The Politics of Not Touching It

06 Sep, 2026

Table of Contents

Preamble

I use these tools constantly, for real work, and I'm not going to pretend otherwise to make anyone comfortable. I also think the technology has permanently changed the material conditions of writing, research, and organizing, and that there's no version of the future where that reverses. Given that, the only serious question is how to use it well. Most of what passes for AI criticism right now isn't answering that question. It's avoiding it.

There are real, material critiques of these tools: who owns the compute, who owns the training data, whose labor got fed into the model without consent or payment, which company signs which contract with which government agency. Those critiques deserve to be engaged on their own terms, not waved off. What I'm arguing against is something else, the moralism that has grown up alongside those real critiques: the position that using the tool at all is a metaphysical stain regardless of who controls it, what it's used for, or how carefully. That position isn't a sharper version of the material critique. It's a retreat from it, dressed up to look like the opposite.

I. What I Am Not Arguing

Say the honest version first, because a piece that skips this becomes indistinguishable from uncritical boosterism, which isn't what this is either. OpenAI signed a deal on February 28, 2026 to deploy its models on the Pentagon's classified network, hours after Anthropic publicly declined to.¹ ICE's resume-screening tool runs on GPT-4.² These are facts, not vibes, and the backlash that followed, a boycott campaign that claims over 2.5 million supporters and drove a 295 percent single-day spike in ChatGPT uninstalls, was responding to something real.¹ Training data was taken from working writers and artists without consent or compensation, and the people who built the models did not ask permission. Data centers draw real water and real electricity, and the mineral extraction behind the chips is a real supply chain with real costs somewhere. I'm not going to litigate the exact numbers in this piece, that's a fight worth having on its own terms with its own evidence, but leaving it out entirely would make my own opening promise, that I'm naming what's real rather than dismissing all of it, a lie by omission. None of it is what I'm defending.

What I'm arguing is that the tactic of individual boycott and individual purity is a different thing from the grievance that motivates it, and that collapsing the two, treating "I quit ChatGPT" or "I would never touch this tool" as itself a political act rather than a consumer gesture, is where the argument goes wrong. The grievance is real. The response to it usually isn't structural at all.

II. The Machine Question

Marx worked through a version of this exact question in the "Fragment on Machines" section of the Grundrisse.³ His subject was industrial machinery, not language models, but the actual claim is sharper than "machines are neutral, ownership decides." His point is that fixed capital absorbs what he calls the general intellect, the accumulated, inherently social knowledge of the whole species, language, prior culture, the sum of what everyone before you figured out, and turns it into something a private owner can hold. That absorption creates what he calls a moving contradiction: capital needs to shrink necessary labor time toward zero by building ever more of that collective knowledge into the machine, while at the same time it still needs labor time to be the measure of value, the thing wages are paid against. Whether that contradiction resolves toward more free time for everyone or toward more extraction from the people still doing the work isn't written into the machine. It depends on who captures the surplus the machine creates.

A large language model is general intellect in almost too literal a sense. It's the totality of what's been written down, compressed into something a company owns and rents out. The Nota News case in Section VII is the moving contradiction made concrete: the same absorbed knowledge could have meant less work for the same pay, and instead meant the same pay for an impossible amount of work. That outcome was never going to be decided by the model. It was decided by who set the quota.

This is the argument the moralist position skips. Quinn, the water-guilt crowd, the comment-section accusers, all of them locate the moral content in the tool itself, theft, taint, disqualification, rather than in the ownership and labor relations that determine what the tool actually does to somebody's life. Marx's own answer to "is machinery good or bad" was neither. It was: that's the wrong axis. The right axis is who captures the surplus.

III. The Same Error at Every Scale

Susan Kaye Quinn, a solarpunk writer working from inside the same broad political tradition I write from, made the purity case plainly: you categorically "can't write stories of a better future using tools like genAI that are based on theft right now, today," full stop, regardless of what the finished work says or does.⁴ That's a real grievance (the theft) run through a totalizing conclusion (therefore any use is disqualifying) that doesn't actually follow from the premise. Training data theft is a fact about the company that built the model. It isn't a fact that migrates into every subsequent use of the tool, the same way theft doesn't migrate into every subsequent transaction made with stolen money once it's been recovered and redistributed. Quinn also makes a second, separate argument I'm not taking up here: that genAI can only produce pastiche, incapable of real novelty. That's an aesthetic and epistemic claim, not a moral one, and it deserves its own treatment rather than a rebuttal borrowed from the theft argument, so I'm naming it and setting it aside rather than pretending I've answered it.

