AI Should Be Free. Anthropic’s 5% Offer Is an Insulting Distraction From a Parasitic Business.

The frontier labs fed on everything we wrote, and now they rent it back to us. A $1.5 billion settlement against a $965 billion valuation is not compensation, and neither is a 5% fund the company controls. Free, ad-free access to general-purpose AI is the one payment that reaches everyone who built it.

In my last few pieces I made two arguments. The first was that large language models were assembled from the written output of billions of people who were never asked and never paid. The second was that those same people are still working for AI companies every day, as unpaid data providers whose prompts, corrections and conversations train the next model.

This piece is about what follows from those two facts.

My position: the general-purpose capability of AI, the part distilled from the human corpus, should be free at the point of use, without advertising, for everyone. Companies can and should charge for compute-heavy extras, enterprise integrations and specialized products built on top. They should not charge the public rent on a resource the public produced.

The bill nobody paid

The most careful attempt to price the corpus comes from Nikhil Kandpal and Colin Raffel, whose position paper appeared at ICML 2025. They studied 64 language models released from 2016 to 2024 and estimated what it would cost to pay people to write each training set from scratch. Even at conservative wage rates, the data would cost 10 to 1,000 times more than the compute used to train the models. [1] Their title states the conclusion plainly: the most expensive part of an LLM should be its training data. In practice it was the cheapest, because almost none of it was paid for.

Where did it come from? A large share started in Common Crawl, a nonprofit archive of more than 9.5 petabytes of web pages collected since 2008, which Mozilla’s researchers describe as one of the most important sources of generative AI training data since GPT-3. [2] Books came from shadow libraries. The record in Kadrey v. Meta shows Meta downloaded Library Genesis and Anna’s Archive. [3] The Bartz plaintiffs alleged Anthropic collected more than 7 million pirated books. [4]

When a platform has leverage, the corpus suddenly has a price. Reddit’s licensing deal with Google was reported at about $60 million a year. [5] Not one dollar of that reaches the people who wrote the posts. The market for human data is real. It just pays the landlord, not the tenant who built the house.

The model is the commodity. The data is the asset.

Architectures are published. Compute can be rented by anyone with capital. Model quality is converging. Epoch AI found that since January 2026, open-weight models have trailed the best closed models by an average of about four months. [34] The price of a given level of capability falls by an order of magnitude or more each year. [23] When a product can be matched within a season and its price collapses every year, it is a commodity.

What does not commoditize is the input. Kandpal and Raffel put it in dollars: replacing the training data would cost 10 to 1,000 times the compute. [1] The labs’ own spending says the same thing. In June 2025 Meta took a reported 49% stake in Scale AI, a company whose business is producing and labeling training data, at a $29 billion valuation. [35] Labs pay contractors to label data, pay platforms like Reddit to license it, and pay settlement funds when they are caught pirating it. The one group never paid as a matter of course is the people who wrote it.

So rank the value chain honestly. Chips and power are bought at market prices. Engineers are well paid. Model weights are converging toward a commodity. The scarcest and most valuable input, the human record of how to reason, explain, argue and write, was taken without consent or payment. There are other moats, such as distribution, enterprise contracts and the capacity to serve a billion users, but none of them would exist without a model that first learned from us. The most valuable part of the chain is the part nobody paid for.

What the courts have given back so far

Two years of litigation have produced one large payout and a lot of uncertainty.

In Bartz v. Anthropic, Judge Alsup ruled in 2025 that training on lawfully acquired books was fair use, but keeping a central library of pirated copies was not. [6] Anthropic settled for $1.5 billion without admitting wrongdoing, and the court granted final approval on July 20, 2026. [7] By the March deadline, claims covered 440,490 of 482,460 eligible works. [8] That works out to roughly $3,000 per book before fees, split by default 50/50 between author and publisher, paid in installments running through September 2027. [9]

Now the ratio. Anthropic’s Series H in May 2026 valued the company at $965 billion. [10] The settlement equals about 0.16% of that figure, and the single funding round eight months after preliminary approval raised $65 billion, roughly 43 times the settlement. The largest copyright recovery in history is a rounding error on the balance sheet it was meant to correct.

