AI's New Battlefield: Why 'Model Distillation' Just Became the Hottest Flashpoint in US-China Tech Relations

The Short Answer

The model distillation dispute is both technically fascinating and strategically consequential. Here's the practical bottom line:

Distillation is a real, widely used technique — not a euphemism for theft. Knowledge distillation has been a standard machine learning practice since at least 2015, used by researchers at Google, OpenAI, Meta, and universities worldwide. It's how you make AI models smaller, faster, and cheaper while preserving most of their capability. Every major AI company uses it.

The dispute is about scale, intent, and terms of service. The US accusation isn't that distillation itself is wrong — it's that Chinese companies have used it at industrial scale, against explicit API terms of service, to build directly competing products. The Chinese counterargument is that API terms of service restricting distillation are anticompetitive — an attempt to use legal fine print to prevent competition that technology alone can't stop.

There's no clear legal resolution in sight. US law doesn't clearly address whether training a model on another model's outputs violates anything — copyright law, computer fraud law, and trade secret law are all imperfect fits for the problem. API terms of service are contracts, not statutes. The dispute may ultimately be resolved through regulation (new rules about what you can and can't do with AI model outputs), industry norms (voluntary agreements not to distill competitors' models), or continued escalation (technical measures to prevent distillation).

Your practical takeaway: The distillation fight won't directly affect most business AI users in the near term. But it could shape which AI models are available, at what cost, and under what restrictions over the medium term. If the US restricts distillation, Chinese models may become less competitive with US models — reducing choice. If distillation continues unrestricted, Chinese models will keep improving rapidly — increasing competition and lowering costs. Either way, the smart strategy is the same: use multiple AI providers, don't bet exclusively on any single company or country's models, and stay informed as the regulatory landscape evolves.

What Is Model Distillation? A Plain-Language Explainer

The concept is simpler than the name suggests. Here's how it works:

The basic idea: You have a very large, very capable AI model — the 'teacher.' It's expensive to run and requires massive computing infrastructure. You want a smaller, cheaper, faster model — the 'student' — that's nearly as good. So you feed the teacher millions of prompts, collect its responses, and use those prompt-response pairs to train the student. The student learns to mimic the teacher's behavior without needing to understand why the teacher answered the way it did.

An analogy: Think of a master chef (the teacher) and an apprentice (the student). The apprentice can't replicate the chef's decades of experience, intuition, and creativity. But if the apprentice watches the chef make 10,000 dishes and carefully notes every ingredient, technique, and timing, the apprentice can produce dishes that are nearly indistinguishable from the chef's — without ever understanding the culinary principles behind them. That's distillation.

Why it's valuable: Training a frontier AI model from scratch costs hundreds of millions of dollars in computing alone, plus years of research. Distilling an existing model's capabilities into a smaller model costs a tiny fraction of that — thousands or tens of thousands of dollars in API fees. It's the difference between building a car factory from scratch and buying a finished car to study its design.

Why it's controversial: The teacher model's capabilities represent billions of dollars of investment by the company that built it. Distillation allows a competitor to capture much of that capability at minimal cost — without doing the underlying research, without building the infrastructure, and without taking the risks the original developer took. The question at the heart of the dispute: is this fair competition, or is it taking a free ride on someone else's massive investment?

The Accusation: What Anthropic and OpenAI Are Alleging

The US companies' complaints, which surfaced prominently in July 2026, make several specific claims:

Systematic, not incidental: Anthropic and OpenAI aren't alleging that a few Chinese researchers occasionally used their APIs for distillation experiments. They're alleging systematic, industrial-scale extraction — millions or billions of API calls designed specifically to capture frontier model capabilities for use in training competing Chinese models.

Terms of service violations: Both Anthropic and OpenAI's API terms of service prohibit using their outputs to train competing models. The allegation is that Chinese companies knowingly violated these terms, using intermediary accounts, shell companies, or third-party services to obscure the scale and purpose of their API usage.

The specific companies named: DeepSeek (known for highly competitive open-weight models), Moonshot AI (creator of the Kimi K3 model), and MiniMax (a major Chinese AI company) have been identified in reporting as targets of the accusations. All three have either denied the allegations or defended distillation as legitimate research practice.

