locusmag.comCommentary: Cory Doctorow: TechnopoliticsCory Doctorow (Copyright Julia Galdo & Cody Cloud)
Itβs one thing to be βanti-AIβ in the sense of believing that AI-generated slop, disinformation and deepfakes are bad; that AI companies that hope to con your boss into firing you and replacing you with AI that canβt do your job are bad; that betting the economy on AI is bad; that betting the human-habitable planet Earth on AI is bad; that running roughshod over local environmental rules and seizing farmland for data centers is bad. Those are all good reasons to be βanti-AI.β
But being βanti-AIβ in the sense of βopposing the statistical inference techniques we call AIβ is weird. So is being anti-AI in the sense of βopposing the use of models that were created using large language corpuses because itβs inherently wrong to compute the frequency with which words occur in writΒten language.β And itβs also weird to be anti-AI in the sense of βopposing AI because someone, somewhere used AI in a way that was gross, harmful or exploitative.β
Today, the two kinds of βanti-AIβ are mostly a distinction without a difference, because nearly all the worldβs AI output comes from one of the βfrontier modelsβ made by a giant AI company. No matter why you oppose AI, you probably donβt like those frontier models.
But after the AI bubble bursts, the distinction will be much sharper. Long after the collapse of the seven AI companies that make up 35% of the S&P 500 stop trading the same $100 billion IOU around really fast while pretending itβs in all their bank accounts at once, there will still be cheap bankruptcy-sale hardware, legions of unemployed skilled AI workers, and the open-source models released by todayβs AI giants will still run, and will continue to be improved by the people who rely on them.
The useful things AI does will keep happening. Youβll still be able to prompt up a single-serving, personal piece of software to perform one task with moderate reliability for your use, and when it stops working, youβll be able to do it again. Youβll still be able to transcribe audio, or summarize a document. Youβll still be able to use a chatbot as an interactive journal, or as a supercharged spellchecker.
None of that will change β and neither will the limits on all those activities. If you donβt know enough about a document or its subject to spot obvious errors in an AI summary, youβll be periodically misled. If youβre so desperate or uninformed that you trust summaries you lack the capacity to evaluate in critical situations, you may be led to disaster. If you donβt know anything about computers and you turn a piece of vibe-coded software loose on your machine, you might delete something important, or irreversibly expose sensitive data to the public internet.
AI isnβt exceptional, itβs normal. Itβs not exceptionally dangerous. Itβs not exceptionally evil. Itβs also not exceptionally useful or exceptionally good. AI represents a significant advance to computer science β a moderate variation to existing machine learning techniques that generated a much larger improvement in those techniques.
In the absence of the AI investment bubble, weβd call the things AI can do βplug-insβ β as in, βI just got a new plug-in for my word processor that generates plausible word salads. I can see how that might be useful for generating placeholder text, and I can tell that a lot of people are going to use it in ways I find supremely annoying.β Iβve had a lot of plug-ins for my word processors over the past 40 years, and most of them fit that description. None of them ever made me want to bet the global economy on it, or to feed all the writers into a woodchipper.
AI can do many useful things. Some of them are _very_ useful: Friends who work in advanced statistics tell me that AI coding assistants help them produce high-quality, reliable code that solves real problems, improving the quality of the work they do. I find AI-generated transcripts of audio very useful for locating quotes in podcasts Iβve recently listened to, and I use a chatbot that runs on my computer to check my essays for typos. This bot often catches typos that my text editorβs checker misses, among a smattering of false positives that I have the experience and discernment to disregard (you know, like every spell checker).
But AI doesnβt exist in the absence of the bubble β it very much exists in the white-hot center of the most dangerous investment bubble in the history of markets. A year ago, the AI sectorβs total expenditures were about $700b. Today, theyβre over $1.4t, and Gartner projects that will rise to $2.4t by the end of the year. This is an astonishing sum of money to commit to a sector that grosses a mere $50b worldwide, annually. The sectorβs failure to produce anywhere near as much value as it is consuming has not dimmed its backersβ ardor, and they are committed to spending, and spending, and spending, until either they run out of investors or they somehow transform $50b worth of revenue on $2.4t in expenditures into profit.
