarios in cultural production and business, but like years down the line.
Yes, lets!
I think you're exactly right with the "multiple scenarios instead of just one." I'll add a level to it that its a question of *when* different scenarios get activated.
Which is a riskier way to spend my time: revising book chapters in anticipation of a TT job that may never materialize, or writing a business plan for a startup that may never get off the ground?
Extraordinary new public humanities resource that curates in near real-time over 50,0000 items from sources/channels of humanities-related scholars who create pubic-facing essays, podcasts, videos, blogs, etc.: publicscholarship.org
Lots more findings and discussion in the pre-print.
And please, share your feedback with us! The article has been accepted for publication at New Literary History, and we consider the pre-print to be part of our peer review process. Thx ✌️
arxiv.org/abs/2603.01220
In order to make sense of these findings, we interpret them by way of recent literary theory on the nature of fictionality. This helps us (1) answer why characters should feature prominently in LLMs learning from novels and (2) articulate a theory of representation for the AI era.
One major finding is that an early model -- BERT -- learns about gender as a dialogical construction between characters.
More surprising is that the relations are characterized by heightened affect: gender differences are represented through anger, confusion, erotics in high stakes situations.
We know that current AI models train on large swaths of novels, thanks to reporting in The Atlantic.
Richard and I trace fiction's role back the first generation LLMs and we empirically test: What do AI models actually learn from fiction, as opposed to non-fiction sources like Wikipedia?
In the paper, we argue that training data is a new and consequential frontier for cultural representation.
It is a truism that LLMs only know about the world what they learn from their training data, but the insight is rarely tested and, as a result, not well theorized. We aim to rectify that.
Excited to share the pre-print for a forthcoming article in NLH with @richardjeanso.bsky.social 🎉
Generative AI & Fictionality: How Novels Power Large Language Models
arxiv.org/abs/2603.01220
A dh tutorial web app
tl;dr: I build little wonky things to help with my own teaching and research. If all goes well, they turn out to be useful/interesting for other people too. This post is in the spirit of that sharing. A worsening problem I am having is an overall decline in basic digital […]
2016: entertaining myself by watching -LL/token go down in MALLET sessions
2026: entertaining myself by watching Claude Code google keywords from my prompt
Fired up the old blog for this post. Glad to see DH Now is still kicking!
Wowza. Well, I stand by my initial assessment: that post is a litany of broken things in academia (with varying degrees of self-awareness)
A pop discourse that represents the imaginary relationship of college students to their real conditions of trudgery!
As a tea drinker, the energy I use to boil water each month probably matches that of driving a car ten miles. Wild!
I get the sense you are looking for disruptions that are real “creative destruction” — interventions that reorganize the playing field. Any eg’s you think are (potentially) transformative?
Maybe notebook LM has gotten closest so far, as a diff way to engage library holdings/course readings
The flip side is that insights about “the way things work” tells you about what is broken: e.g. Einstein directed attention to the terrible incentives students have in the ed system.
He “digesteth harde yronne” (takes a D3+iron supplement)
Hold the phone! La Sagrada Familia finished being built last week
www.theguardian.com/world/2026/f...
DOGE allegedly used ChatGPT to identify 1,400 NEH grants it said were DEI. Grants were terminated April 2025, according to a court filing. E.g.
Film: 1873 Colfax massacre
Film: first female pilots flying for U.S. military in WWII
Film: “Untold Story of Jewish Women Slave Labor in the Holocaust"
But the elision of code work, when we offload it to AI, puts pressure on the code's ultimate purpose, its end. At some level, that's all that prompting consists of: naming goals for AI agents to pursue.
The goals are weird overdetermined mixtures of ethical, aesthetic, and analytic dispositions
I don't say it quite this way in the post, but my hunch is that we, as users, are only partly conscious of our mixture of goals for any project, and that reading chat transcripts of vibe coding will reveal what it actually consists of.
The transcripts themselves may be AI's real value to humanists!
But the elision of code work, when we offload it to AI, puts pressure on the code's ultimate purpose, its end. At some level, that's all that prompting consists of: naming goals for AI agents to pursue.
The goals are weird overdetermined mixtures of ethical, aesthetic, and analytic dispositions
