DeepSeek, a Chinese AI startup, released its R1 model about a year and a half ago and achieved frontier-level performance on key benchmarks at a fraction of the cost that major U.S. labs spent on their flagship models. The achievement has triggered accusations from OpenAI and Anthropic that DeepSeek distilled their proprietary models to train its own systems, reigniting debate over a technique that is reshaping how AI is developed globally and raising questions about whether cheaper AI development can coexist with intellectual property protection.

What Is Model Distillation and Why Does It Matter?

Model distillation, also known as knowledge distillation, is a process where a smaller "student" AI model learns from a larger, more powerful "teacher" model. Rather than training from scratch on massive datasets, the student model mimics the teacher's behavior and reasoning patterns, dramatically reducing the computational cost and time required to build capable AI systems.

Think of it like learning from a mentor. A flagship model like ChatGPT or Claude has already done the hard work of understanding language, math, and reasoning through billions of dollars in training. A distilled model doesn't repeat that expensive process; instead, it watches how the teacher model reasons through problems and learns to replicate that approach. The teacher doesn't just give yes-or-no answers; it shares its probabilistic reasoning, showing the student model the confidence levels behind each decision. These "soft labels" help the student build its own reasoning skills more efficiently.

How Did DeepSeek Achieve Frontier Performance at Lower Cost?

DeepSeek's R1 model approached the performance level of cutting-edge U.S. models on several important benchmarks, but the company reportedly built it at a fraction of the cost that OpenAI, Anthropic, and other major labs spent on their flagship systems. This cost efficiency caught the attention of U.S. AI developers, who quickly accused DeepSeek of using model distillation to extract knowledge from their proprietary models.

OpenAI alleged that DeepSeek distilled ChatGPT to train R1. More than a year later, Anthropic made similar accusations against DeepSeek, as well as against fellow Chinese AI labs Moonshot and MiniMax, claiming they distilled Claude. Anthropic also accused Chinese tech giant Alibaba in June of "illicitly" using Claude for distillation purposes. Notably, Anthropic used the word "illicitly" rather than "illegally," reflecting the murky legal landscape around distillation in the AI industry.

Why Is the U.S. Government Cracking Down on AI Distillation?

The Trump administration has vowed to crack down on Chinese labs distilling U.S.-made models, framing the issue as a matter of geopolitical importance. Michael Kratsios, director of the White House Office of Science and Technology Policy, acknowledged in an April memo that legitimate distillation is "a vital part of" the AI ecosystem when used to create smaller, lighter-weight models. However, he drew a line at what he called "industrial distillation activities that aim to systematically undermine American research and development and access proprietary information".

"AI distillation, when legitimately used to produce smaller, lighter-weight models from more advanced systems, is a vital part of the AI ecosystem. Industrial distillation activities that aim to systematically undermine American research and development and access proprietary information, however, are unacceptable," stated Michael Kratsios.

Michael Kratsios, White House Office of Science and Technology Policy Director

The distinction reflects a broader anxiety among U.S. policymakers and AI companies: if Chinese labs can achieve frontier-level AI performance at a fraction of the cost by distilling U.S. models, the competitive advantage that American companies have built through massive R&D spending could erode rapidly. This concern has made distillation one of the most contentious issues in the AI industry.

What Are the Safety Risks Beyond Geopolitics?

The concerns about distillation extend beyond competition and national security. When flagship models are trained on vast amounts of internet data, they absorb not only useful patterns but also deeply ingrained human biases. Through careful fine-tuning and additional training, developers can gradually teach models to recognize and avoid those biases. Similarly, models can be fine-tuned to detect when users are attempting to misuse them for harmful purposes, such as generating bioweapon recipes, and to refuse such requests.

The problem is that these safety safeguards don't always transfer cleanly when a model is distilled. A student model learning from a teacher may not inherit the full depth of the safety training that went into the teacher model. This creates a potential gap: a distilled model could be more vulnerable to generating biased or harmful outputs than the original, even if it matches the teacher's performance on standard benchmarks.

How Does Model Distillation Compare to Building from Scratch?

• Training Cost: Building a flagship model from scratch requires billions of dollars in computing resources and access to vast amounts of training data. Distillation dramatically reduces this cost by leveraging an existing teacher model's knowledge.
• Development Speed: Training a flagship model can take months or years. Distillation accelerates the process because the student model doesn't need to rediscover fundamental patterns about language, math, and reasoning.
• Safety and Alignment: Flagship models undergo extensive fine-tuning to reduce bias and harmful outputs. Distilled models may not inherit these safeguards fully, potentially creating new risks if not carefully managed.
• Democratization: Distillation makes it possible for smaller companies and labs with fewer resources to build capable AI models, potentially democratizing AI development beyond the handful of well-funded U.S. labs.

Steps to Understanding the Distillation Debate

• Recognize the Legitimate Use Case: Model distillation is not inherently unethical or illegal. It's a standard technique in machine learning that allows researchers to build upon prior innovations, much like how the Wright Brothers' airplane design enabled future aviators to iterate and improve.
• Distinguish Between Legitimate and Illicit Distillation: Using a publicly available model or one you have permission to use for distillation is legitimate. Systematically extracting proprietary information from a competitor's model without authorization is what regulators and companies view as problematic.
• Consider the Geopolitical Context: The U.S. government's focus on cracking down on Chinese distillation reflects broader concerns about AI leadership and technological advantage, not just intellectual property protection.
• Evaluate Safety Implications: When assessing distilled models, consider whether they have undergone the same rigorous safety testing and fine-tuning as their teacher models, since safeguards don't always transfer automatically.

The distillation debate reveals a fundamental tension in AI development: the technology that could democratize AI and make it more accessible also threatens to undermine the competitive advantages of companies that invested heavily in building flagship models. As the U.S. and China compete for AI dominance, how regulators and companies navigate this tension will shape the future of AI development for years to come.