$1.5 billion!?!
Somewhere in the Library of Congress, the break-in already happened. Investigators walked the stacks, ran their fingers along the spines. Every book accounted for. Nothing missing, they wrote. The thieves walked out with something the catalog didn’t track: the arguments the author had with herself, the conversations between the books, the meanings and life-worlds made manifest and corporeal and steal-able. Untraceable bearer-bonds for making AI models.
The price was set in September 2025 when Anthropic settled the Bartz v. Anthropic class action over training on pirated books. Roughly $3,000 per eligible work before fees and costs, across approximately 500,000 titles. The largest copyright settlement in U.S. history, for arguably the largest heist of intellectual property ever completed.
What does it mean that there is now a negotiated price for the corpus of English-language books?
We built towards this question in two earlier posts in this series. In Part 1, we argued that consumption is productive labor: buyers make prices true; platforms inherit the audience commodity and add measurement; wanting is constituted in circulation, not pre-given in private souls. In Part 2, we argued that fan communities collectively invent the genres capital later acquires. Royal Road, Archive of Our Own, the Twilight and Dramione fandoms, the science readers who shaped The Martian: each is a venue where unpaid work produces conventions, audiences, and writers in dialogue. Both posts converged on the claim that the labor of wanting isn’t individually owned, and that the politics of reclaiming it can’t be individual or proprietary.
The Bartz settlement is what happens when that question stops being theoretical.
I have a stake in this question. My blog, anotherpanacea.com, was included in the Google C4 dataset used to train several major language models, along with a substantial fraction of the philosophy blogs that existed in the late 2010s. Twenty years of writing (including things I’d probably disavow, things I no longer remember writing, and a lot of writing about other people’s writing about other people’s writing) scraped, tokenized, and now apparently part of how some chatbots think about Hannah Arendt.
AI training makes the metaphor literal
For decades, the autonomist claim had a measurement problem. The labor was everywhere, just diffuse. The extraction was happening, just hard to isolate. The tag-wrangler keeping AO3 navigable was doing unpaid labor that fed the machinery, but the path from her work to anyone’s profit was hard to trace.
AI training, however, changes the situation. In Part 2, we ended with a question borrowed from Federici’s strategic suspicion: what happens when capital learns to measure previously-illegible productive forces? This is the answer. The measurement is the enclosure.
Now there is a corpus available for every training run, to be incorporated into any model. There are weights. (A heavy piece of jargon, no? It’s calling out for some pseudo-Derridean metaphorical reading.) The general intellect is no longer only a metaphor for social knowledge embedded in production. It’s social knowledge embedded in a commodifiable artifact. Every story on AO3 has a small data-cloud around it. The whole cloud contributed to a corpus that gets scraped, tokenized, trained on, and sold back as capability.
In Part 2, we framed this as Dickens reborn as Patreon, paid by word count, with readers in the loop. The infrastructure becomes raw material for AI training.
The autonomists were right but couldn’t prove it in the form capital recognizes. Training infrastructure proves it in capital’s own language. The Bartz settlement, $3,000 per eligible work before fees and costs across approximately 500,000 titles, is both a fine for the heist and a price tag on a corpus. Rob the Louvre, but leave the art and just take the general intellect instead.
This is why AI scraping fights feel different from ordinary copyright fights. They are copyright fights, legally, but politically they’re fights over the enclosure of socialized productive force. The corpus wasn’t merely a pile of individual works. It was a community’s accumulated intelligence: ours collectively, and no individual contributor’s alone. (Imagine the anxiety of influence for AGI….)
Copyright can recognize the book because the book has edges. It has a title page, an author, a publisher, a theory of damages and market substitutes. The corpus does not have edges in the same way. It includes the review that taught a stranger how to read the book, the forum thread that stabilized the trope, the recommendation list that made the subgenre navigable, the archive tag that let desire find its object, the argument in the comments that taught the next writer what not to do. AI training doesn’t just ingest the works itself, it ingests the social machinery by which works become meaningful.
