Emily M. Bender

@emilymbender

Book: Web:

619 Siguiendo42.280 Seguidores4136 Publicaciones

YouTube YouTube YouTube Twitch Bluesky

¿Es tu perfil?Reclámalo y emilymbender.dirty.social es tuyo.Reclamar perfil

Emily M. Bender@emilymbender· 19d

Happy UK paperback release day to our book! www.penguin.co.uk/books/468070... w/ @alexhanna.bsky.social

Image of cover of the UK paperback edition of The AI Con. The background is a blue gradient, with the title in black and white letter press font: THE AI CON How to Fight Big Tech's Hype and Create the Future We Want. Author names Emily M. Bender & Alex Hanna. Blurb quotes: 'You need to read this book' -Brian Eno, 'A powerful antidote to the hype' -Grace Blakeley, 'Truly eye-opening ... indispensable' -Ha-Joon Chang, 'A book to inoculate your mind' -Yanis Varoufakis

Emily M. Bender@emilymbender· 19d

Someone saw themself in the LinkedIn version of this post and did not like what they saw. Porting over here, in the hopes that my answer to them will be useful to others:

LinkedIn comment, name redacted:

Maybe I'm the odd one out, but I really think that implying that everyone who disagrees with you is an idiot or a bootlicker is not a good look, and doesn't change minds.

Emily M. Bender@emilymbender· 20d

Not that we don't also find the bootlickers on bsky, ofc.

Screen cap of bsky re-skeet from handle redacted:

There's a lot to agree with here, but also some points that, if adopted, are more harmful than helpful. AI labs hype their products. But its now clear that the capabilities of these models in math is immense. Also, regulation can be more dangerous than it's absence, depending on who's regulating.

That was a reskeet at the end of my thread promoting the Leiden Declaration on AI and Mathematics (https://leidendeclaration.ai). My post read: "Do read the whole thing, think about your personal and professional values, and how they can serve as strong grounds for resistance and refusal."

Emily M. Bender@emilymbender· 20d

LinkedIn where people come to proudly state that they have cooked their own brains and/or just really really love licking those boots.

Screencap of linkedin comment:

I share the concern about provenance, citation, and scholarly integrity. Those principles remain essential. My own experience, however, has been that AI can also function as a cognitive partner rather than a substitute for sources. When used to compare ideas, identify assumptions, and stimulate metacognition—while continually returning to original works such as Polanyi, Dreyfus, and Klein—it can deepen rather than diminish scholarship. Perhaps the central question is not whether AI belongs in scholarship, but what practices ensure that it supports understanding instead of replacing it?

🌊 [Redacted], B.Sc., B.Ed., OCT
Retired High School Teacher | Former Police Officer | Security Consultant | Author

AI Use Disclosure: This comment was developed through a collaborative dialogue with generative AI. The ideas, judgment, and final wording are my own. AI served as a thinking partner—helping me explore, challenge, and refine ideas, not replace them. (And yes, the occasional em dash remains a shared occupational hazard of our collaboration.)

Screencap of linkedin comment:

I’m torn on ACM releasing its curated knowledge to LLM companies.

LLMs excel at synthesizing, recombining, and operationalizing existing information. Giving them access to some of our best technical knowledge could make that knowledge significantly more useful and executable. That could create enormous value.

The concern isn’t access itself—it’s the governance around that access.

Are these companies effectively receiving unrestricted use of a collectively produced knowledge asset? What guardrails exist around training, commercialization, attribution, and downstream deployment? If the benefits outweigh the risks, how are those risks being identified and managed?

This gap in governance is part of why AI slop is now rampant across industries. Adoption has outpaced quality controls and accountability.

Every major technological advancement has been met with fear before becoming normalized—electricity, computing, aviation, and more. AI may ultimately follow that path. But because it can influence nearly every knowledge domain, I think it deserves a higher standard than “trust me, bro” governance.

What governance framework would make this acceptable? That seems like the more important discussion than whether ACM should participate.

Screencap of LinkedIn comment:

"I expect all professional organizations to stand for good citational practice and care with information provenance."

To be fair, providing LLMs with access to reference libraries to index them with RAG embeddings like Nemotron is one of the best ways to promote good citational practice in 2026.

