Terminology matters: Why conflating distillation attacks with a core AI technique risks regulatory overreach

The recent alarm over "distillation attacks" conflates legitimate industry practice with actual malicious behavior, risking severe unintended consequences for the Western AI ecosystem. Model distillation—training a smaller model on outputs from a larger one—is a foundational technique used by frontier AI companies to produce cost-effective versions of their models and remains standard practice in post-training workflows across the industry. While certain Chinese laboratories have indeed abused APIs through jailbreaking and unauthorized access to extract additional training data, the core issue stems from these evasion methods, not from distillation itself. The technique has existed in regulatory gray area for years: major AI providers like OpenAI and Anthropic formally prohibit competitors from using their APIs to build rival products, yet enforcement remains minimal and inconsistent. Companies including xAI have distilled from competitors, and smaller research organizations routinely distill from multiple open and closed model sources to build specialized systems. The current policy response—congressional bills, executive orders, and regulatory scrutiny—threatens to permanently associate distillation with criminal conduct. This could trigger restrictions on open-source models developed using Chinese models, decimating the independent research community and smaller enterprises while likely doing little to deter Chinese companies from continuing the practice. The removal of open-source alternatives would accelerate researchers and companies toward proprietary platforms, consolidating power rather than protecting it. The path forward requires precise terminology: address API abuse and jailbreaking as security violations without stigmatizing distillation as a research and development method. Broad restrictions on a core technique would amount to regulatory overreach that harms Western innovation more than Chinese advancement.