C. Thi Nguyen, in The Score, offered us some vocabulary for this thinning. Metrics get their power by being context-stripped enough to travel between strangers, and what travels is never what mattered most. Nguyen diagnoses an attendant pathology, value capture, which runs through the agent: the runner reshapes herself for the FitBit, the academic for the citation count. Training corpora invert the arrow. The writers on Royal Road, the convention-builders, the readers in dialogue weren’t optimizing to be ingested. They were doing the unflattened thing. The flattening happens to them, after the fact, by a metric they never consented to and a tokenizer I don’t think can carry the weight and pluralism of the gift-economy norms that made the prose worth reading in the first place. The community gave things weight, carrying meanings that mattered. The weights make (some of?) it countable, legible, transformable, financialized. A carry-forward.
Once trained, however, the model doesn’t need to keep paying rent to the archive it absorbed. That creates a temporal problem. This is partly dead-labor extraction. The training run reads the corpus; the model carries forward a transformed residue. Withdrawing future labor doesn’t recover what’s already been taken. The strike analogy fails if we already spent our leverage.
The dead-labor frame, however, is too clean. Models need updates, fine-tuning, preference data, reinforcement learning, evaluation, fresh corpora, and user feedback. Living labor keeps re-entering the picture. The politics depend on which layer we are talking about. Pre-training creates one kind of enclosure. Fine-tuning creates another. User feedback creates another. Platform dependence creates still another.
This helps explain why the live fights look the way they do: terms-of-service battles, scraping disputes, opt-out registries, artists using Glaze and Nightshade, authors suing AI companies, newspapers suing over training corpora, fan archives trying to preserve noncommercial gift economies against extraction. These aren’t marginal skirmishes around a technology whose politics lie elsewhere. We’re fighting over the boundary between gift economy and market. I’d call that politics.
That’s also why AI-generated relational goods are so unsettling. AI companions, personalized stories, synthetic tutors, fake-warmth assistants, infinitely responsive characters: each is trained on actual relational labor. The machine learns from care, fandom, flirtation, pedagogy, therapy, humor, and confession, then produces scalable pseudo-relations that compete with the human ones. The clearest case so far is the Scarlett Johansson incident. OpenAI asked for her voice; she declined, twice. The company launched ChatGPT 4o with a voice called Sky that sounded enough like her that Sam Altman tweeted “her” on the day of release, referring to the 2013 Spike Jonze film in which Johansson voices an AI companion the protagonist falls in love with. The company was openly building the product the film treated as a thought experiment and trying to attach the very performance the film was about. Johansson’s lawyers pushed back and OpenAI pulled the voice. The next case will be someone with worse lawyers. Maybe dead folks with greedy estates.
There’s a nice tight loop, too. HR software run by LLMs is screening resumes and cover letters written by LLMs, and the LLMs prefer their own outputs! A recent study found 67 to 82 percent self-preference bias across major models. In that study, candidates whose resumes were written by the screener’s own LLM were 23 to 60 percent more likely to be shortlisted than equally qualified applicants with human-written resumes. The human applicant is the medium, not the message. The medium has to use the right LLM. The Nguyen mechanism re-enters on the user side: agents reshape themselves toward a metric, including the metric of what an LLM-screened application looks like, written by an LLM the screener will recognize. The first round of extraction took unshaped corpora. The second round takes corpora already shaped by the first.
The hopeful possibility, which Alex Imas develops in his recent work on structural change, is that as commodity production gets cheaper, the relational sector absorbs more value and the human-marked grows more valuable. The darker possibility is that capital invents pseudo-varieties of the relational faster than human wanting can reallocate toward the authentic. The general intellect of fan labor becomes the input for synthetic relational goods that crowd the human version into a luxury niche. The Eliza effect has become an industry…
(That dark turn is its own essay.)
What reclamation could mean
If wanting is work, and if that work is appropriated, the obvious question is what we can reclaim and how.
Part of the answer turns on time. The extraction has already happened for many corpora. The trained models exist, and pulling future labor from them doesn’t recover what was already taken. What’s being defended now is partly the corpora not yet enclosed and partly the recovery from what was. Reclamation has to work in both directions.
Familiar answers fail in the same way. Each accepts the system’s units instead of contesting them.