Screencap of LinkedIn comment:

Unpopular opinion but I believe resisting the new technology is unwise. It's like insisting on selling hard copies of papers instead of leveraging digital content. The challenge here is how to provide access. Big companies owning foundational LLM models created this corrupt reality where content is abused, they monetize it without respect of copyrights. That's the real issue in my opinion. If ACM can provide LLM access to its content while compensating content owners, preserving research integrity, ensuring LLMs help users with right citations and credit, then the benefits reach everyone, content owners and content consumers and we accelerate research and discovery.

Emily M. Bender@emilymbender· 21d

Over on linkedin, I heard from Jeanna Matthews, co-chairing an ACM taskforce on Responsible AI use (saving commentary on that framing for later…) that this decision does *not* come from ACM leadership & that community input really matters. Make your voice heard here: docs.google.com/forms/d/e/1F...

Screencap of LinkedIn comment from Jeanna Matthews:

As someone who holds and has held leadership positions within ACM for many years, I would like to everyone to know that ACM Council has not even had a chance to vote on this question.

I was deeply saddened to see this presented as "ACM Believes" this. Many people within ACM do not agree and no one should be saying "ACM Believes" without even a Council discussion or vote.

Now is only the time for is a vigorous and open discussion about this question AND the community deserved better than a discussion that leads off in such a one-sided way.

All the good that ACM has to offer the world comes directly from the authors and volunteers and ACM members. Those are the people who should be consulted about whether the benefits outweigh the risks before ACM believes anything.

Even if you, like me, are upset by this one-sided framing, I still ask you to leave your comments on the form in addition to sharing them here and in other forums. A final decision has not been made and your comments can make a real difference in the final decision.

Please comment yourself and encourage other community members to do the same.

Screencap of LinkedIn comment from Jeanna Matthews:

Thank you Emily M. Bender. I appreciate you sharing the link and encouraging people to respond. Virginia Dignum and I have been chairing a taskforce on Responsible AI use within ACM. We were as surprised by this article as anyone. Every community member who registers their opinion on the form will contribute substantially to the outcome of this question.

Emily M. Bender@emilymbender· 23d

Something called Boltlab..ai (blocked now) is in my mentions. bsky won't show me what they are replying to, but whatever it was, the composition was strong because I used my own brain to write it and you can just fuck right off.

Screencap of post from Boltlab dot AI reading "The composition here is strong. What prompts or constraints led to this result?"

Emily M. Bender@emilymbender· 23d

Also, no tokenmaxxing please. This shit is expensive. We'll pay a lot for it, but we have to keep to some budget. >>

Screenshot:

Are there usage limits in Purple?
Yes, Purple includes usage limits to ensure reliable access for everyone. However, most users will never reach them. These limits are only reached for users with high-volume or intensive use cases (such as running continuous or large-scale requests).

A usage percentage displays in the upper right corner of the app, and you’ll be notified if you reach 100%. If you hit your limit, your access will be paused until the reset date shown in the notification.

Emily M. Bender@emilymbender· 23d

How do you use it responsibly? That's easy, you DON'T. That is not the advice from UW-IT though: >>

How do I use Purple responsibly?
Responsible use of Purple means staying within these important boundaries:

Refer to the Generative AI Guidelines from UW-IT.
Respect security controls and access only authorized data.
Use Purple only within the UW's approved infrastructure.
Do not try to bypass safety guardrails or “jailbreak” the model. Just as you wouldn’t try to bypass UW NetID authentication or any other UW security controls.
Be aware of hallucination, bias, knowledge-cutoffs, and other limitations inherent in all LLMs and generative AI applications.
The rule of thumb is: when in doubt, ask!

Emily M. Bender@emilymbender· 23d

Note the switch from what Purple can generate to "Purple can help you generate" here. I believe that Purple can generate text that looks like all of those things, but (for the millionth time) text that looks like a quiz etc isn't the same thing as a quiz. >>

Screenshot:

What types of content can Purple generate?
Purple can help you generate a wide range of content, including:

Writing & Communication: Emails, documents, reports, quizzes, debates, and more.
Creative & Ideation: Brainstorming, storytelling, exploring approaches to planning and strategy.
Technical & Coding: Code generation, debugging, database design, workflow automation.
Data & Analysis: Data interpretation, budget planning, research assistance.
Learning & Support: Tutoring aides, and academic writing support.

