When you look back at how this year initially was for Anthropic, you will remember the fights it picked with the Pentagon and the White House because they didn’t want their AI being used for mass surveillance and autonomous weapons. That nobility did not just win public favour back then but also gave them a moral high ground that made them the so-called “conscience” of the AI industry. Dario Amodei gave interviews about existential risk with the calm energy of a man who flosses daily and means it. Now in mere days, the same “safe AI” company has lost the good-will of the people and maybe a big chunk of their user base really soon.
Also read: I thought NotebookLM was just for research, then I discovered these 5 uses
Several court documents were revealed this month that showed how Anthropic has spent more than a truckload of money to purchase as many physical books as they could only to then splice their spines off, send the pages for scanning and recycling the leftovers. This was Project Panama and it had one goal, “destructively scan all the books in the world.” Somewhere between the moral high ground and the finish line, Anthropic apparently detoured through a paper shredder.
Let’s take a trip down memory lane and go back to where it all began in February, 2026. The US Department of Defence gave Anthropic a simple deal – the US Military gets full access to Claude for whatever they want to do including but not limited to mass domestic surveillance and fully autonomous weapons systems. The July before that Anthropic had signed a 200 million dollar contract with the Pentagon making it the first frontier AI approved for classified networks. Despite that history, they outright rejected the DoD’s offer.
Pete Hegseth, the Defense Secretary, did not like that rejection and almost immediately the Trump administration all but banned Anthropic’s products and called them a “supply chain risk,” which is a label that is usually reserved for companies that are suspected of being fronts for foreign adversaries.
Anthropic brought suit. The courts found a middle ground. The firm was disqualified from DoD contracts alone, allowed to retain its government business beyond that while the litigation proceeds. The firm’s CFO informed a federal judge that the entire affair could amount to billions in lost revenue in 2026.
And amidst all of this, the tech community largely gave a thumbs up. Anthropic was following through on the principle it always claimed it would follow through on – drawing its line despite the billions being lost as a consequence of drawing its line. It made the firm look good.
As the Pentagon debacle died down to a mere legal footnote, Anthropic rolled out their most flashy new models to date, Claude Fable 5 and Mythos 5, the first models in a range above the already existing Opus series. However, within three days, Commerce made Anthropic shut them down globally due to a jailbreak that allowed the unlocking of hazardous cyber capabilities.
Anthropic argued back strongly, stating that the jailbreak in question was too narrow to justify such a move, and that it could likely be done with rivals’ products, which do not come with any export restrictions at all. It took roughly three weeks of publicly visible arguments between the parties before Commerce silently re-enabled access.
Once again, Anthropic was shown as a company willing to face the government head on instead of backing down. Quite admirable, depending on whether one takes into account the standard “argues with the US government and survives” scale.
Then came the fall from grace because this one incident could have done more damage to Anthropic’s reputation than anything whatever goodwill they gained with the Pentagon fight. Anthropic had purchased a truckload of books but what was even more interesting was they wanted books written before 2022 only. Books prior to the Gen-AI boom and the ChatGPT era. The reasoning is a little too on the nose because anything published after late 2022 had more likely than not some level of involvement and cleanup by a chatbot before it ever went to print. It could be one chapter, a small edit or even just a punctuation fix but there was some AI DNA in there. You feed enough AI generated content to train another chatbot and it will, at some point, be equivalent to a snake eating its own tail. What philosophers might call an Ouroboros while researchers call it model collapse. An AI that trains itself on its own output will eventually become dumber and plainer almost like a photocopy of a photocopy. Books pre-2022 held value for that very reason. They were the last pieces of literature that could be certified as 100% human. Today, that is difficult to verify. You can’t even verify if this article was entirely human written because all you have to go off of is me saying I wrote it myself. Sure, there are AI content detectors but they are largely inaccurate.
Coming back to the books, Anthropic wasn’t hoarding them out of a love for print media and literature, they were strip mining the last vein of what I would call uncontaminated human writing before the whole quarry got infected with AI slop. It also needed these books destroyed so that there can’t be any argument of unlicensed physical inventory.