At least Quinn argues something. A lot of what I actually run into doesn't bother. "AI is a moral issue" gets stated as a complete sentence, full stop, no premise, no conclusion, nothing to disagree with because nothing was claimed. That's not an argument I'm losing. There's no argument there to lose. What it's doing instead is signaling membership: it tells an audience that already agrees exactly what frequency the speaker is broadcasting on, and it treats anyone who asks "which part, specifically, and why" as thereby proving they're tuned to the wrong one. Quinn's version is wrong in a way you can actually engage, trace the premise, show where it overreaches. The bare assertion isn't wrong in that sense. It's not making a truth claim at all. It's making a loyalty claim, and loyalty claims don't need arguments, they need an ingroup that already recognizes the frequency.

QuitGPT is the more serious case, because its grievance is concrete and verifiable rather than metaphysical: a specific vendor signed a specific contract with a specific war-and-surveillance apparatus. But the tactic itself, uninstall the app, announce it, move to a competitor, is still consumer-choice politics. It doesn't touch who owns the compute, who profits from automated labor, or which government agencies get access to which capabilities. It relocates the same underlying dynamic to a different vendor, and the boycott's own reported outcome shows exactly how little it moved: users left OpenAI over its Pentagon deal and Anthropic's Claude usage more than doubled in the days that followed.¹ I'm writing this piece using Claude. I'm not hiding that to avoid a gotcha. I'm naming it because it's the actual point, and it cuts harder than a simple disclosure once you look at who I switched to. Anthropic isn't a clean alternative that stayed out of this. It already held roughly $200 million in Pentagon contracts before this boycott ever started, Claude was reportedly the first frontier model deployed on the Pentagon's own classified networks, and it's been reported in use in an active US military operation. What Anthropic actually declined was one specific clause, language explicitly ruling out autonomous weapons and domestic mass surveillance, not defense work as a category. Even the vendor treated as the principled alternative is this enmeshed in exactly the apparatus QuitGPT is protesting. That's not a defense of Anthropic. It's the point restated at full strength: there was no clean vendor to switch to, because vendor-switching was never going to touch the actual question. Boycotting your way to the "good" AI company doesn't exist as an option, not because you chose badly, but because the option itself was never on the table.

This isn't a new argument for me. I've made a version of it before about consumer boycotts generally: a boycott is fine as a personal choice, but it functions ideologically the moment it gets treated as a substitute for organizing, because it lets you feel clean without changing who owns anything. Pretending the tradeoff doesn't exist, or that opting out resolves it, is ideology doing maintenance work on a structural problem it has no actual power to fix. Purity-position AI moralism and vendor-switching AI boycotts aren't opposite cases, a bad one and a legitimate one. They're the same category of move at different scope, individualizing an ownership question that individual consumption can't answer.

IV. On Theft, From Someone Who Has Actually Thought About Theft

I've been a pirate my whole life. Music, movies, software, books, games, a personal archive built almost entirely from files I never paid for, and I don't carry guilt about it, because I worked out why a long time ago rather than deciding the guilt was inconvenient and dropping it. That history is relevant here, because "theft" is doing a lot of load-bearing work in this argument, and I've actually thought about what the word should mean, in a way most people invoking it against AI haven't had to.

The test copyright maximalists use, and the test Quinn uses against genAI, is the same test: was the copy made without authorization. If yes, it's theft, full stop, and the scale, the actor, and who actually gets hurt don't enter into it. That's not a materialist test. It's property-rights absolutism, and it treats a teenager downloading an album identically to a corporation scraping the internet, because both cases satisfy the same formal definition. The deeper problem with that test is that it ignores what digital goods actually are: reproducible at close to zero marginal cost. A movie costs fractions of a cent to copy. The forty-dollar price tag isn't the cost of production, it's rent, extracted to fund marketing budgets and shareholder dividends on top of a good that could be priced at what it actually costs to make and distribute. I couldn't have afforded my own library on any wage I've actually earned, and that isn't a confession, it's the empirical proof that the pricing has nothing to do with production cost. The test I actually use is who profits, who's harmed, and at what scale. A person copying a song for personal use redistributes culture horizontally. Nobody's labor gets displaced by it, the artist doesn't lose a sale to someone who was never going to pay retail anyway, and the party that "loses" is a label that already extracted the value from the artist's own labor and priced access above what most people can or will pay. That's piracy, and on my own test, it's defensible.

