What is AI model distillation, and why is it so hard to stop?
Anthropic and OpenAI accused Chinese AI developers of mining their models’ answers to train cheaper copycats. Here’s how distillation works and why it’s so hard to stop
AI model distillation involves shrinking a large AI model down to a smaller one that mimics the original's abilities. Companies like Anthropic and OpenAI say Chinese developers are distilling their flagship models. To distill, developers train a smaller "student" model using prompts and responses from a larger "teacher" model. Distillation allows extracting the teacher's learned knowledge without collecting new data.
It's a way to get a cheaper, smaller version of a big model. Distillation isn't entirely illicit; smaller models are often used to save money and resources. In recent years, distillation has been linked to accusations of model theft, particularly against Chinese developers like DeepSeek and Moonshot AI. Anthropic and OpenAI have accused Chinese labs of distilling their models, calling it "illicit" and "unauthorized."
Researchers say distillation typically involves collecting thousands of AI responses to different queries and using them as training data for the smaller model. The goal is to reverse-engineer the original model's capabilities.
Written by urgent.news from Scientific American's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