To this end, the sectorβs masters are committing all manner of unforgivable sins. These range from conning your boss into firing you and replacing you with an AI that canβt do your job; to convincing your town council to seize your home and build a data center, without any environmental review or public input; to swindling everyday savers with nonsensical IPOs and sweetheart deals with stock exchanges that let insiders cash out their stupid AI bets with your retirement savings.
These sins are baked into this kind of AI production. There is no way to raise trillions in subsidies that allow AI companies to sell computation at a subsidy that amounts to selling $100 bills for $1 each without also insisting that the end state of all of this is that every job will be swallowed by a chatbot. You canβt raise trillions on this promise withΒout βblitzscaledβ data-centers β accompanied by climate-shredding gas turbines. You canβt raise trillions without cramming AI into every part of every personβs digital life, without any safeguards for how vulnerable people might use them to sink deep into dangerous delusion.
These are the βtechnopoliticsβ of AI: To make this AI, the people inΒvolved _had_ to inflict all these harms upon us. The important questions are: to what extent are these the _necessary_ technopolitics of AI, and; can those technopolitics change?
Langdon Winnerβs 1980 article, βDo Artifacts Have Politics?β is the Ur-text for questions like these. In this seminal essay, the rock critic turned tech scholar Winner demands that we move beyond the simple framing of tech having βintendedβ and βunintendedβ consequences, and instead insists that politics come baked into some technologies.
Winner divides the politics of technology into two broad categories: politics that are derived from the specific way that a technology has been introduced into the world; and politics that are either necessitated by, or strongly compatible with a technology by dint of its technical charΒacteristics.
For the first category β objects whose politics are determined by deΒployment decisions β Winner cites a set of notorious Long Island freeway overpasses. These were specified by Robert Moses, and in _The Power Broker_ , Robert A. Caroβs seminal biography of Moses, we learn that Moses delibΒerately chose to make these overpasses too low for a bus to pass beneath, in order to prevent a public transit route from being established between New Yorkβs Black neighborhoods and Long Islandβs public beaches, favored by the white bourgeoisie.
According to Winner, a freeway overpass isnβt necessarily political, but these overpasses are. Whatβs more, their politics is durable, outlasting Robert Mosesβs tenure as New Yorkβs most powerful political figure, even outliving Moses himself.
For the second category β technologies whose politics emerge from their intrinsic technical characteristics β Winner compares solar and nuclear power. Nuclear reactors, he argues, require a central authority to build, operate, and maintain (we might add here that they also require a conΒstant stream of fissiles that can be extracted only from certain unevenly distributed mining sites). By contrast, solar can be installed anywhere the sun shines, established and overseen by small, decentralized groups. For Winner, nukes are irreducibly the tech of large, centralized states; while solar is potentially the tech of autonomous, nonhierarchical societies.
Between these two technological poles, Winner discusses a third exΒample: a tomato harvester. The harvester enabled farmers to fire many of their pickers, giving power to bosses at the expense of workers. At the same time, the harvester only worked with hardier varietals that traded flavor for durability, yielding a cheaper, worse kind of tomato that drove out competing tomatoes that tasted better but cost more to raise because their cultivation wasnβt as amenable to cost-saving automation. The tomato harvester wasnβt created to weaken labor and worsen tomatoes, but that was its effect, and so the politics of this harvester included a certain kind of farm, a certain kind of labor relation, and a certain kind of foodway.
AI certainly has technopolitics. At a miniΒmum, AI has the technopolitics of the first kind, those Robert Moses bridges. It has been capitalized specifically on the basis that it can be used to erode workersβ wages and labor power and to foist inferior goods on the public. There is no other way that the AI sectorβs unimaginably vast capital losses can be turned into profits. The sums involved are so large that virtually everything everyone does, all the time, will need to be replaced with AI, and so it is also in AIβs technopolitics that AI must be crammed into every activity we undertake, whether or not it belongs there, and whether or not it presents a danger to the user, or to the people who depend on the userβs work (think here of cramming AI into automated systems for generating criminal accusations, which then prompt the police to make false arrests).