Project Panama was run more discreetly and seriously than the supply chains of entire companies. They hired Tom Turvey who used to run book acquisitions for Google Books to lead this project. There are also claims of brokers like ISBNdb helping source and ship these books, some under NDAs. The sellers were trained to frame it a little softly with words like digital preservation. In reality, the books were being physically disbound with hydraulic cutting machines just chopping off the spines of these books. Then the loose pages were run through high speed scanners and that paper was then recycled so I guess you could say there is some sort of an eco-conscious effort. An internal document did surface with the court documents that said the company didn’t want this to be public knowledge. It is not something companies say while doing something totally uncontroversial.
Also read: Claude Opus 5 vs Sonnet 5: 3 ways Opus is worth the upgrade
Legally, however, things start getting weird. Judge William Alsup decided that the “buy-scan-destroy” technique used by Anthropic was fair use, as it involved buying every single copy legally, which allows you to legally scan them and recycle. The reason Anthropic spent a penny, on the other hand, was the previous, unrelated crime. One of Anthropic co-founders, Ben Mann, had managed to download more than seven million pirate copies of those books from LibGen and other shadow libraries back in 2021, when Project Panama did not even exist yet. As a result, the lawsuit led to a 1.5 billion dollar settlement with the authors, announced just recently, which worked out to be about 3,000 dollars per book pirate copy. Project Panama, in its turn, was Anthropic’s way of getting clean data, since the shadow library strategy failed spectacularly.
This is not actually a tale of corporate wrongdoing in violation of the law. The court claims that they have not done anything illegal, at least not the scanning part. This is a tale of a corporation which for an entire year claimed itself to be the adult in the room when it comes to AI safety, preferred to forfeit a lucrative defense contract rather than violate their core values, while secretly conducting a covert operation whose own employees have been forbidden to talk about, dedicated to vacuuming up and then disposing of forever the irreplaceable copies of out of print and rare books just because the pre-ChatGPT prose has become scarce. Fair use is a legal protection. Fair use is not a PR tactic. And there is no quicker way to ruin the reputation of a safety company than having the public realize that your idea of the proper way of doing things ends up in a wood chipper.
Not to mention that they didn’t build a digital archive using these and a lot of these books could possibly have been the final or some special edition available in the market now lost to humans forever. So whatever the law maybe, Anthropic has lost their footing in the court of public opinion.
The first two stories were absolute wins for Anthropic’s reputation. Saying no to the Pentagon on autonomous weapons, even at a cost of billions in revenue, that is brave. Fighting Commerce over Mythos and getting the shutdown reversed in three weeks makes them look principled rather than reckless. By the middle of this year, Anthropic had banked more public goodwill than the entire notion of Generative AI, precisely because it kept proving it would rather lose money than lose its spine.
Then the books story spent all of it in one news cycle. It was not an unexpected reputational crisis resulting from some scandal that caught Anthropropic off guard. They cultivated the reputation, over the course of several months, by taking two expensive and genuinely commendable positions against one of the most powerful governments in the world. But then, it turned out that the very same company for two years had been quietly buying up all the literature in the world that had been written before ChatGPT and feeding it into the shredder, while explicitly instructing its employees not to discuss what it was doing. The fact that the company is right regarding autonomous weapons, and the fact that the company is right regarding Mythos does not change that at all. It only makes the contrast greater.
I like Claude, it has always been one of the better and more intuitive chatbots in my experience and it can be really useful once you understand how it works but Anthropic has made it extremely difficult to not root against them. GenAI since its inception has been made the villain and everytime you think the pros just might outweigh the cons, something like this happens and more people start losing trust in this new era of tech.
The company that spent an entire year arguing that it is the only lab that will do the difficult right thing has now made it easier than ever for its critics to argue against it: “fine you said no to the DoD, but did you also spend two years buying up every clean, pre-ChatGPT book you could find and feeding it into a machine so nobody could prove you still had them.” Anthropic earned its good guy reputation the hard way, one costly stand at a time. It just spent that reputation the easy way, in a warehouse with a hydraulic cutter and a gag order.
Also read: RBI’s draft AI model ignores agentic AI risk, experts warn