October 18th, 2026
This post is one of my advice & arguments pages about the harms and hazards of the AI Hype Movement.
A while ago, Cory Doctorow revealed that he uses a local large language model to check for typos in his posts. To pull out some quotes from that article:
Here’s the discussion of using an LLM:
““” There is one technology that has made my POSSE life better, and it might surprise you. This year, I installed Ollama – an open-source LLM – on my laptop. It runs pretty well, even without a GPU. Every day, before I run Loren’s python publication scripts, I run the text through Ollama as a typo-catcher (my prompt is “find typos”). Ollama always spots three or four of these, usually stuff like missing punctuation, or forgotten words, or double words (“the the next thing”) or typos that are still valid words (“of top of everything else”). ““”
Note that these are all tasks at which standard, non-LLM spelling and grammar checkers have excelled for years. A few of these actually benefit from some kind of “model of language”, but they tend to work better with small and precisely-applied language models than with large stochastic ones. For completely incomprehensible reasons, Doctorow has decided to use a very inefficient and harmful tool to do a very simple job. He goes on to write:
““” Ollama is a pretty good typo-catcher. Probably half of the “errors” it points out are false positives, which is better than the false positive rate for Google Docs’ grammar-checker. ““”
Well that’s interesting. Confirms exactly that the tool isn’t very good at this job, and compares accuracy at a typo-and-grammar check to accuracy of a grammar checker, which is the harder part of that task. Then he acknowledges the elephant in the room:
““” Doubtless some of you are affronted by my modest use of an LLM. You think that LLMs are “fruits of the poisoned tree” and must be eschewed because they are saturated with the sin of their origins. I think this is a very bad take, the kind of rathole that purity culture always ends up in. ““”
The comparison to “purity culture” is telling here, since it means that Doctrow experiences the LLM-averse among us (like me) as puritan scolds, which is clearly the logic he’s using internally to avoid actually engaging with our arguments about harms. By deciding that our moral objections are unreasonably narrow, he has found a way to ignore them entirely. Incidentally, this is a favorite tactic of the right-wing: painting requests for inclusion or equality as overreaching or extreme. Doctorow continues with this logic, saying:
““” That’s how we make good tech: not by insisting that all its inputs be free from sin, but by purging that wickedness by liberating the technology from its monstrous forebears and making free and open versions of it […] ““”
Sadly, “open-source LLMs” are mostly a fantasy because of the material inputs necessary to create and sustain them, and even if they were possible, their non-harmful applications would be extremely narrow. This isn’t a technology that we can “liberate” without bringing a lot of its harms along for the ride. That’s like saying “just make sure handguns are freely available to everyone and all the ethical issues are fixed.” I’ve got my own complicated thoughts about gun control and oppressive governments, but even I think there’s a lot more nuance necessary than that. And of course the specific technology matters here! Free scissors for everyone? Sure. Free paper, or free water, excellent? Things like the dangers inherent in the technology, the real costs of its production, and the ends to which the technology can be put matter a lot when deciding whether we should “liberate” it for everyone. Even in the realm of software where such actions are relatively more possible, a free music player app is different from a free prison administration app, or a free app for silently tracking people using their phones.
Doctorow continues on this line of argument, turning into a deflection:
““” Refusing to run an LLM on your laptop because you don’t like Sam Altman is as foolish as refusing to get monoclonal antibodies because James Watson was a racist nutjob ““”
Although I definitely don’t like Sam Altman, that’s not my primary reason for not running an LLM on my laptop, so this is a straw man argument that plays into the “purity culture” aspersion cast earlier. But Doctorow really wants to justify his own use of LLMs, and says:
““” The AI bubble sucks. AI itself is a normal technology ““”
The problem is, it’s very much not. Per the examples above, technologies are not inherently neutral, which is what Doctorow is trying to claim here and what a lot of mainstream thinking loves to imagine (spurred on by inventors who don’t want people to consider the politics of their inventions). Technologies are in fact inherently political, and their politics depend on both their intrinsic properties and on how they’re actually invented, produced, and sustained in the world.