Now, none of this is intrinsic to the technoloΒgy. Remember, the technology weβve been calling βAIβ for the past decade is a set of small variaΒtions on existing machine learning techniques that yielded surprisingly large dividends relative to the size of those variations. This improvement was scalable β meaning that when computer scientists pushed these variations harder, they got concomitantly outsized improvements in the machine learning systemsβ performance. This is one of the best kinds of technological breakthrough to be in the midst of, one where you discover a piece of low-hanging fruit that turns out to be at the edge of a whole _orchard_ full of similarly low-hanging fruits.
Thereβs nothing _technical_ about the discovery of this variation on machine learning that dicΒtates the current AI approach. In an alternate universe, βAIβ could be a $100b joint supercomΒputing effort between the International Panel on Climate Change and Sandia National Labs, used to improve the fidelity of our climate models. This, too, would produce carbon emissions, but not nearly as many, and of course, everything people do has an impact on the environment (the only way to stop impacting the environment is to a) die; and b) decompose). This alternate uniΒverse might also see a proliferation of hobbyist open source / open weight models spinning out of these efforts, with online forums and social media filled with generated text and sounds and images. As with every technology controlled by its users, some users will find ways to sin against the law and ethics. There will be some gross stuff around the edge (voice cloning for fraud, deepfake porn, etc), but for the public, the main AI event would be a bunch of plug-ins that were byproducts of a big public research project. The crimes and sins of AI will become a matter for social and legal control β not problems that we can only solve if we can convince vast, corrupt, wealthy corporations to take action.
Could we have this AI future still? Well, after the AI bubble pops, weβre going to have _something_. Weβre going to have the open-source models that the hyperscalers released as demos for their main event βfrontier models.β Weβll have bankruptcy-sale hardware. Weβll have an army of skilled workers who will no longer be compelled by their bosses to try and cram AI into all of our orifices or to convince our bosses to fire us and replace us with chatbots that canβt actually do our jobs.
The question, though, is whether AI is more like a Robert Moses overpass, a nuclear reactor, or a tomato harvester? Are there _inescapable_ aspects of the current generation of AI models that mean that using them independently of the hyperscalers will still somehow perpetuate the harms of the current AI bubble?
One argument for this is that the fruit of the poisoned tree is forever poisonous. The fact that these AI tools were made by bad people for bad reasons makes everything you do with the tools bad. The most extreme version of this argument is that itβs the mere fact of a sinful origin that disqualifies the hyperscalersβ open AI models from moral consideration.
This is a bad argument. Thereβs hardly any technology β any built infrastructure β that is free from sin. Every microchip you use is filled with billions of linear descendants of William Shockleyβs silicon transistors. Shockley was a monster: a virulent racist and eugenicist who spent his Nobel Prize bounty in efforts to pay Black women to have their tubes tied as part of a disgusting racist project.
Going further back: Europeβs roads owe their existence to Romeβs imperial conquest, a bloody campaign that included mass enslavement and the indiscriminate slaughter of civilians. EuΒropeβs modern roads arenβt necessarily any betΒter: Hitler himself turned over the first spadeful of dirt for Austriaβs Autobahn.
Itβs important to remember these facts and to make reparations for old crimes β but just as you canβt both be alive and avoid impacting the environment, you also canβt be both alive and free from contact with the relics of monsters both modern and ancient. Anyone who claims that they live a life that isnβt steeped in blood both old and fresh is kidding themselves. By all means, do everything you can to right wrongs, but if you only permit yourself to act in ways that avoid contact with the product of sin, you wonβt do _anything_ , let alone anything to right wrongs.
