Emily M. Bender

Practicing my blocking skillz --- today's target: Rando on the internet who thought he had a good suggestion for what should go into a paper I obliquely referenced.

@jasonkoebler.bsky.social

10h

"AI" (specifically "law and AI") in bio, ofc.

Emily M. Bender reposted

Jason Koebler

17h

the idea of some random guy making his chatbot write a press release about how his chatbot is a bad employee then emailing it to a bunch of journalists is still cracking me up, it’s so good. I really liked this one

Jason Koebler@jasonkoebler.bsky.social

The wildest one of these I got was a three PDF "press release" rollout from someone who said his ChatGPT sometimes got something wrong and that OpenAI didn't respond to his support ticket and fashioned himself as a kind of big tech whistleblower

Thinking about the argument from convenience for LLMs & related technologies in the context of a paper I'm working on. And how people might say, "Without [software], it would take to long to ___" My usual reaction to this has been: What did you do four years ago? >>

And when the pressure to do them is external --- that's likely a workplace conditions issue and one we shouldn't just roll over and accept. (Even if non-acceptance looks like playing along for now, while organizing for better.)

p.s. As always, the typos prove the authenticity. Sigh.

But I think there are additional directions to go here: It's okay for things to take time, it's okay to just not do some things (and in some cases even preferable). >>

My standing search on the phrase "stochastic parrots" regularly turns up people labeling it a "thought-terminating cliché". I think maybe they're right, but not the way the think they are: >>

The phrase has become a symbol for those they perceive as heretics (though usually without acknowledging that, since one feature of this cult is that it sees its faith as logic). >>

I guess my point here is, that's on them, not on me. When I coined this phrase, I could see that I was dealing with hype (even in late 2020) and general lack of understanding of how language works & how we work with language. I didn't know I'd be dealing with a cult. /fin

My standing search on the phrase "stochastic parrots" regularly turns up people labeling it a "thought-terminating cliché". I think maybe they're right, but not the way the think they are: >>

bloodinthemachine.comThe two words the AI industry hates more than anythingI talk with computational linguist Emily M. Bender, co-author of the influential 2021 “stochastic parrots” paper, about why AI industry leaders can’t stop seething over the term.

So in the face of a short, catchy phrase that undermines all of that, they are angry. www.bloodinthemachine.com/p/the-two-wo... >>

That anger doesn't stem from the phrase itself, of course, but from the challenge it presents to their belief system. Rather than engage with the critique, rather than look beyond the marketing of the tech companies/the gospel of the TESCREALists, they opt to self-terminate their toughts. >>

My standing search on the phrase "stochastic parrots" regularly turns up people labeling it a "thought-terminating cliché". I think maybe they're right, but not the way the think they are: >>

For all they will also point to things like code "assistants" (highly engineered harnesses around LLMs) as proof of the effectiveness of their beloved software, the idea that scaling up the systems designed to output plausible looking text will get to "AGI" is still core to the belief system. >>

Because even though a system for outputting code is also key to the myth of recursive self-improvement, I'm guessing at the end of the day, the chatbots that seem to speak our language remain a really important hook. >>

My standing search on the phrase "stochastic parrots" regularly turns up people labeling it a "thought-terminating cliché". I think maybe they're right, but not the way the think they are: >>

I haven't looked deeply at every case, but the overall gestalt is that this remark tends to come from people who are deeply committed to the idea/ideology of "AI" as something all-powerful that will either solve all of our problems or kill us all. >>

Emily M. Bender reposted

Replying to @emilymbender.bsky.social

I think the idea of automated peer review rests on both wishful thinking about "AI" systems (gimme those magic beans) and a misunderstanding of what we're doing when we review papers. >>

Emily M. Bender reposted

Replying to @emilymbender.bsky.social

So in the end, it doesn't matter what kinds of experiments people do comparing review-shaped objects extruded from LLMs to actual reviews written by people. Comparing them is and always will be a category error. >>

Emily M. Bender reposted

Replying to @emilymbender.bsky.social

There are no magic beans, and any scientific community that behaves as if there were ... isn't doing science anymore. /fin

Emily M. Bender reposted

insidehighered.comGenerative AI Is Not Our PeerAutomation can make many good things possible, but let’s not pretend it’s something it isn’t.

Replying to @emilymbender.bsky.social

But notice that none of the above use cases involve chucking the manuscript into an automated system and asking for any evaluation of it, nor "draft a review for me", nor "check this paper for correctness". Because that's not how peer review (nor science) works! bsky.app/profile/emil... >>

Emily M. Bender@emilymbender.bsky.social

"I understand that there have been some experiments showing that LLM review may have similar outcomes to human peer review, but who cares? Either scholarship, research and publishing are human enterprises or they aren’t." www.insidehighered.com/opinion/colu... >>

Emily M. Bender reposted

Replying to @emilymbender.bsky.social

While we're wishing, can I also have world peace, an immediate end to the climate crisis, a teleportation machine, two extra hours in every day and an endless supply of dark chocolate? >>

