Emily M. Bender
LightNews LightNews
banner
emilymbender.bsky.social
Emily M. Bender
@emilymbender.bsky.social
36K followers 570 following 2.2K posts
Book: https://thecon.ai

Web: https://faculty.washington.edu/ebender
Posts Replies Media Videos
emilymbender.bsky.social
Feeling annoyed at the cognitive cost of having to dismiss all the offers of "AI" assitance every time I use Acrobat to provide feedback on documents my students wrote. (NO I do NOT want an "AI" summary of this. In what world???)

Decided to check settings:
Image of "Preferences" pain for Adobe Acrobat, with the "Generative AI" menu item selected. This provides check boxes that can be unchecked, including "Enable generative AI features in Acrobat" and also "Show AI query bar on document", "Show AI video cards (this will replace the starred documetns, requires restart)", and "Show AI Summary Bar on top of the document".
November 10, 2025 at 6:37 PM
emilymbender.bsky.social
Beautiful tree on my walk this morning
Japanese maple tree with leaves in shades of red and gold, early morning light streaming through almost horitzontally
November 9, 2025 at 1:17 AM
emilymbender.bsky.social
Wow -- look at OpenAI trying to shift responsibility to the users:

🧵>>

www.ctvnews.ca/sci-tech/art...
Screencap reading: "OpenAI dispelled suggestions that it’s changing its terms around legal and medical advice. In a statement to CTVNews.ca, it said “this is not a new change to our terms. ChatGPT has never been a substitute for professional legal or medical advice, but it will continue to be a great resource to help people understand legal and health information.” The AI research company updated its usage policies on Oct. 29 to clarify that users of ChatGPT can’t use the service for “tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional.”"
November 6, 2025 at 1:09 AM
emilymbender.bsky.social
And we do see that happening!

But sometimes it is also so totally cringe, as in today's (in print, Nov 10th's) New Yorker piece:

>>
Screencap reading: There is understandable resistance to such a simplistic and triumphant account of A.I. The case against it was well argued by Ted Chiang, who wrote an article for this magazine in early 2023 titled “ChatGPT Is a Blurry JPEG of the Web.” He meant it in a more or less deflationary way: that’s all ChatGPT is. You feed the whole internet to a program and it regurgitates it back to you imperfectly, like a copy of a copy of a photograph—but with just enough facility to fool you into believing that the program is intelligent. This spring, a similar argument was made in a book, “The AI Con,” by Emily M. Bender, a linguist, and Alex Hanna, a sociologist. Bender is perhaps best known for describing L.L.M.s as “stochastic parrots.” “Large language models do not, cannot, and will not ‘understand’ anything at all,” the writer Tyler Austin Harper declared in a book review in The Atlantic. Models “produce writing not by thinking but by making statistically informed guesses about which lexical item is likely to follow another.” Harper buttressed these technical arguments with moral ones. A.I. enriches the powerful, consumes enough energy to accelerate climate change, and marginalizes workers. He concluded that “the foundation of the AI industry is a scam.” Source: https://www.newyorker.com/magazine/2025/11/10/the-case-that-ai-is-thinking
November 3, 2025 at 2:52 PM
emilymbender.bsky.social
In case anyone is curious how Kurtis replied:
Reply from Kurtis Loken in the same LinkedIn thread, reading: Emily M. Bender I am trying to learn. You are insulting me. Why? To me it seems you have checked YOUR brain at the door if you won’t even debate one of these supposed flawed and humanity killing machines. What if your assumptions don’t actually fit the results you claim…will you keep your draconian stance? Energy consumption aside, the arguments that AI will be a negative vs a net plus is something that has not been proven at all or people wouldn’t build it. I know you won’t agree with that last sentence.
November 3, 2025 at 4:59 AM
emilymbender.bsky.social
Screenshot of the title author list and publication date of Bender, Gebru et al 2021 "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"
November 2, 2025 at 3:59 PM
emilymbender.bsky.social
Every day it feels more and more like a zombie apocalypse over on LinkedIn, in which you start a conversation with someone who you hope is an ordinary person and then a couple of turns in discover that they have surrendered their brains to the "AI" mass delusion.
Comment on a linked in post, from user Kurtis Loken who presents themself as "Experienced Metrology Engineer | AI/ML real world experience | Patented Solutions | R&D | Mass Production" Comment reads: Emily M. Bender You are a linguist. LLMs have proven some theories disproved others about how language works. I would think that would be an amazing time to be a linguist. Have you ever actually talked to an AI…I mean really talked to one? The Turing test has been passed. (Sonnet 4.5 has particularly impressed me when discussing ideas with it…definitely has passed my own Turing test.) I would really love to see a conversation between yourself and it discussing your views. That would make a great podcast. Have these models explain to you why your idea to ban them may or may not be a good idea. These things probably have their own opinions on the topic if someone would listen to them. Maybe that’s you?
November 2, 2025 at 2:34 PM
emilymbender.bsky.social
Mystery AI Hype Theater 3000, Episode 65:
www.buzzsprout.com/2126417/epis...