Here's the part that actually explains the ferocity of the people defending the old arrangement, and it's the same mechanism running through everything else in this piece. Someone who pays full price for a game console and a streaming subscription has fused their own identity to that spending. If piracy is easy and morally neutral, their spending was wasted, and rather than sit with that, they retroactively decide the spending was virtuous, "supporting the creators," "playing by the rules." A pirate's mere existence threatens that story, so the response isn't just personal compliance, it's active policing of anyone who didn't comply. They need someone else to be a criminal so their own submission reads as a choice instead of a loss. That's not really about piracy. It's the identical shape as the class tell two sections from now and the accusation culture after that: people who paid a cost defending the arrangement that charged it to them, needing someone else's noncompliance punished so their own compliance still means something.

FOSS is the same underlying refusal, formalized instead of quiet. Rather than taking what enclosure shouldn't have locked up in the first place, it refuses the enclosure at the point of creation: build the thing, give it away, license it so nobody downstream can re-enclose it. Linux, Wikipedia, a huge share of the actual infrastructure the internet runs on, exists because people decided the general intellect Marx describes in Section II should stay common property instead of becoming somebody's exclusive asset.

I'll be honest about where my own practice falls short of my own theory, because a piece this concerned with what separates real materialism from a comfortable story owes itself the same scrutiny. Piracy that only accumulates, that fills a hard drive and stops there, is expropriation without socialization: I took the thing out of the rentier's hands, but I didn't put it into anyone else's. A personal archive nobody else can access is still a private archive, even if nothing was paid for it. The actual next step, the one that finishes the move instead of freezing halfway through it, is distribution, seeding, sharing, building the infrastructure that makes the commons a commons instead of one person's collection. I don't always do that either. Worth naming, not smoothing over, because it's the same test I'm about to apply to AI.

Sci-Hub is the case that actually maps onto AI training data, more precisely than a mixtape does. Academic researchers write papers, often unpaid for the writing itself and sometimes charged their own page fees to get published, funded in the first place by public grants and peer-reviewed for free by other academics, and a publisher like Elsevier locks the finished work behind a paywall and sells it back to the same universities whose researchers produced it, at a markup those universities pay year after year. Sci-Hub doesn't take anything from the researcher. It reclaims labor the researcher already gave away, from a publisher charging rent on knowledge it didn't create. That's decommodifying something that should never have been commodified, and it's the closest existing case to what a genuinely careful complaint about AI training data should be aiming at.

Now apply the same test to AI training data, honestly, not defensively. A company scraping the writing of working authors at industrial scale, without payment, and then selling access to a commercial product built on that scraped labor, in direct competition with the same authors for the same readers, sits much closer to the Elsevier case than to the mixtape case. It isn't horizontal redistribution between people with no power over each other. It's an intermediary extracting labor and selling it back against the people who produced it, the exact structure Sci-Hub exists to break, just running in the opposite direction, enclosure instead of a raid on enclosure. And the fix isn't "pay the writers a licensing fee," the same way the fix for academic publishing was never "pay Elsevier more fairly." Creators already get crumbs from platforms; a licensing scheme just formalizes the crumbs at a slightly better rate. The actual correct line, the same one Sci-Hub and FOSS already point at, is socializing the distribution infrastructure itself, models and training data and compute held in common, not a better-negotiated price for admission to something that should never have been enclosed in the first place. This isn't me walking back Section I's credit to the material critique. It's me applying the exact same framework I've used my whole life, consistently, in both directions. Copying isn't the crime. Enclosure and extraction, at scale, for private profit, is, whether that's a record label suing a fan, a publisher renting out a paper it didn't write, or a trillion-dollar company monetizing labor it never paid for.

V. The Class Tell

There's a version of this critique that isn't really about theft or politics at all, and it's worth naming precisely because it hits differently. One commentator on AI shame culture noticed something worth sitting with, offered as a passing observation rather than a finding, that people who never felt shame about paying humans to write for them may be the ones most eager to shame AI users.⁵ It's not just an observation about hypocrisy. Hiring a ghostwriter, an editor, a research assistant, an agency, has never been treated as an authenticity violation. It's normal. It's sometimes prestigious. Money buying human labor to produce your words was never shameful.

What changed the moment a cheap or free tool started doing something similar is that the option became available to people who couldn't previously afford to solve the production problem by paying someone else. That's not a coincidence, and it's not just status anxiety among threatened cultural workers, though that's part of it too. It's a live class boundary being defended: who gets to produce "legitimate" work is being quietly redefined as whoever can afford a human proxy, with anyone using a machine proxy instead marked as a cheat. The moralizing isn't protecting craft. It's protecting a gate that used to run on capital and now has a cheaper door standing next to it.

