A phenomenon I've noticed recently is people trying to occupy some untenable middleground wrt to the use of systems sold as "AI" -- this is a position where people try to recognize the harms of this tech but also hold space for "responsible" or "ethical" use.
🧵>>
When someone is trying to hold this untenable position, a few things tend to come up (not everytime, not everyone):
1- Defensiveness. People read criticism of the systems and proposed uses of the systems as accusations that users are "bad people". Thus a criticism of the tech lands as criticism of the user, and tensions flare.
>>
2- Righteousness. People do have legitimate needs, often unmet needs, and the synthetic text extruding machines can *look like* a solution. But just because the problems are real doesn't mean the solution is beneficial, effective, or worth (not always externalized) costs. Unfortunately, pointing out any of this is taken as the same as saying you don't care about the legitimate needs.
>>
3- Whataboutism. This is used to brush off concerns about the externtalities of these systems. You eat meat, you fly on airplanes, etc, etc, how dare you talk about the impacts of data centers?
>>
4- Tone policing. People who are trying to occupy that uncomfortable, untenable space will claim that clear statements of harms/strong principles against use of these systems will "turn others away" as if the centrists are the ones actually pushing for more ethical practice.
But this "other people won't listen" remark I think is really a way of saying "This makes me uncomfortable" while trying to claim to be on the right side of history at the same time.
>>
5- Wishcasting. Some folks will point to scientific results from fields outside their own (usually media coverage thereof) that are marketed as having been done with "AI" and ask: How could you take a hardline against "AI" when it has provided XYZ?
>>
6- Exceptionalism. "I know this can be dangerous for people in general, but I know how to use it carefully."/"I know how to verify every output, and I am not deskilling myself." How do you know? Also, if you acknowledge the dangers to others, what example are you setting by talking about/talking up your use?
>>
So what is the best way out of that uncomfortable, untenable space? I think one key step is disaggregating the (non-coherent) set of technologies sold as "AI". If you don't call the stuff you work with "AI", you aren't saddled with trying to defend any of the rest of it.
>>
The most recent iteration of this conversation I was involved in turned in part on a strange, over-expansive definition of "genAI" which included, for ex, optical character recognition (OCR).
>>
OCR can be a useful tool for many research projects! OCR is also the kind of technology that gets better with better language models, i.e. more fine-grained models of which word(parts) go where. That has been true since before "genAI" and will be true after.
Just because you can use the synthetic media extruding machines to approximate the task of OCR, however, doesn't mean that that task can or should be used to justify the use of "genAI" in research.
>>
I think another important step is a values examination. What is important to you? How are those values supported or not by entering the discourse in a way that holds space for OpenAI/Anthropic/Google/Meta and all the other actors in this massive push to shove "AI" into every part of our lives as "not all bad"?
What are your research goals, what do you value about participating in scholarship, how can you meet those goals/act in accordance with those values and what obstacles are in your way?
>>
Part of what makes that middle ground untenable and uncomfortable, I think, is that it requires carrying water for these clearly bad actors. You can set that bucket down and step out onto firmer ground.
This does require going against the mainstream, but that gets easier when a) you find you're not alone and b) you see how much of mainstream opinion on this is actually the result of marketing.
/fin for now
@emilymbender good thread. very good thread. many thanks for it. we do see this tension play out a lot, it's helpful to have positive advice for people in the middle of it.
Human beings have a genius for mean-spirited subtle insults.
In a recent meeting, a coworker was profusely praised for using CoPilot to summarize some very dry & voluminous documentation. It was kind of over the top.
Soon my puzzled coworker clued that they were obliquely putting her down. They were implying she was lazy and insufficiently bright enough to accomplish the task without a crutch like AI.
What's concerning was Copilot praising her in a similar fashion
@ireneista @emilymbender problem is: the whole world is crazy. It's difficult to stay sane. The fight against AI is the fight against capitalism. Thats difficult when one is employed in a capitalist society.
@emilymbender Good thread, not sure if I am in alignment with every detail, but certainly with the overall philosophy.
@emilymbender
Good thread!
@emilymbender Heard an interview today with two of the execs from Anthropic. The way they describe their products is maddening. They use words like thinking, intuition, and understanding. None of that is going on. I was screaming at the podcast.
@xvf17 @emilymbender Eh, don't get worked up. There's lots of bad stuff out there, no use wasting time on hate.
Denounce it, work against it, but hate is way too much effort only to the detriment of your well being.
I'd see hate being useful to stir to action at a rally but the podcast is pretty much ether.
@emilymbender @cohentheblue Not hate, just frustration. The people building this stuff should understand it, is my view.
@xvf17 @emilymbender Oh, don't you worry, they understand. Games inside games, hands washing hands etc.
Jun 24, 2026, 11:22 AM@cohentheblue @xvf17 @emilymbender You do the fascists' job for them friend, but you can stop at any time.