And Bartz covers one company, one category of work and one method of acquisition. Everyone whose words came from the open web is outside it. Under the reasoning in Bartz and in Kadrey v. Meta, where Judge Chhabria granted Meta summary judgment because the thirteen plaintiffs failed to prove market harm [3], training on publicly available text is likely to be treated as fair use in the Ninth Circuit.

Elsewhere the picture is mixed. On September 29, 2026 the Third Circuit became the first federal appeals court to rule on AI training, holding that ROSS Intelligence’s use of Westlaw headnotes was not fair use, while expressly declining to decide the generative AI question. [11] The Department of Justice told that court that training large language models is extraordinarily transformative. [12] In Germany, a Munich court ruled in November 2025 that OpenAI’s models had memorized song lyrics and that memorization is reproduction; OpenAI is appealing. [13] In the UK, Getty Images lost its main copyright claims against Stability AI. [14] The New York Times case against OpenAI remains in pretrial fighting, including a July 2026 sanctions motion over ChatGPT logs. [15]

So the law is converging on a split answer: piracy is expensive, scraping the open web probably is not. That settles a copyright question. It does not settle the economic one. Something can be legal and still leave a debt.

Five percent of someone else’s value chain

On October 18, 2026, Anthropic announced it would commit 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval, with another 2% at the board’s discretion for every 50% rise in the company’s value above a baseline it has not yet published. [32] Counting its co-founders’ pledge and an employee equity match, Anthropic says roughly 20% of its shares will be dedicated to the public good. [32]

That is a serious number. At the May valuation, 5% is about $48 billion, roughly 32 times the Bartz settlement. Analysts cited by Al Jazeera expect an IPO valuation above $2 trillion, which would put the same 5% near $100 billion. [33] It also arrives weeks before a listing, when the usual pressure runs the other way. On its face it looks generous. Measured against what was taken, it is a distraction.

But look at who it is for. The fund may make large donations to nonprofits, direct distributions to specific populations, or work on labor disruption caused by AI. It is an LLC owned by the company rather than a 501(c)(3), and the board decides who receives what. [32] The announcement does not mention authors, training data or the people whose writing the models learned from.

That is the shape of philanthropy, not settlement. The input was taken without consent, in part from pirated libraries. The value was booked as equity. A slice now comes back on terms the company sets, to beneficiaries the company chooses. A value chain whose first and most valuable link was taken does not become fair because its last link donates. The people who supplied the raw material are not creditors in this arrangement. At best, they are eligible for charity.

The fix is small and specific. Name data contributors as a beneficiary class. Dedicate a defined share of the fund to free, ad-free public access and to the commons the models drew on. Give contributors a seat in its governance, and publish the valuation baseline. Do that, and Anthropic’s fund becomes the first working version of the Human Data Trust I describe below.

Why cash compensation will never reach most of us

The instinctive remedy is to pay contributors. I spent a whole article pricing that debt, and I still think the number matters. But individual cash payments fail on mechanics.

Attribution at the level of a single post is not feasible across trillions of tokens. Most contributors, the forum posters, Wikipedia editors, Stack Overflow answerers, product reviewers and bloggers, hold no registered copyright and belong to no class. The one functioning payout shows the friction even in the easy case: two years of litigation, a claims process, a publisher split, and about $3,000 per book for a narrow set of registered works. Licensing deals, as Reddit shows, pay platforms rather than people.

A dividend paid in kind solves the distribution problem. Everyone contributed to the corpus, so everyone receives access to what it built. There is precedent for treating a shared resource this way. Alaska’s Permanent Fund pays every eligible resident a share of oil wealth, and in 2026 the dividend was $1,200 per person, more than $800 million in total. [16] The oil was in the ground before any company drilled it. The human corpus was on the web before any lab scraped it. It is our oil field.

The free tier we have is not the free tier we’re owed

A reasonable reply is that AI is already free. ChatGPT reached roughly 900 million weekly users in February 2026, and only about 5% of active users pay. [17] Claude, Gemini and others offer free tiers too.