Doctorow continues with this doozy:
““” It’s not “unethical” to scrape the web in order to create and analyze data-sets. That’s just “a search engine” ““”
Doctorow is perfectly aware of the difference between extremely
aggressive scrapers run by the AI hype faction which ignore conventions
like robots.txt and scrape even pages that hosters try
actively to hide away, and normal search engine scrapers which treat the
digital commons with respect. I’m pretty sure he’s even written about
these problems in the past. But now in order to psychologically justify
supporting those harms, he dismisses them and disingenuously equates
them with much more benign behavior.
Doctorow proceeds to try to narrow the harms of AI purely into the economic realm:
““” There’s plenty of useful things people can do with AI. There’s plenty of useful things people will do with AI. AI is bad because it’s an economic bubble and a grift, but not because we’ve created a bunch of utilities that would – under normal circumstances – be called “plug-ins” ““”
In a way, this sounds similar to my position on the main page: that I’m opposed to the hype movement but not to AI itself. However, what I mean to exclude from criticism is AI broadly, whereas I think that LLMs specifically are indeed a “wicked technology”.
It’s possible to just tackle Doctorow’s argument head-on here. He wants us to see LLMs as a “normal” technology that’s not “wicked.” I set out to show here why it is in fact “wicked.”
LLMs as designed/deployed by the handful of companies big enough to invest in frontier models plus the gaggle of hangers-on developing “open” models are in fact an “exceptionally wicked” technology for at least three different important reasons, and only if you intentionally confound them with other adjacent technologies can you escape these.
Large Language Models cannot be built without massive amounts of plagiarism. As someone who is a proponent of information sharing and piracy at times, what do I mean by plagiarism here, and why not use the word “theft”? I mean not that “the files are stolen” since “stealing” a file by copying it doesn’t immediately or directly harm the person who has the original copy. I mean that the intellectual investment and chance for engagement with others is stolen when an LLM regurgitates a victim’s work without making any connection to them that the “user” of the AI can see. It’s a bigger loss for the AI user in a way because they’re missing out on knowing who the author of the idea is and any opportunity to explore similar ideas, understand the context and politics of that author, etc. But it’s also a loss for the author, since they lose the opportunity to connect with that reader.
If I were to copy Doctorow’s blog every day and republish it under my own name, he’d be justifiably mad, and I’d be injuring him, with or without the legal framework of copyright. The fact that when LLMs do this they (often but not always) mix together ideas from multiple sources doesn’t change the logic at all. A plagiarism free LLM-adjacent technology that we could have been building with even a fraction of the effort invested into “AI” would be essentially a better search engine that would give you properly-sourced quotes of relevant ideas along with authorship information and links in response to a query. It could even still use a transformer model on the inside. But that’s a different technology than the one that is being built, and that opportunity cost is also part of the wickedness of LLMs.
LLMs are a “wicked” technology because they cannot exist without an underclass of psychologically scarred knowledge workers, and because the knowledge-filtering these workers do also helps centralize control of information in a way that asymmetrically aids causes which need to rely on propaganda for survival, like fascism. To be clear, you can build a “medium-sized language model” that’s trained only on known-good data sources that are already curated by more-ethical means (e.g., Wikipedia, even though even its curation practices aren’t perfect). You cannot build a Large Language Model that way. You could probably build a large language model over the course of a decade or so and with a few billion dollars by hiring and fairly compensating curators, including taking appropriate precautions around all the psychologically dangerous stuff they’d need to experience in order to filter it out, but that model would always remain on average several years out-of-date, which is why I consider it a meaningfully different technology than what we have. Nobody I know of has even proposed building such a thing, with all of the “ethical AI” projects that address training data taking massive shortcuts and/or choosing very different tradeoffs.
Even if you could wave aside the labor issue as Doctorow is now doing, the centralization of control inherent in the filtering step by itself makes the technology “wicked”. When a single LLM producer gets to control the policies and training that the knowledge filterers work from, it gets to inject all kinds of biases into the result. Just look at the blatant white supremacism of Grok for an example of what can be achieved. The fact that OpenAI and Anthropic have chosen biases slightly more acceptable to the average neoliberal does not change this fact.
