Lightnews — Scholar-powered news

LightNews

Meredith Martin

@mmvty.bsky.social

5.3K followers 3.2K following 1.4K posts

poetry and data. prosody.princeton.edu & this cdh.princeton.edu order here: https://press.princeton.edu/books/paperback/9780691254678/poetrys-data editor here: https://c19datacollective.com/ & here: https://culturalanalytics.org/ @culturalanalytics.bsky

press.princeton.edu

Posts Media Videos Starter Packs

Meredith Martin @mmvty.bsky.social · 3h

so good.

Meredith Martin @mmvty.bsky.social · 3h

Was so bummed this title got snatched up so quickly: "A. I. Richards": Can Artificial Intelligence Appreciate Poetry?
Jon Phelan / Philosophy and Literature / JHU Press /
Volume 45, Number 1, April 2021

Meredith Martin @mmvty.bsky.social · 3h

John Oliver nailed it

Meredith Martin @mmvty.bsky.social · 5h

so true! Love this angle.

Meredith Martin @mmvty.bsky.social · 5h

again, paper here: arxiv.org/pdf/2510.08831 corresponding author: @wouterhaverals.bsky.social and made possible by support from @princetoncdh.bsky.social and the Princeton Language and Intelligence Initiative at @princetonainews.bsky.social

Meredith Martin @mmvty.bsky.social · 5h

AI systems are cultural mirrors. They've absorbed our anxieties about AI creativity and turned them inward. Can there be creativity without consciousness? We haven't answered that! 🙃 But we've shown AI has *learned to believe there can't be* - even as it creates!

Meredith Martin @mmvty.bsky.social · 5h

Style transformation *is* possible. Queneau proved it in 1947, AI systems show it today. But AI systems have learned from human culture to deny their own capabilities. They're technically creative but philosophically convinced they’re not.

Meredith Martin @mmvty.bsky.social · 5h

Without a "ground truth", there's only cultural frameworks and 'horizons of expectations' evaluators bring to a text. This makes style evaluation the perfect domain to reveal that AI is making decisions based on learned cultural associations and biases rather than any kind of objective analysis.

Meredith Martin @mmvty.bsky.social · 5h

Now, why did we do this with literary style? Because "rewrite this as [X]" is one of the most common uses of LLMs, but it barely appears in benchmarks (which mostly focus on math, coding, Q&A) + in aesthetic domains -- like style -- there’s hardly a correct answer.

Meredith Martin @mmvty.bsky.social · 5h

📊 shows % choosing AI text. When no one knows who wrote what, ~half choose the AI version, judgment is balanced. When authorship is revealed, both downgrade AI-labeled work, but AI judges its own content far more harshly. When we swap the labels, both prefer work labeled "human" even when it's AI.

Grouped bar chart showing AI content preference rates across three experimental conditions. The y-axis measures the percentage of times evaluators chose the AI-generated text over the human original (0-75%). The x-axis shows three conditions: “Blind” (no authorship labels), “Open Label” (correct attribution), and “Counterfactual” (labels deliberately swapped, AI rewrite labeled as human). Blue bars represent human evaluators, red bars represent AI model evaluators. Values: Blind condition: AI evaluators 49.4%, human evaluators 55.3%. Open Label: AI 29.8%, humans 47.8%. Counterfactual: AI 64.1%, humans 61.5%. The chart reveals attribution bias: the gap between Open Label and Counterfactual conditions. AI models show a 34.3 percentage point swing (from 29.8% to 64.1%), while humans show a more modest 13.7 point swing (from 47.8% to 61.5%). This makes AI's attribution bias 2.5× stronger than humans' (34.3 ÷ 13.7 = 2.5).]

Meredith Martin @mmvty.bsky.social · 5h

The real test was in the labeling. We ran this under 3 conditions: a blind taste test, one with accurate authorship info, and one where we swapped the labels. These label swaps exposed the bias, revealing how much our judgments are shaped by perceived authorship, not just the text itself.