… in which @deccamuldowney.bsky.social joins me and @alexhanna.bsky.social to talk about what it means to actually do journalism.

Thanks to Ozzy Llinas Goodman for production!
October 31, 2025 at 1:02 PM
emilymbender.bsky.social
I'd love to see someone try to estimate just how much time and money has gone into research that is either fully undermined by reliance on LLMs or fully pointless --- because obvious if you start from an understanding of what LLMs actually are.

www.pnas.org/doi/10.1073/...
Screen cap from linked article, with heading Significance and then text: Large Language Models (LLMs) are used in evaluative tasks across domains. Yet, what appears as alignment with human or expert judgments may conceal a deeper shift in how “judgment” itself is operationalized. Using news outlets as a controlled benchmark, we compare six LLMs to expert ratings and human evaluations under an identical, structured framework. While models often match expert outputs, our results suggest that they may rely on lexical associations and statistical priors rather than contextual reasoning or normative criteria. We term this divergence epistemia: the illusion of knowledge emerging when surface plausibility replaces verification. Our findings suggest not only performance asymmetries but also a shift in the heuristics underlying evaluative processes, raising fundamental questions about delegating judgment to LLMs. Sentence starting with "While models often" is highlighted in blue.
October 18, 2025 at 10:11 AM
emilymbender.bsky.social
Good morning from McLeod Ganj, near Dharamsala India
A monkey sits on a green pole under Tibetan prayer flags, with a forest in the background. The monkey has a black face and light fur. Its tail hangs down below. A vertical shot of a grey mountain with a light dusting of snow behind green forest covered foothills. The foreground has evergreen trees.
October 13, 2025 at 3:37 AM
emilymbender.bsky.social
Here's a rule of thumb: If "AI" seems like a good solution, you are probably both misjudging what the "AI" can do and misframing the problem.

>>
Comment by Tom Diettrich on a linkedin post reading: "You can't "test-in quality" in engineering; you can't "review-in quality" in research. We need incentives for people to do better research. Our system today assumes that 75% of submitted papers are low quality, and it is probably right (I'll bet it is higher). If this were a manufacturing organization, an 75% defect rate would result in bankruptcy. Imagine a world in which you could have an AI system check the correctness/quality of your paper. If your paper passed that bar, then it could be published (say, on arXiv). Subsequent human review could assess its importance to the field. In such a system, authors would be incentivized to satisfy the AI system. This will lead to searching for exploits in the AI system. A possible solution is to select the AI evaluator at random from a large pool and limit the number of permitted submissions. I imagine our colleagues in mechanism design can improve on this idea." Original: https://www.linkedin.com/feed/update/urn:li:activity:7381685800549257216/?commentUrn=urn%3Ali%3Acomment%3A(activity%3A7381685800549257216%2C7382628060044599296)&dashCommentUrn=urn%3Ali%3Afsd_comment%3A(7382628060044599296%2Curn%3Ali%3Aactivity%3A7381685800549257216)
October 11, 2025 at 12:13 PM
emilymbender.bsky.social
Oh FFS. Because what overburdened peer review systems really need is more noise between the reviewers & ACs. Yeah, that's the ticket.

aaai.org/aaai-launche...