Imagine a new generation of buses could handily fit under Robert Mosesβs Long Island overpasses, such that the overpasses could no longer be used to racially and economically segregate the cityβs beaches. At that point, the overpassesβ technopolitics change. They may have been made for an evil purpose, but now that they no longer serve that purpose, theyβre just bridges. The mere fact of their historical sin is no reason to tear them down today.
A less extreme version of the sinful origins arΒgument is that the ongoing use of a sin-steeped object is the source of fresh evil in the world. Museums that display the looted remains of indigenous people donβt merely dishonor those long-dead people: They inflict ongoing harm on living indigenous people whose elders are reduced to colonial curiosities, making indigΒeniety itself a relic, as opposed to a living force in the world that is critical to the lives and thriving of millions of people.
I think a version of this can be applied to some of the hyperscalersβ models. Models that have been trained on β and retain β the ill-gotten private information of millions of people are indeed so wicked that they should be deleted. Just as the mere existence of pornographic images of children is an ongoing harm to the people who were abused to create them, the mere existence of databases of ill-gotten private information is an ongoing source of harm and risk to the people in the database. If thereβs a chance that an AI model can be made to yield up your personal, sensitive information, we should destroy it.
But if there are databases β models β that were trained on private information, _in such a way that they no longer retain that information_ , then they are more like the Autobahn than they are like a trove of child sex abuse material. They are steeped in sin, they should not have been made in the way they were, but the ongoing use of them causes no new harm to anyone, and so if they are useful, we should use them β even as we erect protections to prevent anyone else from engaging in mass privacy invasions for any purpose, including AI, ever again.
Think here of the thousands of miles of fiber-optic cables that Worldcom buried under the streets of cities all over the world. Worldcom was a massive, wicked, fraudulent enterprise, whose executives lied to the public in order to steal billions from everyday working people, leaving them bankrupt or separated from their retirement savings and doomed to a miserable old age. For this crime, Worldcom CEO Bernie Ebbers was sentenced to prison, and he died there.
And yetβ¦the fiber is still under the ground. It got there by the most disgusting, harmful means imaginable. But treating those cables as morally untouchable doesnβt help any of those bankrupt pensioners. At my home in Los Angeles, I have a 2GB symmetrical fiber link from AT&T, who bought and lit up some of Worldcomβs fraudulent, pensioner-destroying fiber. I would rather live in a world where I had slower internet and the pensioners still had their money, but thatβs not an option. At this point, the fiber is just fiber β its use doesnβt harm anyone anew, nor would refusing to use it bring any benefit to anyone.
Some peopleβs anger over AI modelsβ training data is couched in claims about copyright, not privacy. The argument goes that by making transient copies of billions of public works and then tabulating the relationship of the elements in them, the AI companies are stealing from creative workers. I think this represents dangerΒously misdirected anger. Itβs totally legitimate to be angry at a tech executive with billions in investor capital who publicly announces their intention to destroy your livelihood β even moreso if they use your own creative output to accomplish this end.
But the bad conduct here lies in the intenΒtion and action to destroy your livelihood. It does _not_ lie in making copies of your work in order to count the words or the pixels in them. Copying and analyzing every public work is how every search engine works. Whatβs more, this practice wasnβt invented by search engines. Long before the first digital computers, the Oxford English Dictionary examined the entire corpus of publicly available written work to trace the etymology of our language. Long before Google News, the library where I worked subscribed to a quarterly, hardbound subject index of all the articles published in every general-interest English-language newspaper and magazine.
A language model is a collection of embedded inferences that certain words are likely to follow from other words. You may rightfully regard the use of a model trained on your work to attack your wages as unethical, but thatβs about the modelβs _operation_ , not its _creation_. The fact that the model has derived some of these inferences by counting all the words you ever published is not theft β no more than itβs theft for the OED to identify the first times I used the word βenshitΒtificationβ in its entry on my word. Counting words β even all the words ever uttered β isnβt theft. We call that βa search engine.β
If an open model, operated and hosted by someone other than the hyperscaler that shipped it, is sold or used to destroy your liveliΒhood, then the person using that model is bad, and we should have social, economic and legal sanctions for that person. But the issue here is the _use_ of the model. A neofascist who uses the Autobahn to travel to a racist rally is a Nazi β but theyβre a Nazi because of what theyβre doing, not because theyβre doing it on Hitlerβs highway.