What if we pay people for their individual data? Call this first answer “data dignity.” It has intuitive appeal. If platforms profit from our traces, why shouldn’t we get a cut? The framework above makes the premise unstable. Wanting isn’t individual in the relevant sense. The datum that matters isn’t mine alone. Its value comes from aggregation, relation, comparison, and pattern. Platforms don’t need my isolated click. They need my click in relation to millions of other clicks. They need the social form. (Not X, but Y and Z. It’s getting harder to resist epanorthosis!)
Data dignity tries to sell back to the producer something that they only own individually in a very limited and liberal sense. And worse: companies will always value the data less than the writers and creators did. Anthropic’s $1.5 billion Bartz settlement is roughly 0.17 percent of the $900 billion valuation the company is now negotiating. Sam Altman has floated “universal basic compute”: everyone gets “a slice of GPT-7’s compute” that they can use, resell, or donate to cancer research. No wonder Marx is making another comeback: expropriate the labor, then pay people back in the very product made from their labor, denominated in the company’s own scrip.
The second is a wage demand for fan labor. This tracks the productive-labor claim better, but it risks destroying the thing it wants to honor. Fan communities aren’t failed labor markets. They are gift economies with their own norms, obligations, pleasures, hierarchies, exclusions, and forms of recognition. The Organization for Transformative Works has been explicit that AO3’s noncommercial structure isn’t incidental. It’s part of what protects fanwork as fanwork. Federici helps here: the strategic point of demanding wages was to make the hidden labor visible so it could be contested and refused, not to properly price every intimate activity. Sometimes pricing is the enclosure.
The third is aesthetic refusal. Read slowly. Buy from independent stores. Cultivate tastes outside the recommender. Sound advice for a life well-lived, perhaps, but it goes nowhere politically, any more than buying local small-batch beers; private discipline can’t solve a collective infrastructure problem. You can refuse to train your own feed, but we’re social creatures and our preferences are social, too. Your doors are locked but your neighbor’s gate is open, so the heist continues.
A different orientation is also worth naming. Some writers, far from resisting capture, are courting it. The economist Tyler Cowen has argued that “you’re an idiot if you’re not writing for the AIs. They’re a big part of your audience, and their purchasing power will accumulate over time.” His most recent book is, by his own account, written largely for AI readers. Many academics write for posterity already. Many papers have few human readers. The AIs offer a strange new gift economy: trawl me, remember me, let me become part of how you think. Let me know a small bit of Ea Nasir’s digital immortality. Whether this is solution or surrender depends on what you wanted the writing to do.
Each, therefore, gets the unit wrong: the individual datum, the individual worker, the individual consumer. What we want is collective and infrastructural, and the Bartz settlement gestures toward how to reach it.
Bartz was a class action. Roughly half a million authors won a collective recovery from a company that trained on their work without permission. The mechanism worked. It also missed half the problem. Bartz paid authors per book, which honors the legal form of authorship while leaving uncompensated the reviewers who taught audiences to read in new genres, the fans who kept the wiki navigable, and the readers of all stripes who held the gift economies together while writers above them broke into commercial visibility. The corpus is a layered infrastructure that authorship sat on top of.
Bartz-writ-large is the version that uses the class-action mechanism to compel recovery and then routes the recovery through cy pres into the infrastructure that produced the corpus. Public libraries, public-service recommendation systems, cooperative archives, community-owned platforms, mutual-aid discovery networks, AO3-style non-market cultural infrastructure: these are what the recovery should fund, in acknowledgment of where the value the AI companies extracted actually came from.
When a class action settlement can’t be cleanly distributed to individual class members, because the class is too diffuse or the harm is too collective, courts approve distribution to organizations whose work serves the affected interests. That’s cy pres doctrine. Applied to AI training settlements, cy pres distribution to gift-economy infrastructure does what direct per-work payouts can’t: it recovers from extractors and rebuilds the conditions of further production.
This suggests two moves. First: infrastructure defense for what isn’t yet enclosed. AO3 is noncommercial, donation-funded, volunteer-built infrastructure for a gift economy. It’s a working example of exit from the dominant platform form. Didn’t Marcel Mauss teach us that gifts bind, obligate, rank, and discipline as well as sustain? Gift economies are more than just primitive markets waiting for prices. Defending them is institutional work, not sentimentality.
The second move is contesting training itself. Maybe that looks granular and procedural rather than visionary, but that is part of what might make it work.