>>
Screenshot of AAAI automated review announcement reading: 1. Supplementary First-Stage Reviews: LLM-generated reviews will be included as one component of the initial review stage, providing an additional perspective alongside traditional human expert evaluations. 2. Discussion Summary Assistance: LLMs will assist the Senior Program Committee (SPC) members by summarizing reviewer discussions, helping to highlight key points of consensus and disagreement among human reviewers.
October 6, 2025 at 7:17 PM
emilymbender.bsky.social
As a birthday bonus, here’s one more audiogram for Mystery AI Hype Theater 3000 Ep 63

www.buzzsprout.com/2126417/epis...

w/ @alexhanna.bsky.social
production by Ozzy Llinas Goodman
October 2, 2025 at 1:07 PM
emilymbender.bsky.social
Happy birthday to our little pod! In Mystery AI Hype Theater 3000 Ep 63 @alexhanna.bsky.social and I reflect on three years (three years!) of ridicule as praxis & answer listener questions.

www.buzzsprout.com/2126417/epis...

Thanks to Ozzy Llinas Goodman for production!
October 2, 2025 at 1:01 PM
emilymbender.bsky.social
Coming up in Seattle (and live streaming). Registration:

www.washington.edu/lectures/eve...

w/@alexhanna.bsky.social
Graphic advertising up coming talk 10.21.25 The Graduate School Office of Public Lectures University of Washington The AI Con How to fight big tech's hype and create the future we want Emily M. Bender Professor of Linguistics UW Alex Hanna Director of Research, Distributed AI Research Institute (DAIR) Graphics include photos of Alex and Emily (jacket photos from the book). Alex is wearing hoop earrings and a black shirt and has her arms crossed. Emily is wearing glasses and small drop earrings and a blue/green patterned shirt. Both are faintly smiling and looking straight at the camera.
September 27, 2025 at 1:46 PM
emilymbender.bsky.social
And lol at this part -- in other words, depending on design, at best the "AI" didn't hurt?
Screencap: One often-cited study, a large-scale experiment with nearly 1,000 high-school students, measured how well students in three groups solved math problems: students who used no technology while solving them, those who could use an AI chatbot that functioned like regular ChatGPT, and those who could use an AI chatbot built with detailed instructions to act like a tutor, such as by giving students hints rather than answers to the problem. They found that students who used the standard AI chatbot performed better during practice sessions than students who had not used AI, but the standard-AI users performed worse when tested unassisted. Students who used the AI that functioned like a tutor performed the best during practice sessions and performed about the same on tests as students who did not use any AI.
September 26, 2025 at 2:18 PM
emilymbender.bsky.social
How does stuff like this make it past a fact checker in late 2025? "Accurate"? Really?

>>
Screencap: Tutoring technology has long intrigued educators because it offers the possibility that students can learn from an accurate and always-available guide. Now AI has made such tools much easier to create. Greg Kestin, the physics lecturer who created the Harvard AI tutor, said at the time of the study’s initial release that a lot of the students’ comments focused on how they “could go at their own pace and could have this unfettered access to a nonjudgmental instructor who would not get frustrated or annoyed no matter how many times they asked.”
September 26, 2025 at 2:16 PM
emilymbender.bsky.social
And this strikes me as especially disingenuous. If what's offloaded in the cognitive offloading is the educational experience then the education does. not. happen.
Screencap: Where our understanding — or misunderstanding — of brain activity intersects with generative AI often has to do with cognitive offloading, the process of relying on something else — whether it’s ChatGPT, Google Maps, a notepad, or another person — to remember things for us. That frees up our brains for other work, but whether that’s harmful or helpful isn’t a simple yes or no answer.
September 26, 2025 at 2:12 PM
emilymbender.bsky.social
Back to CHE:

NO ONE WAS SAYING THAT THE TELEPHONE OR VIDEO GAMES WERE GOING TO REVOLUTIONIZE EDUCATION AND WE HAD TO HAVE THE KIDS USE THEM TO BE PREPARED FOR THE FUTURE.
Screencap: New technology has always brought with it social panics. Commentators argued that the telephone would be ruinous to families because women would spend more time gossiping, says Lee Rainie, director of the Imagining the Digital Future Center at Elon University. The advent of video games led to fears of rampant violence and addiction. When the internet and search engines began to be an integral part of life, people worried about their effect on our memory and intelligence. As a viral 2008 article put it, “Is Google Making Us Stupid?”
September 26, 2025 at 2:10 PM
emilymbender.bsky.social
This part is mostly true. ("Technology is changing too fast" is bullshit though.)

But the article seems to be putting the burden of proof in the wrong place.

The companies selling this as ed tech are the ones who should be expected to show its benefits and safety.

>>
Screen cap from article, reading: And what about the dozens of other studies that also examine AI use? Can experiments even capture how technology is affecting the traits that make us human: our creativity, our analytical skills, and our overall intelligence? Teaching experts say it’s unlikely we’ll find definitive answers in a lab. Technology is changing too fast. AI is now embedded in so many products that it’s impossible to isolate how it is shaping us. And, perhaps most significantly, learning is a complex process that is hard to define, let alone measure. Trying to determine the interplay between our use of technology and our own thinking is tricky.
September 26, 2025 at 2:06 PM
emilymbender.bsky.social
Hey @nytpitchbot.bsky.social did you write this one for CHE?
Screen cap of newsletter from the Chronicle of Higher Education. Graphic is an image of a human brain sitting atop the white letters AI and casting a shadow below all against a yellow background. Text above the image: DIGITAL MINDS Professors Fear AI Will Rot Students’ Brains. The Research Shows It’s More Complicated Than That. By Beth McMurtrie Text below: Learning is a complex process — and so is measuring it. Though research shows we have cause to be concerned about what happens when students use AI, the devil is in the details.
September 26, 2025 at 1:55 PM
emilymbender.bsky.social
TFW your little accidental podcast hits a quarter million downloads!

Here's to the next quarter million, @alexhanna.bsky.social ! 🥳

And huge gratitude to producers @whatulysses.bsky.social and Ozzy Llinas Goodman!
Badge from Buzzsprout with text "250,000 Podcast Downloads" below an image of the Mystery AI Hype Theater 3000 logo against a white field with color confetti falling from the top.
September 16, 2025 at 5:35 PM
emilymbender.bsky.social
Mystery AI Hype Theater 3000 Episode 62

www.buzzsprout.com/2126417/epis...

Come to hear @mharrisondupre.bsky.social , @alexhanna.bsky.social and I dismantle the misplaced enthusiasm for chatbots in therapy, stay for the literal word sausages.
September 12, 2025 at 2:13 PM
emilymbender.bsky.social
This detail is also really haunting. Google claims to "organize the world's information." This is not that.

>>
screencap from linked article: One work day, her task was to enter details on chemotherapy options for bladder cancer, which haunted her because she wasn’t an expert on the subject. “I pictured a person sitting in their car finding out that they have bladder cancer and googling what I’m editing,” she said.
September 11, 2025 at 2:42 PM
emilymbender.bsky.social
Some important reporting on data work.

www.theguardian.com/technology/2...

Surprising but telling to see Google flat-out denying the contribution of data workers to their products:

>>
screencap from linked article: Google said in a statement: “Quality raters are employed by our suppliers and are temporarily assigned to provide external feedback on our products. Their ratings are one of many aggregated data points that help us measure how well our systems are working, but do not directly impact our algorithms or models.” GlobalLogic declined to comment for this story.
September 11, 2025 at 2:42 PM