The tell shows up in a more literal form too, among writers who saturate their own prose with exactly the markers an AI detector flags, the em dash chief among them, on purpose, before anyone accuses them of anything. The sequence runs: perform the tell, wait for the accusation, then flip it, claim the marker predates the machine, that it was theirs first, that the accuser has just insulted a human artist by mistaking them for the thing that stole from them. The critic meant to expose synthetic labor and instead gets recast as the aggressor policing an authentic voice they never actually threatened.

The claim underneath that gambit is the same enclosure move this piece keeps finding on the other side of the argument. Em dashes, aphoristic rhythm, the discursive cadence of "serious" prose are inherited conventions built up over generations of edited professional writing, by underpaid MFA-trained freelancers, journalists, and adjuncts, not any one writer's personal invention. A model absorbed that corpus by ingesting it wholesale, uncompensated, and a writer claiming personal title over "their" em dash is asserting the identical kind of property claim: individual enclosure of something that was actually produced socially. There was never a distinct human style for the machine to steal. There was only a statistically dominant register, and the model reproduces it because it was trained on that class fraction's collective output. Fighting over who gets to claim the em dash is easier than fighting over who owns the corpus and who captures the value of automating the labor that built it, and it's a lot more fun to play literary detective than to write something actually worth stealing.

VI. Shunning, or Cancel Culture Finds Its Object

I can't open a comment section anymore without tripping over an accusation. That's not an exaggeration for effect. It's the actual texture of being online in 2026, and it isn't random. It's cancel culture's mechanism finding a new, particularly well-suited object.

Cancel culture runs on a specific fuel: identify a surface marker of transgression, publicize it, extract social standing from having caught someone, and stop there. No organizing step follows the accusation, because the accusation is the politics. AI-use accusation fits that mechanism almost perfectly, because the "evidence" looks technical and objective (a detector score, a stylistic tell) even when it isn't, which makes the accusation feel like exposure rather than what it actually is: a moral performance dressed in forensic language.

The cases are concrete and they're bad. Jamir Nazir's short story "The Serpent in the Grove" won the Commonwealth Short Story Prize and was published in Granta. Social media flagged a single metaphor as an AI "tell," and the detector Pangram scored the whole story 100 percent AI-generated. The accusation stood for weeks before the Commonwealth Foundation reviewed his actual drafts and time-stamped notes and cleared him. He composes using text-to-speech.⁶ A researcher quoted on detector reliability put the stakes plainly: "even a false positive rate of 1% is high for an industry like publishing, where an accusation can send a career up in smoke."⁶ Mia Ballard's novel Shy Girl was pulled by Hachette over an internet rumor. She's a Black author, and she says only her hired editor used AI tools, not her.⁷ Jerry Falade's multimillion-dollar book and film deal was withdrawn over the same kind of uncertainty. He's also a Black author, and he argues, in his own words, that Black writers face disproportionate AI-use scrutiny compared to white ones.⁷ Granta ended its long-standing partnership with the Commonwealth Short Story Prize entirely over the Nazir controversy, a real institution damaged by an accusation that later turned out to be false. That's the pattern in miniature: the cost lands before the proof does, and sometimes it never gets reversed even after the proof arrives. Orion Newby, a student in a disability-support program at Adelphi, was expelled over a Turnitin false positive before a New York court reversed it.⁸ And when the Authors Guild ran the same pre-2022, verifiably human articles through multiple detectors, the results ranged from 0 percent to 100 percent AI depending on which tool was used.⁸

None of this is a story about a few bad detectors. It's a story about what happens when a real question, was this actually produced with integrity, has no reliable technical or legal answer, and social punishment rushes in to fill the vacuum left by the missing proof. That's not incidental to the moralism I'm arguing against. It's the same mechanism, at the scale of a single person's reputation instead of a company's stock price: individualized judgment substituting for a structural question nobody has actually solved, which is whether AI authorship can be verified at all.

VII. Yes, Most of It Is Slop

I'm not going to pretend the visible evidence doesn't exist, because I see it too, constantly, and most of what I see genuinely is slop. But the reason for that isn't a property of the technology. It's a property of who is economically incentivized to produce cheap, high-volume, unaccountable writing, and AI just collapsed the marginal cost of doing that at scale.