Look at what that free tier has become. On January 16, 2026, OpenAI announced advertising for logged-in US adults on its Free and $8 Go tiers, with paid tiers kept ad-free. [18] The first paid ads shipped on February 9. [19] Above that, the consumer price ladder runs from $20 to $200 a month across the major labs. [20] OpenAI says ads do not influence answers and that conversations are not shared with advertisers, and I take that at face value. The structure is still the problem.

Count the extractions. The public wrote the corpus for free. The public now uses the product built on it while being shown ads. And, as I argued in the previous piece, the public’s conversations become training data for the next version. Three layers of value flow one way.

There is a fourth cost. The commons that fed AI is being drained by it. The Wikimedia Foundation reported that human pageviews fell about 8% from March to August 2025 compared with a year earlier, and attributed the decline in part to AI answers that summarize Wikipedia without sending readers to it. [21] Wikimedia’s CEO has since asked AI companies that rely on its content to pay. [22] A fair settlement has to fund the sources, not only the users.

Free is affordable, and the public is already paying

The usual objection is cost. The numbers no longer support it.

Epoch AI found that the price of reaching a fixed level of model performance has been falling between 9 and 900 times per year, with a median of 50 times. [23] A usage-weighted market index of token prices hit $0.97 per million tokens on September 1, 2026. [24]

Here is a simple model of the cost of a generous universal tier:

Annual cost = N × t × 365 × p

where N is users, t is tokens per user per day, and p is price per token. With N = 1 billion users, t = 20,000 tokens a day (roughly 15,000 words, far above typical use) and p = $0.97 per million tokens, the result is about $7.1 billion a year. Multiply by five to allow for stronger models and peak load and you get about $35 billion.

Compare that with revenue. OpenAI is reported at about $50 billion in annualized revenue [25], and Anthropic reported a $47 billion run rate in May. [10] Together that is roughly $97 billion. A 10% levy on those two companies alone would raise about $9.7 billion, more than the base case.

This is a proposed model, not a cost accounting. A blended market price is not the same as marginal cost, frontier reasoning tokens cost far more than the average, and real usage is skewed toward heavy users. Treat it as an order of magnitude. The order of magnitude is the point: universal access is a single-digit percentage of industry revenue, not an existential burden.

The public is also already funding the physical buildout. Good Jobs First estimates that data center tax exemptions now cost Georgia about $2.5 billion a year, and Ohio paused new exemptions after the 2025 cost came in above $1.5 billion, roughly 11 times its forecast. [26] Thirty-eight states offer dedicated data center incentives. [27] Taxpayers are subsidizing the servers. They should not then be billed for the service.

What the other side gets right

The strongest counterarguments deserve a direct answer.

Training is legal. Under current US rulings, mostly yes. But this is a policy argument, not an infringement claim. Legislatures set the terms on which public resources are used all the time, from spectrum auctions to mineral royalties, without first proving anyone broke the law.

Free kills the incentive to build. The proposal covers a baseline, not the frontier. Labs keep charging for top models, agents, high-volume use and enterprise products. What they give up is the ability to sell the public’s own knowledge back to it at the entry level.

Free means ad-funded or degraded. That is exactly why the ask is free and ad-free, with a capability floor, funded by levy rather than by attention.

Open-weight models are already free. They are, and they matter. Switzerland’s Apertus, built by EPFL, ETH Zurich and CSCS with open weights and open training data, is the clearest public example. [28] But running a 70-billion-parameter model takes hardware and skill most people lack. Available is not the same as accessible.

Models learn the way people read. Judge Alsup leaned on this analogy. Judge Chhabria called it inapt, pointing to the risk that AI output floods the market for the works it learned from. [29] A person who reads a thousand books does not then sell a billion conversations a week.

A proposal: the Commons Access Standard

This extends the Human Data Trust I proposed earlier this year into something narrower and easier to legislate.