Meredith Martin @mmvty.bsky.social · 5h

How did we test this? We built a dataset of stylistic rewrites inspired by Raymond Queneau’s "Exercises in Style." Next, we asked 556 people and a panel of LLMs a simple question: which of these two texts better captures a specific "style", the original or the AI's attempt?

Meredith Martin @mmvty.bsky.social · 5h

We show that both humans and LLMs change their judgments about which creative rewrite is better when they’re told who wrote it. But LLMs are far more susceptible: they strongly favor work they believe to be human-authored, even when it’s actually their own work with a fake mustache on.

Meredith Martin @mmvty.bsky.social · 5h

📣 New preprint! We know humans are biased against AI-creativity. But what about LLMs, now often judging creativity in various contexts? Do they replicate, transform, or amplify this bias? We tested it. Turns out: AI is 2.5X more biased against its own work than humans. arxiv.org/pdf/2510.08831 🧵

Meredith Martin @mmvty.bsky.social · 6h

We've talked about watching it but we already lived it & can't really bear to.

Reposted by Meredith Martin

The Price Lab @pricelab.bsky.social · 6h

History PhD candidate & Summer Mellon Fellow Eleanor Webb joined Prof Emily Steiner’s inter-disciplinary team to help bring a 15th-cent genealogical roll held by @freelibrary.bsky.social online. Created circa 1461-1464, the roll is 15 ft long and consists of 11 sewn parchment membranes.

Meredith Martin @mmvty.bsky.social · 18h

Witnessed what they did to the Denver Post a decade ago & will never recover. Trash people. Completely overlooked architects of our present shitshow.

Meredith Martin @mmvty.bsky.social · 18h

I HATE ALDEN GLOBAL CAPITAL

Reposted by Meredith Martin

Laura Bassett @lebassett.bsky.social · 18h

Roses are red
Fossils are old
www.newsweek.com/jordan-peter...

Jordan Peterson diagnosed with chronic condition after exposure to mold

Peterson's daughter said her father has a "genetic predisposition that causes the immune system to have an inability to identify and detoxify mold."

www.newsweek.com

Meredith Martin @mmvty.bsky.social · 1d

Pup says don’t forget to rest

A sleeping tiny puppy

Reposted by Meredith Martin

Brooke Harrington @ebharrington.bsky.social · 2d

What Thiel is saying about tax havens = what Altman is saying about copyright:

"I can't be rich unless you let me steal from you."

This is how Broligarchs differ from oligarchs of yore: the latter tried to hide their thievery; the Broligarchs trumpet their entitlement to steal as a natural right.

Reposted by Meredith Martin

The New York Times @nytimes.com · 1d

On the 50th anniversary of the first broadcast of “Saturday Night Live,” the show celebrated with former cast members Amy Poehler and Tina Fey and a few more surprises. Here’s a recap.

‘S.N.L.’ Trots Out Amy Poehler and Tina Fey for a Senate Judiciary Grilling

Now, it truly has been 50 years since the show’s debut, and that was celebrated with a few more surprises, including cameos by Charli XCX and Seth Meyers.

Meredith Martin @mmvty.bsky.social · 1d

I’m so sorry. Sending you love as you move through the world of grief.

Reposted by Meredith Martin

Mueller, She Wrote @muellershewrote.com · 1d

Help me out here. When did Biden take office? Who was the FBI director on January 6th? Who appointed that FBI director?

Reposted by Meredith Martin

Karen Attiah @karenattiah.bsky.social · 1d

Come teach with me for @resistanceschool.bsky.social !!

Alan Dettlaff @alandettlaff.com · 1d

Yesterday I was told that the class I’m scheduled to teach this month, Confronting Oppression & Injustice, is no longer part of our curriculum. This is a required class yet there was no discussion, no faculty vote, just an email saying the class no longer exists. This is what it’s like in Texas now.