Emily M. Bender reposted

It seems like the #nlproc community is overdue for a conversation about how we should and shouldn't use text processing in reviewing, and also, I guess, what peer review & science even are. #emnlp2026 , 🧵>>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

So we have to look elsewhere for solutions to the crisis in peer reviewing, and that elsewhere will involve lot of person-to-person work, in mentoring, in adapting systems, in raising standards of what counts as an MPU, and in organizing against the bean counters who incentivize over-submission. >>

There are no magic beans, and any scientific community that behaves as if there were ... isn't doing science anymore. /fin

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

The words aren't the work, not in teaching, not in writing, and not in peer review. >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

This applies to all parts of scholarship, including peer review. Science is an enormous conversation, happening at many scales and through many channels (in classrooms, in conference presentations, in the hallway track, and in peer review). >>

youtube.comResisting Dehumanization in the Age of "AI": The View from the Humanities, Emily M. BenderYouTube video by Simpson Center

Attempting to replace some of those interlocutors with machines is the same kind of dehumanizing move that appears anytime someone suggests that we are replaceable because a machine can output words similar to ones we might say. www.youtube.com/watch?v=T7Lc... >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

Not only are all of these contingent notions, they also misframe what it is we are doing when we review papers. Yes, we are evaluating papers, but we are evaluating them not against some ground truth of the world, but against our own understanding, as scientists and scholars, of our fields. >>

This is a fundamentally human process because scholarship is a human process. There is no knowing without knowers. bsky.app/profile/emil... >>

Emily M. Bender@emilymbender.bsky.social

A beautiful essay indeed, and great jumping off point for people across all fields of scholarship to reflect on what it is we are doing and why. A short thread >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

I think the idea of automated peer review rests on both wishful thinking about "AI" systems (gimme those magic beans) and a misunderstanding of what we're doing when we review papers. >>

Some folks seem to have the idea that there is an objective truth as to which papers "should" be accepted into the conference, that such papers are the ones that are "correct", "novel" and "interesting". (Yeah, objectively interesting...) >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

Likewise, we need appropriate expectations on how much background reading a reviewer might be doing when considering each paper in their pile. >>

insidehighered.comGenerative AI Is Not Our PeerAutomation can make many good things possible, but let’s not pretend it’s something it isn’t.

Emily M. Bender@emilymbender.bsky.social

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

For each of these, of course, there are some cautions. 1) I can imagine cases where e.g. a position paper might get especially hostile reviewers through text similarity. 2) To the extent that these rely on translation into some formal language, there is room for error. >>

3) We need to calibrate expectations on just how thorough literature review can possibly be. Reviewers recommending a paper get rejected because they used a tool that turned up some paper the authors didn't cite could be problematic, especially if reviewers don't read it to check. >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

More seriously, here are some ways I can imagine using text processing and/or theorem-proving & other mathematical support in reviewing: >>

1) Matching papers to reviewers, based on similarity metrics between paper to review and reviewers' work. 2) Checking proofs (and maybe other equations) 3) Finding relevant further papers in some repository >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

So I understand the wish for some way to magically get fair, thorough, "objective" reviews from a source that furthermore doesn't ever get tired. >>

While we're wishing, can I also have world peace, an immediate end to the climate crisis, a teleportation machine, two extra hours in every day and an endless supply of dark chocolate? >>

Maria Antoniak@mariaa.bsky.social

A dataset of the responses will hopefully be coming soon. Summary of the results:

Let me start by saying I understand the pain of peer review, from all sides. Authors often feel like their paper was handled by unqualified and/or not careful reviewers & ACs. Reviewers have impossible numbers of papers to get through. >>

And the people bearing the brunt of the massive increase of interest in ACL conferences are the volunteers managing the whole ARR process. That's difficult, time-consuming, often thankless work. >>

techpolicy.pressThe Edge Dialogues with Dr. Alondra NelsonThe first in a series of events on ideas upstream of policy hosted by Tech Policy Press and the Privacy Center at Georgetown Law.

Really enjoyed this ep of @techpolicypress.bsky.social 's The Sunday Show with @alondra.bsky.social , Emily Tucker and @justinhendrix.bsky.social www.techpolicy.press/the-edge-dia... >>

"but the alternative is to realize that the people driving the bus are headed off the cliff and that it's up to you to stop them, which is really scary." >>

"marketing by gaslighting" is a fantastic phrase to describe the inevitability argument in all its forms ("don't get left behind", "it's here, get used to it", etc etc). Thanks, Emily Tucker!

techpolicy.pressThe Edge Dialogues with Dr. Alondra NelsonThe first in a series of events on ideas upstream of policy hosted by Tech Policy Press and the Privacy Center at Georgetown Law.

Really enjoyed this ep of @techpolicypress.bsky.social 's The Sunday Show with @alondra.bsky.social , Emily Tucker and @justinhendrix.bsky.social www.techpolicy.press/the-edge-dia... >>

Absolutely banger quote from Emily Tucker towards the beginning: >>

"There's a kind of marketing by gaslighting that's going on where the industry is feeding us a rehash of stale ideas and telling us we're on the cutting edge and to make yourself believe that entails a kind of soul destroying nihilism, >>