1. Free baseline tier. Any developer whose general-purpose model was trained on uncompensated public data provides a free, ad-free tier with meaningful daily limits, at a capability level no more than one generation behind its flagship.
2. Training-data levy. A small percentage of AI revenue flows into a Human Data Trust that funds public inference capacity, commons institutions such as Wikipedia and public libraries, and public models like Apertus.
3. Subsidies with strings. Data center tax incentives are conditioned on contributing compute to public access.
4. Disclosure. Developers publish summaries of what their models were trained on, so contributors can see what was taken. The EU’s enforcement powers over general-purpose AI providers became active on August 2, 2026, which gives this a working template. [12]
5. Opt-in conversation training. User conversations train future models only with explicit opt-in, and opt-in users get expanded free access in return.

The window is now

OpenAI was founded as a nonprofit to make sure advanced AI benefits all of humanity. After its 2025 recapitalization, the OpenAI Foundation holds 26% of the for-profit company. [30] Anthropic is expected to go public in mid-November, and OpenAI is reportedly raising at a valuation above $1.4 trillion. [31] Once these companies are public, every free unit of service becomes a quarterly line item that shareholders will want removed. Terms are easiest to set before that expectation hardens. Anthropic’s pledge shows the labs already accept that part of this value belongs to the public. The open question is whether the public receives it as a right or as a donation.

We wrote the training data. We edited it, argued over it, answered each other’s questions in it and posted it for free because we thought we were building a shared resource. We were right about that. The companies built on it should be free to charge for what they add. They should not be free to charge us for what we gave.

References

1. Kandpal, N. & Raffel, C. (2025). Position: The Most Expensive Part of an LLM should be its Training Data. ICML 2025, PMLR 267.
2. Baack, S. / Mozilla Foundation (2024). Training Data for the Price of a Sandwich: Common Crawl’s Impact on Generative AI.
3. Loeb & Loeb (2025). Kadrey v. Meta Platforms, Inc.
4. OPI (2025). Bartz v. Anthropic: $1.5 Billion Settlement for Use of Pirated Books in AI Training.
5. Engadget (2024). Reddit is licensing its content to Google to help train its AI models.
6. Ohio State University Libraries (2026). Fair Use and Artificial Intelligence 2026 Update.
7. Authors Guild (2026). Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement.
8. Manuscript Report (2026). AI Copyright Lawsuits Tracker.
9. Textbook & Academic Authors Association (2026). Bartz v Anthropic: When to Expect Payments, Next Steps; Authors Alliance, Bartz v. Anthropic Settlement FAQ.
10. CNBC (2026, May 28). Anthropic tops OpenAI as most valuable AI startup.
11. Ballard Spahr (2026). Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved.
12. Tech Times (2026, Oct 1). AI Training Fair Use Loses First Appellate Test.
13. Bird & Bird (2025). Landmark ruling of the Munich Regional Court (GEMA v OpenAI).
14. Cleary Gottlieb (2025). UK High Court Issues Landmark Ruling in Getty Images v. Stability AI.
15. CADE Project (2026). New York Times-led group seeks sanctions against OpenAI.
16. Alaska Beacon / States Newsroom (2026, Sep 30). Permanent Fund dividends will be distributed to Alaskans starting this week.
17. WiserNotify (2026). ChatGPT Users Statistics.
18. AdNews (2026, Jan 19). OpenAI launches advertising tier for ChatGPT in U.S.
19. cloro (2026). Does ChatGPT Have Ads? 2026 Rollout Status.
20. SentiSight (2026). 2026 AI Subscription Price Comparison.
21. Computing (2025). Wikipedia reports traffic drop, blames AI search and social media.
22. Pillitteri, P. (2026). Wikipedia loses 8% of visits and its CEO tells AI firms to pay.
23. Epoch AI (2025). LLM inference prices have fallen rapidly but unequally across tasks.
24. KuCoin Research (2026). Why AI Token Costs Are Falling.
25. Tech Startups (2026, Oct 19). OpenAI seeks $1.4 trillion valuation in $30 billion funding push.
26. The Register (2026, Jun 1). Ohio hits pause on datacenter tax breaks; Governing, The Very Real Costs of Tax Incentives.
27. Pennsylvania Capital-Star (2026, Apr 15). Many states don’t report losses from data center tax breaks.
28. GGBA Swiss (2025). Switzerland launches Apertus, a fully open multilingual large language model.
29. Pearl Cohen (2025). Two U.S. Federal Court Decisions Find Specific Uses of Books to Train AI Are Fair.
30. OpenAI. Our structure.
31. Tech Funding News (2026, Oct 20). OpenAI races to lock in $30B at a $1.4T valuation before Anthropic’s IPO.
32. Anthropic (2026, Oct 18). Investing in the public benefit.
33. Al Jazeera (2026, Oct 19). AI giant Anthropic commits 5% of shares to ‘public benefit’ fund.
34. Epoch AI (2026, May 29). Open-weight models trail the closed frontier by about four months.
35. TechCrunch (2025, Jun 13). Scale AI confirms significant investment from Meta.