Nota News is the cleanest case I've found, because it isolates the variable. Eleven AI-generated "local news" sites launched in September 2025 to cover news deserts, staffed by two part-time editors. Poynter's investigation found more than 70 plagiarized stories drawing on the reporting of at least 53 journalists across 29 outlets, including one story that republished two-year-old news as current.⁹ One of the editors, Jorge Rodríguez, was paid $30 an hour with a quota of 10 to 15 stories per day, which he called "impossible to meet using traditional journalistic techniques." There were no editorial guidelines. When he asked about photo credit, he was told, "Don't worry about that now." Leadership, in his own account, "emphasized quantity over quality."⁹ The second editor at the same company, Dulce Ramos, did the job honestly, sourcing only public materials, and produced 10 to 15 stories a week instead of a day.⁹

Same tool. Same company. Same technology available to both of them. The only variable that actually explains the difference between honest work and plagiarized slop is the quota. AI didn't create the incentive to produce cheap, high-volume, unaccountable content. A pay-per-story economy that can't survive honest sourcing created that incentive, decades before any of this technology existed. The tool just made the volume achievable. Blaming the AI for what a $30-an-hour, 15-stories-a-day contract produces is like blaming a word processor for a content mill. The tool is downstream of the arrangement, not the cause of it.

There's a second, quieter version of this same error, which is the "you're too lazy to write your own grocery list" complaint. That one's actually onto something real. There is a documented effort-reduction effect: a preregistered study of 3,628 people found that merely labeling a slogan as AI-assisted, with nothing else changed, cut discretionary effort by roughly 13 percent relative to baseline.¹⁰ But even there, the honest reframe isn't "the tool made you lazy." It's that reduced effort on a task you've been told a machine already handled is a rational response to labor that was already devalued and unrewarded, the same logic that governs why nobody puts real effort into a task they know won't be checked or credited. That's a labor-conditions story, not a character flaw, and it's a story about the task's own stakes, not about the technology.

VIII. What Discipline Actually Requires

I don't think the difference between good and bad use is a matter of clever prompting, and I want to be honest that my own prompts, most of the time, are not sophisticated. What actually does the work is something less visible: a rigorously built environment, standing rules I've written down and hold myself to, that make verification and correction the default rather than something I have to remember to do in the moment.

Here's a concrete instance, from the actual writing of this piece. Partway through, I offered myself a plausible-sounding theory for why AI writing looks bad so often. Maybe good, well-integrated use is invisible by design, and what critics generalize from is a biased sample of nothing but failures. It's a clean explanation. It's also not true, at least not as the whole story, because I see the bad output directly and constantly, not just the flagged cases, and I said so. The theory got dropped, not defended, the moment it failed a direct check against what I actually observe. That's not a special skill. It's a habit, backed by a written standing rule that treats verification against direct observation as non-negotiable rather than optional, the same discipline that governs everything else I build with these tools.

This is the same argument I'd make about durable institutions generally. A council beats a charismatic leader, bylaws beat goodwill, and a written procedure beats trusting one person to stay accountable, not because any individual using the structure is smarter, but because the structure keeps working when the individual isn't at their best. The same principle applies to a single person's own working practice with a tool. Good AI use isn't a personality trait. It's whether you've built any structure around the practice at all, rules that catch you when you're tempted to accept a plausible answer instead of the true one. The moralist frame treats AI use as a character test, asking whether you're the kind of person who cheats. The actual variable is whether anything you've built holds you accountable regardless of what kind of person you are on a given day.

IX. What's Actually Left

How you use the tool and who owns it are two different questions, and this piece has been answering both on purpose rather than pretending one collapses into the other. Section VIII was about the first: your own conduct, disciplined by structure you build rather than character you're born with. This last part is about the second, and it's the one that actually determines whether the first even matters at scale.

Strip out the moralism, and strip out the quality complaint that's actually a complaint about incentive structures wearing a technology costume, and what's left standing is the real question the whole time: who owns the compute, who owns the models, who captures the value when labor gets automated, and which vendor signs which contract with which state apparatus. Those are organizing questions. They get answered by building power over ownership, not by refusing to touch a tool, not by switching which company gets your subscription fee, and not by scoring someone else's prose for a stylistic tell that a detector can't reliably distinguish from a human being who happens to write the way a machine writes this year.

I've already spent my whole life in two traditions that already know how to answer this. Piracy is the informal version, taking back what enclosure shouldn't have locked up. FOSS is the formal version, refusing to let it be enclosed in the first place. Neither one is a boycott and neither one asks anybody to feel clean by abstaining. Both are about building and using the alternative that keeps the commons a commons: open-weight models, training data that's actually public or actually contributed rather than scraped, compute owned cooperatively instead of by three companies. That's the same organizing answer piracy and free software have been giving for decades, aimed at a new target, and it's the same unfinished business I admitted to in Section IV: expropriation without socialization is only half the move. Using the tool well isn't the finish line either, if nothing you do ever feeds back into keeping the thing a commons instead of somebody's asset.

None of that is a reason to stop having the argument about training data, about labor, about which governments get which capabilities. Keep having it. Just have the actual argument, the one about who owns and who profits, instead of the substitute argument about who's allowed to feel clean.

Notes