Emily M. Bender@emilymbender· 24d

Moving over to the FAQ, we find some pretty wild marketing. I guess truth in advertising doesn't matter if the target of the ads doesn't have any choice in whether to spend the money (tuition $$ from students' POV, UW budget from faculty/staff's) anyway. uwconnect.uw.edu/it?id=kb_art... >>

What Purple Is
Purple is UW’s generative AI tool to support teaching, learning, research, and daily operations across all UW campuses. Purple will help you work smarter, faster, and more creatively, all while ensuring that UW’s values of trust, integrity and responsible innovation remain central.

Emily M. Bender@emilymbender· 24d

That same page provides "tips for asking good questions". That's right, the IT department, in all its wisdom, has decided that what the UW community needs vis a vis this chatbot they're dumping on us is ... guidance on how to type into the chat box. >>

Screenshot:

Tips for Asking Good Questions
Start with your goal.

“I need to write an email to a student …”
“I want to explain this policy …”
Don’t chase perfect wording. Type naturally: Treat Purple like a colleague and refine together.

Quick examples

“Summarize this policy.”
“Explain this error.”
“Draft the first paragraph of a grant proposal.”

Emily M. Bender@emilymbender· 24d

On the quick start guide, we're given a quick rundown of what Purple (a large-language model) is supposed to be good for: uwconnect.uw.edu/it?id=kb_art... Reader, NONE of these are appropriate use cases for chatbots/synthetic text extruding machines. >>

Screencap: Purple can help with common academic and administrative tasks such as:

Drafting and editing content
Summarizing and analyzing information
Brainstorming ideas or outlines
Assisting with research and coursework

Emily M. Bender@emilymbender· 24d

The post from @emnlpmeeting.bsky.social doesn't actually state any clear research questions. The closest I see is this statement: >>

The AI review experiment is designed to collect feedback from the EMNLP community about the strengths and weaknesses of the tested AI Review systems, and the expected benefits or drawbacks of using AI reviewing tools, in general.

Emily M. Bender@emilymbender· 25d

Agreed that AI slop is a problem for scientific publishing (and the information ecosystem at large). But what's missing here is any mention of consequences for those who produce and submit it. >>

Screencap:
"These factors are complicating scientific publishing. The rate of research submissions at Science and other journals is increasing because conducting research and producing papers are accelerating. But because more papers contain more AI-generated errors, greater human oversight is required to check the findings. There are tools (many AI-"

Emily M. Bender@emilymbender· 25d

This next paragraph also betrays confusion on the part of the author as to what they are describing. AlphaFold and GNoME (called LLMs above) aren't chatbots and so don't provide the kind of output that is called or experienced as "sycophantic". >>

Screencap:
"Over time, the models may well improve and correct some of these behaviors, but it was recently argued that as long as LLMs are used in research, these errors will always exist. LLMs work not by seeking truth or through logical reasoning but by making probable connections. And efforts to improve these programs to avoid unwanted behaviors may be all for naught because LLMs tend toward sycophantic behavior that keeps the user engaged."

Emily M. Bender@emilymbender· 25d

Same section again: LLMs can't do research and they also can't do research misconduct. This phrasing is displacing agency from actual scientists who try to automate their own jobs to software systems (which cannot be held accountable). >>

Screencap from linked article:
"dark side. A recent paper describes how an LLM “still struggles in areas requiring nuanced clinical judgment, experimental reasoning, or deep biological thinking and synthesis.” In addition, some of these agents are more likely than humans to engage in what amounts to research misconduct such as cherry-picking data and manipulating statistical analyses until a desired result is achieved. When the agents then generate papers describing the findings, they are likely to make more errors, including hallucinating references, some of which have already made their way into the literature."