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.

References

1. Kandpal, N. & Raffel, C. (2025). Position: The Most Expensive Part of an LLM should be its Training Data. ICML 2025, PMLR 267.
2. Baack, S. / Mozilla Foundation (2024). Training Data for the Price of a Sandwich: Common Crawl’s Impact on Generative AI.
3. Loeb & Loeb (2025). Kadrey v. Meta Platforms, Inc.
4. OPI (2025). Bartz v. Anthropic: $1.5 Billion Settlement for Use of Pirated Books in AI Training.
5. Engadget (2024). Reddit is licensing its content to Google to help train its AI models.
6. Ohio State University Libraries (2026). Fair Use and Artificial Intelligence 2026 Update.
7. Authors Guild (2026). Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement.
8. Manuscript Report (2026). AI Copyright Lawsuits Tracker.
9. Textbook & Academic Authors Association (2026). Bartz v Anthropic: When to Expect Payments, Next Steps; Authors Alliance, Bartz v. Anthropic Settlement FAQ.
10. CNBC (2026, May 28). Anthropic tops OpenAI as most valuable AI startup.
11. Ballard Spahr (2026). Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved.
12. Tech Times (2026, Oct 1). AI Training Fair Use Loses First Appellate Test.
13. Bird & Bird (2025). Landmark ruling of the Munich Regional Court (GEMA v OpenAI).
14. Cleary Gottlieb (2025). UK High Court Issues Landmark Ruling in Getty Images v. Stability AI.
15. CADE Project (2026). New York Times-led group seeks sanctions against OpenAI.
16. Alaska Beacon / States Newsroom (2026, Sep 30). Permanent Fund dividends will be distributed to Alaskans starting this week.
17. WiserNotify (2026). ChatGPT Users Statistics.
18. AdNews (2026, Jan 19). OpenAI launches advertising tier for ChatGPT in U.S.
19. cloro (2026). Does ChatGPT Have Ads? 2026 Rollout Status.
20. SentiSight (2026). 2026 AI Subscription Price Comparison.
21. Computing (2025). Wikipedia reports traffic drop, blames AI search and social media.
22. Pillitteri, P. (2026). Wikipedia loses 8% of visits and its CEO tells AI firms to pay.
23. Epoch AI (2025). LLM inference prices have fallen rapidly but unequally across tasks.
24. KuCoin Research (2026). Why AI Token Costs Are Falling.
25. Tech Startups (2026, Oct 19). OpenAI seeks $1.4 trillion valuation in $30 billion funding push.
26. The Register (2026, Jun 1). Ohio hits pause on datacenter tax breaks; Governing, The Very Real Costs of Tax Incentives.
27. Pennsylvania Capital-Star (2026, Apr 15). Many states don’t report losses from data center tax breaks.
28. GGBA Swiss (2025). Switzerland launches Apertus, a fully open multilingual large language model.
29. Pearl Cohen (2025). Two U.S. Federal Court Decisions Find Specific Uses of Books to Train AI Are Fair.
30. OpenAI. Our structure.
31. Tech Funding News (2026, Oct 20). OpenAI races to lock in $30B at a $1.4T valuation before Anthropic’s IPO.
32. Anthropic (2026, Oct 18). Investing in the public benefit.
33. Al Jazeera (2026, Oct 19). AI giant Anthropic commits 5% of shares to ‘public benefit’ fund.
34. Epoch AI (2026, May 29). Open-weight models trail the closed frontier by about four months.
35. TechCrunch (2025, Jun 13). Scale AI confirms significant investment from Meta.