The Accusations and the Reality
The AI industry has a massive data problem. Nobody wants to pay for it.
Recently, Anthropic published a blistering accusation against three Chinese AI companies. They claimed DeepSeek, Moonshot, and MiniMax used a sophisticated proxy network of 24,000 fake accounts.
This network allegedly scraped over 16 million conversations from Claude. Anthropic framed this as a massive security threat.
They warned that “distilling” their model could strip away safety guardrails and lead to unpredictable risks. It sounded like a classic story of corporate espionage.
Then, court documents unsealed in a separate copyright lawsuit painted a very different picture.
- Anthropic accused three competitors of utilizing 24,000 fake accounts to scrape 16 million Claude interactions.
- Leaked documents expose Anthropic’s “Project Panama”, a covert operation systematically destroying books for AI training.
- AI experts argue API scraping is merely imitation, validating that true reasoning requires independent reinforcement learning.
The Pot Calls the Kettle Black
Anthropic is currently pointing fingers at others for unauthorized data scraping. Meanwhile, their internal communications reveal a highly organized effort to ingest copyrighted material on an industrial scale.
They called it “Project Panama”. The internal mandate was clear. They needed high-quality, long-form text to teach their AI how to structure arguments and write properly.
The internet is filled with garbage. Books are heavily edited, proofread, and logically sound.
Inside the Book Destruction Factory
Instead of licensing these books, Anthropic took a brutal approach. They purchased tens of thousands of used books from retailers like Better World Books and World of Books.
Then, they systematically destroyed them in high-speed scanners to feed their models. The process operated like a factory line. They even hired a former Google Books executive to run it.
One internal planning document stated explicitly that they did not want the public to know about this destructive scanning plan.
The secrecy makes sense. When your co-founder is caught enthusiastically sharing links to pirate libraries like LibGen with employees, public relations becomes difficult.
The Distillation Illusion
Let us put the ethical gymnastics aside. Does scraping Claude actually build a better AI?
According to Nathan Lambert, a prominent AI researcher specializing in Reinforcement Learning from Human Feedback (RLHF), the answer is complicated.
Lambert points out that Anthropic lumped three very different operations together to build their narrative.
- DeepSeek: Accounted for only about 150,000 of those scraped interactions. They were likely hunting for chain-of-thought reasoning paths. This volume is a drop in the bucket for a frontier model. It looks more like a small internal test than a massive heist.
- Moonshot and MiniMax: Pulled millions of interactions focused strictly on tool calling and coding capabilities.
Lambert throws cold water on the idea that this “distillation” is a shortcut to top-tier AI. Distillation is just imitation. You are simply teaching a smaller model the shape of the right answer.
Real AI reasoning requires Reinforcement Learning (RL). During RL, a model explores, fails, and learns from its mistakes. It builds its own internal logic.
You cannot copy a competitor’s API and suddenly achieve frontier reasoning capabilities. DeepSeek’s recent success with their R1 model heavily relied on pure reinforcement learning rather than imitation, which perfectly validates Lambert’s technical assessment.
A Race to the Bottom
The current state of AI training is a free-for-all. Tech giants vacuum up every scrap of human knowledge they can find.
When they run out of free websites, they chop up physical books. When their competitors scrape their synthetic outputs, they cry foul.
This is not about safety. It is about building a commercial moat.
Editor’s Note: Anthropic builds incredible tools. I use Claude almost daily for logic testing. But their outrage over data scraping rings hollow when their own servers are stuffed with unauthorized scans of copyrighted books.
If the AI revolution requires stealing from everyone, we should at least be honest about it.
What do you think? Are tech giants justified in scraping public and copyrighted data to build better models, or is it plain theft? Sound off in the comments below.