The Distillation Attack Hypocrisy: AI Giants Cry “Theft” After “Stealing” the Internet
AI companies are screaming “intellectual property theft” over “distillation attacks”.
Here’s the play: You (or Chinese labs) pay for API access to Claude or Gemini. You hammer it with millions of smart prompts. You capture the high-quality outputs and reasoning traces. You train a cheap “student” model to copy the expensive “teacher.”
Anthropic just publicly shamed DeepSeek, Moonshot, and MiniMax for doing exactly this: 16+ million exchanges through 24,000 fake accounts, all violating ToS. Google calls it model extraction and IP theft, threatening legal action and bans.
They’re right that the fake accounts and evasion are shady. But the hypocrisy is nuclear.
These same companies built their trillion-dollar models by distilling the entire internet. Books, articles, code, images, Reddit, etc. without permission or payment. They called it “fair use.” Now they’re getting sued by The New York Times, authors, artists, and publishers.
Yet the moment someone pays their API bills and learns from the outputs they sold, it’s suddenly “industrial-scale theft.”
You can’t vacuum up humanity’s knowledge for free, then clutch pearls when others do the same with data you willingly served (even if they gamed the system).
If your model can be meaningfully distilled by querying your own paid API, your moat was never that strong to begin with.
The internet distilled you first. Now you’re mad others are distilling the distillate?