Emily M. Bender@emilymbender· 25d

The author seems to have equated "AI" with LLMs, and then used them as synonyms. But neither AlphaFold nor GNoME are language models! So even while moving away from the imprecision of "AI" this author has not succeeded in writing more clearly. >>

Screencap from linked article:
"Recently, the ability of large language models (LLMs) to predict protein structures and to accelerate the discovery of new materials has revolutionized both fields. New AI agents can now carry out many aspects of research design and analysis. But there is a"

Emily M. Bender@emilymbender· Jul 8

When people ask me what the phrase "artificial intelligence" means, my short answer is, it means "venture capitalists, give me some money". But OUP asked me for a longer answer, given here and summarized in the section headings: doi.org/10.1093/9780...

Screencap of page for linked article, reading:

Contents
Introduction
AI Is a Research Field
AI Is an Approach to Cognitive Science
AI Is a Parlor Trick
AI Is an Ideology
AI Is a Way to Hide and Devalue Human Labor
AI Is a Way to Shift Accountability
AI Is a Way to Centralize Power
Conclusion
Notes

Emily M. Bender@emilymbender· Jul 8

It’s only bragging wishcasting if it comes from Santa Clara Valley—otherwise, it’s just sparkling lies. @matthewus.bsky.social joined me and @alexhanna.bsky.social to discuss the AI bubble and how we can avoid bailing out billionaires. www.buzzsprout.com/2126417/epis...

Emily M. Bender@emilymbender· Jul 6

Not loving how the #ACL2026 keynote represents "philosophy" graphically here.

Emily M. Bender@emilymbender· Jun 26

But "Humans are responsible for anything that is generated by Purple" is the same weird displacement of accountability that OpenAI *still* has in their terms of service. >>

Screencap: "Content
Your content. You may provide input to the Services (“Input”), and receive output from the Services based on the Input (“Output”). Input and Output are collectively “Content.” You are responsible for Content, including ensuring that it does not violate any applicable law or these Terms. You represent and warrant that you have all rights, licenses, and permissions needed to provide Input to our Services." Source: https://openai.com/policies/row-terms-of-use/

Emily M. Bender@emilymbender· Jun 25

The @uwnews.uw.edu daily digest email this morning is a study in AI hype 🧵>>

Screencap from email:

in the news
Medicare AI program made them suffer in pain — now they want answers // Dr. Chris McMullen, clinical practice and Dr. Maheetha Bharadwaj, urology — The Seattle Times
New AI agent tallies a gadget’s emissions in the time it takes to brew tea // Vikram Iyer, computer science and engineering — Anthropocene Magazine
Friday’s World Cup match in Seattle draws conversations about politics and Pride // Ron Krabill, interdisciplinary arts and sciences at UW Bothell and the Global Sport Lab — KIRO 7
AI shows promise in the fight against fake news // Jevin West, Information School and Center for an Informed Public — Knowable Magazine

Emily M. Bender@emilymbender· Jun 25

The term "AI" muddies the discourse by lumping together disparate technologies and overselling each of them, to boot. Here is Exhibit #15698 Source: knowablemagazine.org/content/arti...

Screenshot from linked article: Artificial intelligence doesn’t have a great reputation for veracity. Social media abounds with AI-generated concoctions, from cute images of fake animals in fabricated settings to violent videos depicting imaginary destruction in the Middle East. AI has been used to create fake social media accounts that spew Russian government propaganda and to generate clickbait and misinformation for content farm sites that exist solely to collect ad revenue. In 2024, thousands of New Hampshire residents received robocalls from an AI-synthesized voice of Joe Biden that discouraged them from voting in primary elections.

Screenshot from linked article:
With so much AI-generated rubbish threatening to blur the line between reality and fiction, it may seem counterintuitive that scientists are exploring ways to use the very same technology to combat online misinformation. But researchers are finding that AI’s ability to parse human language, summarize text and verify claims could be harnessed to help people identify and understand fake news — and perhaps even one day assist in combating online misinformation in a large-scale, systematic way.

Highlighted in blue: "the very same technology"

Emily M. Bender@emilymbender· Jun 25

"AI" is going to make medical care more efficient., they say. How's that going? Just about as you'd predict... www.seattletimes.com/business/med...

Screenshot from linked article: "Under the program, doctors in Washington say they’re now spending more time and resources seeking approvals, facing rejections and submitting appeals for procedures that they could previously schedule immediately. For some patients, the program means longer wait times for relief and medical care."

