Is model distillation always an attack?
No. Distillation is a legitimate training method, including when a provider creates a smaller model from its own outputs. Anthropic uses illicit distillation to describe covert extraction of another provider’s capabilities without authorization.
Answered in
Anthropic’s Distillation Defenses Show Why Extraction Is a Lifecycle ProblemAnthropic describes controls against model-output extraction, from reasoning access to account networks, while acknowledging tradeoffs and adaptive attackers.
Read the full analysisOther questions this article answers
More ai & society questions
- How much did Anthropic say its new controls reduced extraction?
- Does Anthropic's self-assessment block harmful requests?
- How accurate is Claude's self-assessment monitor?
- Are ChatGPT locked chats end-to-end encrypted?
- Do locked chats change model training or retention?
- What did the dissertation study measure?
- Does an AI-detector flag prove a student used AI improperly?
- Has the NBER study been peer reviewed?
Every answer on Crashtech is written by the editor of the article it comes from — never auto-summarised. Browse all answers or the AI & Society beat.