@pip @xvf17 @emilymbender You wot? By advising people not to wind themselves up where nothing other than damaging themselves is the result, I do the job of fascists? What are you on about...
Jun 24, 2026, 6:42 PM@cohentheblue @xvf17 @emilymbender
The result of not using AI is protecting yourself, your family, and your community. The result of not using AI is harm reduction, battling against the devastating social, political, and environmental costs of AI. The result of not using AI, is making it easier for others to refuse to use it as well.
So please, stop using AI.
@pip @xvf17 @emilymbender Where did you get that any of us uses AI?
Have you even glanced at our profiles for that matter?
You come across as an LLM yourself.
Nothing more frustrating than being misunderstood deliberately and repeatedly.
@xvf17 @emilymbender THIS IS SO UPSETTING. I am one of those folks advocating for using certain types of machine learning responsibly, like "identify the patients with cancer" which ML tools can do better than humans, and which were trained using ethical data sources.
But the LLM field is just a cesspool. I can *see* use cases for it, but holy shit fruit of the corrupt root system with that training data.
@SomeVeganCheeseIsOk
Oh no dude. This prof thinks you are an asshole then.
Read what she wrote again. She is full of fallacies herself.
@xvf17 @emilymbender
@Noisecolor @xvf17 @emilymbender I don't see that, really. She asks us to make certain specific things more clear in our speech, which is GOOD because "AI" as a term is so overused and awful, and you can't make good decisions on bad information. LLMs *are* a nightmare and *are* heavily overused and *do* have horrendous ethical issues. But machine learning is not always "AI", and I like people being asked to make the distinction because it's an important step.
@SomeVeganCheeseIsOk
Nightmare, heavily, horrendous... I don't think those are words that make anything clear.
I don't know what ai is when you put it in quotes.
I guess the good ai is the one you like and the bad ai is what bad people use or something?
@xvf17 @emilymbender
@SomeVeganCheeseIsOk @emilymbender @Noisecolor LLMs of today are not AI, in the traditional sense. The ChatGPT folks coopted the term. People (like me) who put it in quotes are doing it as a form of protest.
@xvf17
As far as I know AI was coined as simulated intelligence. I think LLMs are exactly that.
Ai is a term that we used to describe a bunch of systems. Chess engines, opponents in video games,... Now LLMs. Perhaps someday when we have even better systems we will call them ai, but for now ai seems to fit LLMs completely.
@SomeVeganCheeseIsOk @emilymbender @Noisecolor they are not simulated intelligence. That’s precisely my point.
@xvf17
Really? 😀
How is it not a simulation of intelligence?
@SomeVeganCheeseIsOk @emilymbender
@Noisecolor @xvf17 @SomeVeganCheeseIsOk @emilymbender AI is literally the abbreviation of *artificial* intelligence. Not simulated intelligence. That feels like a real "moving the goalposts" moment.
AI as it has historically/originally been used is referencing a legitimate, non-organic intelligence. LLMs are smoke and mirrors, when it comes to intelligence. There is no cognition, no understand.
Personally, I prefer GenAI over scare quotes, but both are more accurate than LLM = AI
@theadhocracy
Not at all. You can freely browse the web or ask a chat who coined the term and why. It will be an interesting read I can assure you and you will understand that ai is not meant to refer to a specific human level intelligence like you mean. It wouldn't make sense at that time anyway.
@Noisecolor @xvf17 @emilymbender talk to me about what you mean by the term simulated. Because I can see a use case where it means "fake, but enough to fool a human" and a use case for it just being a synonym for artificial. One of those is the traditional use, the other is a marketing use.
@SomeVeganCheeseIsOk
It's not a marketing term. We have always used this term. There was no marketing team involved.
Humans need words to describe certain stuff.
When we do choose those words those words are assigned that meaning. It's linguistics, it has nothing to do with who you think is a good or bad guy or what a certain group might like or not.
In certain times words can change or the meaning of words can change. But in this case it's pretty clear I think.
@Noisecolor @xvf17 @emilymbender
Let me clarify: "AI" is an unbounded marketing term with wildly ambitious, ambiguous definitions. No machine performs contextual reasoning like a human at this time, but that's the implication of "AI" to most folks. Machine Learning is a bounded, defined, useful term which talks about compute systems with specific capabilities.
@Noisecolor @xvf17 @emilymbender Large Language Models are a form of machine learning that was built on stolen data by companies who, in part, design it to capture and hold attention for as long as possible, much like social media, to extract paid tokens from users. LLMs are the 1-900 numbers of this age. But worse, because they are simultaneously somehow a social disaster undermining human skills, an ecological disaster, and an economic disaster.
@Noisecolor @xvf17 @emilymbender How you do things matters. What you do with things matters. And everything counts in large amounts.