Now some might say here that I’ve strayed into criticizing the use of the technology rather than the technology itself. But the fact is that large language models (which to everyone’s surprise a few years ago turned out to have meaningfully different capabilities than medium language models) cannot be built without this centralization of control, making them inherently more dangerous than technologies that don’t have that property. Is “inherently more dangerous” the same as “wicked”? Yes in this case, because technologies don’t exist in a vacuum. LLMs have other properties (like requiring massive training resources) which ensure that the sociopolitical context of their production will lead to harmful weaponization of the inherent danger at issue here (among others).
Even in the most optimistic case, the ways that centralized knowledge filtering amplifies the social biases of the filterers is its own related issue. Is there a possible adjacent technology that isn’t wicked in this specific way? Yes! It’s called “the internet”. The interface is a bit different (and quite a bit more ergodic and it doesn’t have exactly the same use cases, of course) but it does have distributed and collaborative knowledge filtering that users are in control over. Does this mean that you are responsible for not clicking on that link to goatse, rather than some underpaid worker leveraged into the role by centuries of colonialism who will be fired when the psychological toll of previewing all the goatse links for those who outsource to them is too much? Yes, and that’s why it isn’t as “wicked” in this way.
(Whew. What of the many remaining wickednesses should I pick as number three? There are plenty to choose from (see, for example my main page on the AI hype movement).
LLMs are a phenomenally energy-inefficient technology for all of their meaningful use-cases. This makes them a “wicked” technology for our specific situation in the midst of a climate emergency. There are two common paths of rebuttal here: first that this is a situational wickedness that’s not inherent to the technology, and second is that if some use-cases are okay, then the technology itself can’t be deemed wicked.
To address the first: the situations in which this critique does not apply are unfortunately not available to us on planet Earth right now. This makes the invention of the technology (let alone proposals for its use) wicked. I don’t mean that the instant the idea of an LLM crosses someone’s mind they’re a bad person; invention is a drawn-out process that’s much more than just ideation. I do mean that seriously proposing and working towards the realization of such technology in a moment of deep climate crisis is wicked, even if those working on it might be too biased to see that.
But again does this mean the technology itself is wicked? Yes, because we cannot separate technologies from the context of their invention and use. Platonic forms don’t exist, and neither do platonically-context-free technologies. Another way to see this: if you lived in a world in which inventing LLMs was not wicked (let’s say: 10,000 years from now, climate crisis already solved by the arrival of magical fairies, who also power fusion reactors giving access to essentially unlimited energy at zero cost), could you possibly invent the same technology? No. There would be innumerable differences for a lot of reasons, but most relevant: the existence of basically-unlimited energy would allow you to invent and actualize much better technologies. In any scenario in which the environmentally-destructive backend of LLM architectures was somehow not necessary, LLMs as we know them would already be obsolete. I’ll admit that this isn’t an ironclad logical proof, and some readers may try to imagine a separation between invention-process and invention-itself to cast the invention-itself as not-wicked, even understanding that an “invention” or “technology” is much more than an “idea”. If you’re in that camp, I think you probably need to sharpen your understanding of the politics of technology, but I’ll let you off at 2 dimensions of wickedness for this post.
To address the second rebuttal to the resource consumption issue: while there are some uses of an LLM that might justify its resource intensity, these largely don’t overlap with the uses to which the technology is being put, or the ones it’s marketed for. Remember that a “technology” is much more than an idea: it encompasses production and distribution systems, as well as the perception in the popular consciousness that drives actual use. The main tasks LLMs are perceived as being useful for are all tasks for which their resource-intensity is wildly unjustified, and this makes it a “wicked” technology. If you could strip out all that marketing and have people generally accept it as a kind of niche system whose inefficiency warrants use in only a handful of ways, you’d then come to the conclusion that it’s not really worth all the effort just for those cases, when other technologies exist that are much easier to deploy.
This kind of opportunity cost is why technologies aren’t fully separable from their marketing: if we didn’t have the fiction of LLMs as magic chatbots, Doctorow would still be using conventional spelling and grammar checkers, because without that fiction nobody would have invested enough to build an LLM and we’d have gotten better technologies instead.