Look at four weeks in the summer of 2026 if you want the whole racket in one screenshot.
On 20 July a federal judge in California gave final approval to the Bartz v. Anthropic settlement: about $1.5 billion for a class of book rightsholders, after Judge Alsup had already held that Anthropic’s download of more than seven million pirated books from LibGen and PiLiMi into a permanent library was ordinary infringement, even while calling the later training “transformative.” The settlement requires the original torrented files destroyed. Anthropic’s lead early-and-often venture backer is Menlo Ventures. Menlo has called Anthropic the firm’s bet-the-firm AI position.
Nine days later, 29 July, Pangram Labs ships Pangram 4 and announces $9 million led by Menlo Ventures. The product is a referee. Schools, publishers, agents, recruiters, Substack. Spero tells TechCrunch the future is “discriminate in favor of human content.” The human now buys a scan to prove it.
Four days after that, 2 August, the EU AI Act’s Article 50 marking rules apply. Anthropic has already signed the Code of Practice. New Claude models begin embedding an invisible statistical watermark in every patch of text, worldwide, no opt-out. The detector is a private API for regulators, cops, media, schools, and enterprises. You cannot read your own brand. Anthropic says this “doesn’t change a user’s rights under our terms.” That is how you describe a product change you did not renegotiate.
Theft. Referee. Brand on the sentence. Same month. Same class of capital. The people who write for a living get the invoice at every window.
This is not a conspiracy. It is a cap table.
Nobody needed a Signal chat titled “double dip.” Venture firms recycle the same thesis: first, scale the model that swallowed the 20th century; then, once the internet is sludge and every classroom is paranoid, scale the tool that sorts the sludge. Menlo did not have to be evil. Menlo had to be early on Anthropic and later on Pangram. That is enough. The cleanup is a market because the mess was a market.
Call them crooks if you want. I will call them what the filings call them. Meta’s own engineer wrote, in a message the Kadrey plaintiffs unsealed, “Torrenting from a corporate laptop doesn’t feel right.” He asked whether they should use a VPN. Another internal note said not to use Facebook infrastructure, so nobody could trace the seeder back to Facebook servers. They still pulled, the authors say, 81.7 terabytes through Anna’s Archive, plus an earlier 80.6 from LibGen. That is the industry pattern: do it, do it off the office IP, then say “commercially licensed” in the model card.
Pangram’s model card says the human pool is owned, commercially licensed, or openly licensed, and not an “unauthorized internet crawl.” They report 7,000,000 book examples and do not name the catalog. Gutenberg, the eval set, has on the order of seventy thousand books, not seven million titles. Chunks can print that number. So can a shadow library. We do not have their torrent log. We have their silence, their method (book versus LLM rewrite of book), and a decade of peers who hid the download. That is the assumption. The next piece argues it in full.
What this does to a person who still writes
You already paid, as a culture, when the labs trained on the books. Some authors will get a settlement check. Most will not feel made whole. Then you pay Pangram, or your school pays Pangram, so a stranger with a budget will not throw the chapter out. Then you pay a humanizer, or an editor, or three more hours of your life, because clean 2026 English lives next to the rewrite bin. Then Claude, if you used it to cut a paragraph, invisibly brands the tokens and will not let you inspect the brand.
That is a money waste on both sides of the desk. The detector company sells fear to institutions. The writer sells labor to the detector. The lab that made the fear sells the watermark as “transparency.” The VC is in all three rooms.
The Atlantic already noticed the cultural part: basically every recent high-profile “this was AI” pile-on starts in Pangram. A novel pulled days before publication. Prize stories. A papal encyclical. WIRED asked whether you should trust the gold standard. The company in its own 2024 paper said do not use the classifier as the sole arbiter of academic integrity. Two years later the score is a verdict anyway, because a yellow highlight is cheaper than reading.
The rest of the series is the brief. Part 2: why “they did it on the side” is the reasonable prior, and what would falsify it. Part 3: the writer tax, dollar by dollar. Part 4: Claude’s watermark, which they did not put in the contract you could refuse.
- Part 2 — They wouldn’t use the office IP
- Part 3 — First you pay to prove you’re human. Then you pay to look human.
- Part 4 — Claude branded your sentences. It wasn’t in the deal.
- Earlier: the double collection (HN essay)
Sources: Authors Alliance on Bartz final approval; Alsup June 2025 order via case writeup; TechCrunch, 29 Jul 2026; TechCrunch on Menlo / Anthropic; Ars Technica / Kadrey emails; arXiv 2402.14873 Table 5; Pangram 4 technical; Anthropic watermark explainer; The Atlantic, May 2026.