The technopolitics of language models are bad because the people selling language models market them for the purpose of bankrupting writers and filling the world with slop. But thereβs nothing intrinsic to language models that decrees their use in this fashion. The computer scientist Pete Warden has a lovely demo of a language model running on a sub-$10 system on a chip (SOC) that can ingest and conversationally access the manual for an apΒpliance. In his demo, Warden uses this system to create an interactive dishwasher manual: tap a button on the dishwasher and ask it a question like βWhat should I do if Iβm getting streaks on my glassware?β and it will locate the correct answer in the manual and tell you what to try. This language model may have been trained on the corpus of all available literature without permission, and the company that did the training may have raised their capital on the promise of bankrupting writers with their own words, but the button on your dishwasher that uses an open, disconnected model to give you advice on getting your dishes clean in no way perpetuates those harms.
If tomorrow someone invented a nuclear reactor that sifted uranium out of seawater and ran safely in your garage, it wouldnβt have the same centralizing technopolitics as the nukes that Winner concerned himself with β and that would be true even if the factory that produced these decentralized pocket reactors ran on a ten billion dollar nuke filled with traditional enriched uranium. The technopolitics arenβt tied to the fact that the nuke produces energy by splitting atoms β itβs tied to the technical reality of the only way we know to accomplish that feat. If we learn new atom-splitting techniques, they can have different technopolitics.
By the same token, a language model β even one released by a despicable Altman or Amodei β has embedded technopolitics because of the kinds of infrastructure it requires, not because it guesses words. Guessing words is fine. Itβs someΒtimes useful. The open models that OpenAI and Anthropic have put into the world and can never claw back will be guessing words long after those companies are dust. We can use those open models for good or ill β we can choose their technopolitics. Technopolitics are a matter of what happens when you use a technology, and what you have to do to maintain and improve that technology. Optimizing open models is a matter for collaborative, open effort β not closed labs fueled by investorsβ centibillions.
AI _has_ technopolitics. But those politics are neither inevitable nor immutable. Sam AltΒmanβs lurking presence in a language model is no more politically disqualifying than William Shockleyβs residue in your microchips.
AI today might be a defective tomato harΒvester, immiserating workers and driving out good tomatoes with bad. But thereβs nothing intrinsic in the idea of a tool for harvesting a tomato that is bad for workers β or tomatoes. Itβs because we let bosses choose which tomato harvester was suitable for use that we ended up in this place. Put the pickers in charge of that decision and weβd have very different harvesting technology β and far more delicious tomatoes.
* * *
#### Cory Doctorow
CORY DOCTOROW (craphound.com) is a science fiction author, activist, and journalist. He is the author of dozens of books, most recently Enshittification: Why Everything Suddenly Got Worse And What To Do About It (nonfiction); and the novels Picks And Shovels and The Bezzle (followups to Red Team Blues). Other notable books include the solarpunk novels Walkaway and The Lost Cause; the tech policy books The Internet Con and Chokepoint Capitalism; the internationally bestselling YA Little Brother series; and the picture book Poesy The Monster Slayer. He maintains a daily blog at Pluralistic.net. He works for the Electronic Frontier Foundation, is an AD White Professor at Cornell University; an MIT Media Lab Research Affiliate; a Visiting Professor of Computer Science at Open University. He co-founded the UK Open Rights Group. In 2020, he was inducted into the Canadian Science Fiction and Fantasy Hall of Fame.
All opinions expressed by commentators are solely their own and do not reflect the opinions of Locus
**While you're here, please take a moment to support _Locus_ with a one-time or recurring donation. We rely on reader donations to keep the magazine and site going. We publish hundreds of reviews and news stories each year, but we can't do it without your help. **
Β©Locus Magazine. Copyrighted material may not be republished without permission of LSFF.