@SomeVeganCheeseIsOk
But that's not precise nor clear at all. You are mixing tech itself with I don't know what that is, some kind of preaching? It sure sounds like you found your external enemy you can pin everything on and hate together with a group. Like a mob or a cult
@SomeVeganCheeseIsOk
It's not a marketing term. It's an old term that people have been using for a while. And I don't think it's ever been used to describe a machine person that's alive. It was always used as simulated intelligence. Which LLMs are. It's a very good description.
@xvf17 @emilymbender The tools do it too and claim they are "thinking" when asking users to wait for their output.
Jun 23, 2026, 8:51 PM@emilymbender Great thread.
Jun 23, 2026, 9:22 PM@emilymbender great thread, I think at different points I've run into each of the points you laid out (especially when I was trying to be more amenable to LLM users). since, I've grown to taking a much more hard-line anti-big-tech-LLM, neutral-to-meh academic/open-model LLM, and recommending searching out all alternatives before turning to the tech as a solution.
thank you for your work in this area, as you and @alex have helped shaped my opinions/knowledge on the subject.
Jun 23, 2026, 10:02 PMi despise AI because of its energy waste, its abuse of source material it rips off without attribution, and how it concentrates power with these amoral techbros who can use it for all sorts of malicious goals
but i always wondered if AI could be a standalone thing, something you ran yourself, at home, so there is no exposure to manipulative outside agendas, if i would have the same objections. assuming sane power usage and respect for creators. and assuming complete privacy
@benroyce @emilymbender
Yes, you can use Ollama and free LLMs from hugging face.
It's slow but it stays private on your own machine. You don't even need a GPU to do it.
Jun 23, 2026, 10:32 PM@hackersquirrel @emilymbender and people don't do it because it's not as powerful? I'm certain that will change. But then the big tech offerings will be more powerful too. Bht if the private AI does enough, good enough, then I suppose we can talk about the evils of corporate AI alone, what we just call "AI" today
It's a start. Once Nvidia no longer has the techbros sucking on its tit, they'll be begging for folks to buy their GPUs.
Jun 23, 2026, 10:39 PM@hackersquirrel @emilymbender and we'll finally be able to buy RAM and SSDs again
@benroyce @hackersquirrel @emilymbender no, because vast amounts of compute are still necessary to train these "open-data" models. Their impact on energy consumption and hardware waste is maybe somewhat less than that of proprietary models, but it is still significant
@benroyce @hackersquirrel @emilymbender folks avoid local language models b/c some combination of:
- don't know about them
- they seem complicated
- usually aren't well integrated into easy-to-consume applications and interface
- the outputs aren't as good as what you get from commercial services, because you're using a smaller quantized model; or
- you need to buy expensive hardware to support running the biggest and best models; or
- the output is sloooooow, like a word per second or worse
Jun 24, 2026, 11:15 AM@ryanprior @benroyce @hackersquirrel @emilymbender
Yes, and they share most of the harms of cloud models anyway.
@benroyce @hackersquirrel @emilymbender IME you need substantial computer muscle for it to run models large enough to be useful. Not that it can't be done, not at all, but it's more accessible if you already have a powerful rig for other reasons.
And then, it doesn't necessarily improve some of the issues. Since models need lots of ram, it's actually less useful to fill, say 32 gigs of ram to serve one user, times 10000 users, than it is to have 320 terabytes of ram serving 10000 users.
@benroyce @hackersquirrel @emilymbender Cause the second solution allows use to split cpu time between users, whereas the first one has one cpu per user that spends most of its time idling. So energy-wise you're not necessarily winning either unless your use case involves AI running around the clock, but then while you might be more *efficient*, you're still using gobs more energy.
Jun 24, 2026, 1:34 AM@renardboy @hackersquirrel @emilymbender
well distant future
like we walk around with computers in our pockets that are 10,000x what we used to send astronauts to the moon. in 50 years, we'll walk around with petabytes and petaflops in our pockets (or stuck in our heads... cyborgs)
@benroyce @hackersquirrel @emilymbender Personally, my hope for better generative "AI" involves distributed models running on home servers that channel their heat usefully.
You'd have a bunch of servers with water cooling that feeds into appliances like hot water heaters, and each would hold some focused fragment of a large language model, and when you'd prompt you'd have some kind of router language model that'd send it to whoever is likely to have an answer.
@benroyce @hackersquirrel @emilymbender I don't dare to expect this but I allow myself to hope.
It's not that far-fetched as I understand, there are existing LLMs that are internally structured as components, they are called "Mixture-of-Experts" (MoE) LLMs.
@benroyce @hackersquirrel @emilymbender all the energy use would displace energy normally used simply to run resistive heaters in hot water heaters. You'd have models sized for a number of average power consumptions (and corresponding heat outputs) to match with water use patterns of different buildings.