The legal front is the most visible. The New York Times sued OpenAI in late 2023 for training on its archive. The Authors Guild brought a class action that overlaps with Bartz. Sarah Silverman, Richard Kadrey, and Ta-Nehisi Coates are litigating against OpenAI on similar grounds, and their case has forced OpenAI to disclose training data in court. Anthropic settled Bartz for $1.5 billion. The other cases will price other corpora.
Workers and artists defend themselves directly. SAG-AFTRA struck for and won AI consent provisions for actors. The Writers Guild won the same for screenwriters. Visual artists fight back through tools they built: Glaze defends a working artist’s style against mimicry, and Nightshade poisons training data that ignores opt-outs.
The infrastructure layer is shifting too. Cloudflare changed its defaults in July 2025: new domains block known AI crawlers, and a “Pay Per Crawl” feature lets sites charge AI companies for access. The EU AI Act requires training-data disclosure for general-purpose models. Fan archives debate at length whether their works should be exempted, and many writers move toward invite-only spaces.
If the corpus is collective labor, the right question is who has authority to move a gift economy into a training market. Who can convert an archive into input? Who can decide that a community’s accumulated form of life is now raw material for weights?
If we try to solve this by saying “pay everyone,” we’ll be stuck with the fact that “everyone” isn’t the right unit, either. The community is closer, but even “the community” is unstable: porous, internally unequal, often pseudonymous, often built around fanworks whose own legal status is contested. That instability is why the claim has to be infrastructural rather than proprietary. But infrastructure defense isn’t a complete answer to that asymmetry, either. Perhaps, though, it’s the name for a politics that can begin from it. (1)
Coda: the unrecognized half
We began this series with a folk picture: producers make, consumers use up.
It’s truer but stranger that production and consumption are both productive. Production makes objects, subjects, needs, and publics. Consumption realizes products, makes prices credible, supplies information, trains systems, stabilizes genres, and produces value through comparison. Wanting is not emptiness, lack, absence, or inadequacy. Wanting is social activity.
This doesn’t make all wanting good. It doesn’t make the customer sovereign. It doesn’t mean every appetite is liberatory or every consumer exploited in the same way. Just this: what we used to call the passive side is never passive.
The work of wanting is the unrecognized half of the economy.
AI training makes that recognition harder to avoid. It’s not enough to crawl the internet, they need to rank it, too. They need Reddit upvotes and Facebook likes. The general intellect can escape its old industrial sense of being embedded in machines. It is embedded in model weights, trained on what people wrote when they weren’t being paid to write. They were often working precisely because they didn’t experience what they were doing as work.
That is the trouble. Labor may be suffering (I’m with Arendt on this one), but work can be joyful. Meanwhile, voluntariness, gift, and consent are sometimes the medium through which extraction becomes durable.
The political fight now visible around AI training is an old fight in a new form. It’s the autonomist fight, finally legible to the machines. The tag wranglers, the patient-forum moderators, the Wikipedia editors, every uncredited maintainer of an internet that runs on attention: these people have been doing the work. They have built publics with taste. The genres capital later acquires were already legible because of them. The corpora now trains the machines made from what they wrote. The infrastructures they have built are the form of reclamation my framework predicts. I’m trying to describe a politics already in motion.
If you’ve ever tagged a story, written a review, or kept reading because you wanted to find out what happens next, then you, too, are in some of these corpora! I know I am. We are inside the politics whether we chose to be or not. What that politics requires from its observers is a separate question, though, and one I’m unable to settle here. Here’s what I think I know:
Sellers post prices. Buyers make them true.
Platforms offer choices. Users train the platform.
Writers make books. Readers make genres.
The world calls this consumption because calling it work changes who owes what to whom.
(1) Alex Imas argues that the relational sector will absorb labor as commodity production gets cheap; this is a producer-side claim about post-AGI employment. The argument I am making here is consumer-side and analytically distinct: the labor of wanting has been productive and has been appropriated, and the question of reclamation is about defending the gift economies that invent forms of life against enclosure. The two arguments are compatible but address different objects. Citing Imas’s empirical mimetic-desire findings doesn’t commit me to his sectoral-change conclusion.