逐字稿 🔗
WEBVTT
1
00:00:00.000 --> 00:00:35.000
This is the Intellectual Property News podcast with Richard Gerhardt. AI learned from us now it doesn't want Other AI learning from it. Can an artificial intelligence company build its technology by learning from other people's work and then object when another AI company learns from it? That is at the center of a new controversy involving Anthropic, the company behind Claude. Anthropic claims that several Chinese technology companies used a process called distillation to copy Claude's most valuable capabilities.
2
00:00:35.000 --> 00:01:02.000
Distillation involves using a powerful teacher model's answers to train a smaller student model. Anthropic alleges that companies associated with Alibaba, Deepseek, and others secretly generated millions of exchanges with Claude, sometimes using fake accounts to to evade Anthropic's restrictions. Anthropic says Alibaba was connected to more than 151 million exchanges.
3
00:01:02.000 --> 00:02:20.000
These are Anthropic's allegations, not court findings. But here's the ironic part. Anthropic has itself faced major copyright lawsuits accusing it of taking other people's work to train Claude. In the court case Bart's vs. Anthropic, a group of authors alleged that Anthropic used pirated books to build its training library. The judge ruled that training an AI model using lawfully obtained books could be fair use, but described Anthropic's downloading of millions of books from pirate websites as piracy and not fair use. Anthropic subsequently agreed to a $1.5 billion settlement. Anthropic was also sued. In Concord Music Group vs. Anthropic, music publishers alleged that Claude was trained on copyrighted song lyrics and and could reproduce those lyrics without authorization. And in the newly filed Sony Music Publishing vs. Anthropic case, Sony and Warner Chapel allege that Anthropic illegally downloaded, scraped and copied thousands of music compositions to develop Claude. Anthropic denies those allegations and says it will defend itself in court. This presents an uncomfortable question for the AI industry.
4
00:02:20.000 --> 00:03:50.000
AI developers argue that their models do not merely copy music, books and artwork. They study those materials and learn patterns from then. But a competing AI company could make a remarkably similar argument. We didn't steal your computer code. Our model simply studied your model's answers and learned from them. Anthropic may still have claims based on breach of contract, fraud, trade secret misappropriation, or computer security laws, particularly if fake accounts or stolen credentials were used. But copyright generally protects original works of expression, not ideas, methods, or reasoning abilities. That is the central contradiction in the AI revolution. AI companies want broad freedom to learn from the world, but strong protections against the world learning from them. I find it annoying that tech companies feel entitled to a special set of rules just for them. So who gets to decide when learning becomes stealing? And should the same rule apply to everyone? And this has been Richard Gerhart for Intellectual Property News. And if you have an idea or an invention you want to protect, the team at Gerhart Law helps entrepreneurs turn ideas into protected assets. You can visit learn more about Patents.com or learn more about Trademarks.com for a free consultation and practical guides to get you started.