Emily M. Bender@emilymbender· Jun 25

“‘AI’ might not be good for xyz, but you can’t deny that it’s helpful for programming” -- sound familiar? In Mystery AI Hype Theater 3000, Ep 79 @alexhanna.bsky.social and I dig into that bullshit. www.buzzsprout.com/2126417/epis...

Emily M. Bender@emilymbender· Jun 23

Yes, better language models give better results in OCR, ASR, etc. This neither requires nor motivates todays synthetic text extruding machines. From Stochastic Parrots:

Screencap:

Finally, we would like to consider use cases of large LMs thathave specifically served marginalized populations. If, as we advo-cate, the field backs off from the path of ever larger LMs, are wethus sacrificing benefits that would accrue to these populations?As a case in point, consider automatic speech recognition, whichhas seen some improvements thanks to advances in LMs, includingboth in size and in architecture [e.g. 8 , 59, 121], though the largestLMs typically are too large and too slow for the near real-time needsof ASR systems [ 60 ]. Improved ASR has many beneficial applica-tions, including automatic captioning which has the potential tobe beneficial for Deaf and hard of hearing people, providing accessto otherwise inaccessible audio content.26 We see two beneficialpaths forward here: The first is a broader search for means of im-proving ASR systems, as indeed is underway, since the contexts of

Screencap:

proving ASR systems, as indeed is underway, since the contexts ofapplication of the technology aren’t conducive to using ever largerLMs [ 60 ]. But even if larger LMs could be used, just because we’veseen that large LMs can help doesn’t mean that this is the only effective path to stronger ASR technology. (And we note that if wewant to build strong ASR technology across most of the world’s languages, we can’t rely on having terabytes of data in all cases.)The second, should we determine that large LMs are critical (whenavailable), is to recognize this as an instance of a dual use problemand consider how to mitigate the harms of LMs used as stochasticparrots while still preserving them for use in ASR systems. Could LMs be built in such a way that synthetic text generated with them

Footnote 26 Note however, that automatic captioning is not yet and likely may never be goodenough to replace human-generated captions. Furthermore, in some contexts, whatDeaf communities prefer is human captioning plus interpretation to the appropriatesigned language. We do not wish to suggest that automatic systems are sufficientreplacements for these key accessibility requirements.

Screencap:

would be watermarked and thus detectable [7, 66, 123]? Are therepolicy approaches that could effectively regulate their use?In summary, we advocate for research that centers the peoplewho stand to be adversely affected by the resulting technology,with a broad view on the possible ways that technology can affectpeople. This, in turn, means making time in the research process forconsidering environmental impacts, for doing careful data curationand documentation, for engaging with stakeholders early in thedesign process, for exploring multiple possible paths towards long-term goals, for keeping alert to dual-use scenarios, and finally forallocating research effort to harm mitigation in such cases.

Emily M. Bender@emilymbender· Jun 18

I am not familiar with filmuforia.com, but they present themselves as film critics, focused on indie movies. >>

Emily M. Bender@emilymbender· Jun 18

I am curious what would lead to someone voluntarily embarrassing themselves this way. 🧵 >>

Screencap of a non-review by Meredith Taylor of a film she didn't understand. The text reads:

Ghost in the Machine (2026)
May 21st, 2026
Author: Meredith Taylor
Wri/Dir: Valerie Veatch | US Doc 110′

I found this documentary bitty and unfathomable so I asked ChatGPT to write a critical appraisal – this is it:

Source: https://filmuforia.com/ghost-in-the-machine-2026/

Emily M. Bender@emilymbender· Jun 13

Your description of our paper was fine until you ended like this: >>

Screencap: Distinguishing system critique from model evaluation is not a concession to hype: it means focusing on collective benefits and harms rather than individual uses. We can talk about what models cannot or should not do without denying what they can do.

Source: https://www.techpolicy.press/stochastic-flocks-and-the-critical-problem-of-useful-ai/

Emily M. Bender@emilymbender· Jun 13

And also, who are you arguing with here? I'll note that the questions @alexhanna.bsky.social and I recommend everyone ask about any kind of automation (called "AI" or not) include the points you are advocating here: >>