The public benefit corporation that did ‘nothing wrong’ is paying $1.5 billion for ‘nothing wrong’

Last year, Anthropic agreed to pay $1.5 billion after being caught with a central library containing more than seven million pirated books. I wrote about it at the time, although the agreement was still crawling through the usual legal swamp of claims, objections, fee disputes and judicial approval. Anthropic had agreed to pay, but the settlement had not yet reached the point where the court had finished stamping the paperwork and telling everyone to get on with it.

That has now happened. On 20 July 2026, the judge granted final approval, entered judgment and dismissed the case with prejudice. Anthropic must fund the entire $1.5 billion settlement, plus interest on the later instalments, while continuing to insist that it did ‘nothing wrong’.

Of course it did ‘nothing wrong’. Companies frequently pay $1.5 billion, destroy the disputed files and accept a permanent court judgment when absolutely nothing untoward has occurred. It is practically routine.

The settlement agreement contains the usual language saying Anthropic admits no liability or wrongdoing. Legally, that is standard boilerplate. Practically, it is hilarious. Anthropic is paying one of the largest copyright settlements ever agreed, but please do not interpret the enormous pile of money leaving its accounts as evidence that anything happened. The $1.5 billion is presumably just an unusually generous donation from the responsible AI company to hundreds of thousands of authors whose books mysteriously appeared in its pirate library.

The complicated legal debate and the extremely uncomplicated piracy:

There is a genuine legal argument around whether using copyrighted books to train an AI model can qualify as fair use. Anthropic won an important part of that argument when Judge William Alsup ruled that training its models on lawfully acquired books was sufficiently transformative to qualify.

That decision did not grant Anthropic a magical fair-use aura that made every step it took lawful. It did not convert torrent sites into bookshops, piracy into purchasing or stolen copies into responsibly sourced research material.

Before buying and scanning books, Anthropic downloaded more than seven million pirated copies from Library Genesis and Pirate Library Mirror. It paid nothing for them, built them into a central library and kept them even after deciding it might no longer use those copies for training. The court was quite clear that the fair-use finding for AI training did not excuse the separate creation and retention of that pirated library.

You can argue about transformation, model weights, market substitution and the finer boundaries of copyright law until the sun burns out. None of that changes where those files came from. Anthropic knowingly downloaded millions of books from two notorious pirate libraries because obtaining them properly would have cost money and taken considerably more effort.

That is piracy. There is no sophisticated AI exception that turns it into something else because the company doing it employs researchers, publishes safety papers and uses the phrase ‘long-term benefit of humanity’ with a straight face.

Anthropic was not a confused teenager clicking the wrong download button. This was a richly funded technology company systematically assembling a massive book collection for commercial research. Its case was about as blatant as corporate piracy gets: acquire millions of pirated copies, keep them in a central library, then argue over which later uses should be protected by fair use.

The settlement covers roughly 482,000 qualifying works rather than every pirated file Anthropic downloaded, because books had to meet the class requirements relating to copyright registration and ownership. Anthropic must also destroy the books it downloaded from LibGen and PiLiMi, along with copies derived from those files, subject to limited legal-preservation requirements.

Those would be the books connected to all the ‘nothing wrong’.

Paying for ‘nothing wrong’ in convenient instalments:

Anthropic does not have to throw the whole $1.5 billion into the furnace in a single afternoon. Even ‘nothing wrong’ can apparently be financed.

The first $300 million was paid on 2 October 2025. A second $300 million is due within five business days of final approval, followed by $450 million on 25 September 2026 and another $450 million on 27 September 2027. Interest accrues on the third and fourth payments from September 2025 until Anthropic hands them over.

The settlement fund is also non-reversionary, so money left over after claims, legal fees and administrative costs does not wander back into Anthropic’s accounts. Once the money enters the fund, it stays gone. There is no refund for being extremely innocent.

That means Anthropic will have paid $1.05 billion of the principal by the end of September 2026, with the remaining $450 million following next year. Class members may have to wait until any appeal period has expired or an appeal has been resolved before distributions begin, but Anthropic’s obligation to fund the settlement is no longer some hypothetical figure buried inside a proposed agreement.

The company has reached the expensive phase of ‘nothing wrong’, where the words remain free but the denial comes with a ten-figure payment schedule.

Humanity’s most responsible pirate library:

Anthropic is a public benefit corporation. According to Anthropic, this structure exists to balance commercial success with social and public good, while its Long-Term Benefit Trust helps keep the company aligned with the responsible development of advanced AI for humanity.

That branding is what makes the entire thing so fucking pretentious.

Plenty of technology companies behave badly. Anthropic behaves badly while presenting itself as the thoughtful adult supervising everyone else. It speaks in the solemn language of safety, alignment, governance, catastrophic risk and humanity’s long-term future, as though it is a committee of philosopher-monks reluctantly operating a nearly trillion-dollar technology company for our collective wellbeing.

Meanwhile, the company’s practical approach to writers was to treat their work as free raw material available from a torrent site.

Anthropic spends an extraordinary amount of time worrying publicly about theoretical future harms from advanced AI. Actual writers losing control over copies of their books apparently did not inspire the same level of caution. Those harms were immediate, measurable and happening to identifiable people, which may have made them less interesting than the glamorous hypothetical disasters that support Anthropic’s preferred image as humanity’s nervous guardian.

Its public-benefit structure did not prevent it from downloading the books. Its safety culture did not prevent it from keeping them. Its governance mechanisms did not make it pay the authors before taking their work. What finally produced alignment was a federal court and a $1.5 billion settlement.

Apparently, the most reliable AI safety mechanism remains an enormous legal bill.

Only $1.5 billion, because the AI bubble has destroyed everyone’s sense of money:

Anthropic raised $65 billion in May 2026 at a $965 billion post-money valuation. The company said that funding would support safety and interpretability research, expanded computing capacity, growing demand for Claude and the scaling of its products and partnerships.

Instead, an amount equivalent to roughly 2.3 per cent of that enormous funding round is being consumed by this settlement.

In the generative AI bubble, $1.5 billion can somehow be described as a manageable expense. That alone shows how completely the numbers have detached from ordinary reality. A billion dollars now sounds like loose change because every funding announcement comes wrapped in valuations that resemble the GDP of a medium-sized country.

The $965 billion valuation is imaginary private-market arithmetic. It represents what investors have agreed the company might theoretically be worth, not a vault containing $965 billion in cash. The $65 billion funding round, however, involved actual capital, and the settlement requires actual liquid assets to leave the company.

Investors gave Anthropic that money because they were promised the future. They were funding chips, data centres, engineers, research, model development, product expansion and the hope that all this colossal spending would eventually produce even more colossal returns.

The settlement produces none of that. It buys no GPUs, trains no models, hires no researchers and creates no product. It does not expand Anthropic’s capacity or generate future revenue. It closes a self-inflicted legal wound and leaves the company with $1.5 billion less cash than it would otherwise have had.

Yes, capital is fungible. Nobody can point to one specific investor cheque and declare that it has been placed directly into the copyright furnace. The financial effect remains perfectly clear: Anthropic’s available resources are being reduced by $1.5 billion plus interest because of decisions made by Anthropic. That money must eventually be replaced through revenue, borrowing, an IPO or yet another funding round.