The competitive impact: US companies argue that distillation at this scale allows Chinese firms to free-ride on American AI investment — capturing the benefits of models that cost billions to develop for the price of API access. This, they argue, undermines the economic incentive to invest in frontier AI research and creates an uneven playing field where Chinese companies can compete without bearing comparable R&D costs.

The security dimension: Some US officials have added a security concern: if Chinese companies can extract frontier AI capabilities through distillation, those capabilities could be transferred to Chinese military or intelligence applications without the safeguards and oversight that US companies build into their own deployments.

The Counterargument: China's Defense of Distillation

Chinese AI companies and researchers have pushed back forcefully against the accusations, making several arguments:

Distillation is standard global research practice. Google's 2015 paper introducing knowledge distillation is one of the most cited papers in machine learning. Every major AI lab — including OpenAI and Anthropic — has published research using distillation techniques. US companies use distillation to train smaller, more efficient models. The technique itself is not controversial in the research community — only its application to commercial competitors' outputs.

API terms of service shouldn't restrict research. Chinese researchers argue that using publicly available API outputs for research purposes — including training better models — is a legitimate activity that API terms of service shouldn't be able to restrict. They draw an analogy: if a company sells a book, it can't prohibit you from learning from that book and writing a better one. If a company sells API access to an AI model, it shouldn't be able to prohibit you from learning from that model's outputs.

The accusations are protectionism. Chinese officials and researchers argue that the real motivation behind the distillation accusations is protecting US AI companies from competition. Chinese open-weight models (DeepSeek, Qwen, Kimi) have become genuinely competitive with US models — and often significantly cheaper. Restricting distillation would remove one of the key techniques that enabled that catch-up, entrenching the advantage of well-funded US incumbents.

US companies also distill — they just don't call it that. Critics note that US AI companies train their models on vast quantities of internet data that includes outputs from other AI models — effectively a form of distillation, whether intentional or not. The distinction between 'training on internet data that happens to include AI outputs' and 'training on AI outputs deliberately' is blurry in practice.

The security argument cuts both ways. If distillation is restricted, the global AI ecosystem becomes more dependent on a smaller number of US-controlled proprietary models. That concentration creates its own security risks — including the risk that US AI companies could restrict access for geopolitical reasons. Chinese officials argue that a diverse, globally distributed AI model ecosystem — enabled by distillation and other techniques — is more resilient and secure than one dominated by a few US companies.

The Legal Landscape: What Law Actually Applies?

One reason the distillation dispute is so hard to resolve is that existing law doesn't clearly address it:

Copyright law: Can you copyright the output of an AI model? US law is unsettled. The Copyright Office has generally taken the position that works created by non-humans aren't copyrightable — which would mean AI outputs aren't protected by copyright and can't be infringed. But this isn't definitively resolved, and some argue that AI outputs should be protected as derivative works of the copyrighted training data. If AI outputs aren't copyrightable, distillation can't be copyright infringement.

Computer fraud law: The Computer Fraud and Abuse Act (CFAA) prohibits unauthorized access to computer systems. But using a publicly available API — even in violation of terms of service — may not constitute 'unauthorized access' under current legal interpretations. The Supreme Court narrowed the CFAA's scope in 2021, ruling that exceeding authorized access doesn't necessarily violate the statute. Whether violating API terms of service to distill a model is a CFAA violation is genuinely unclear.

Trade secret law: If a company can show that its model's outputs reveal trade secrets — proprietary information that derives economic value from not being generally known — distillation could potentially be trade secret misappropriation. But AI model outputs are generally not trade secrets: they're publicly available (to anyone with API access) and the model itself is the valuable asset, not individual outputs. This is a difficult legal theory to sustain.

Contract law: API terms of service are contracts. Using the API in violation of those terms is a breach of contract. But the remedy for breach of contract is typically damages — and proving damages from distillation (as opposed to lost API revenue) is difficult. An injunction prohibiting future violations might be available, but shutting down an already-trained model based on a contract claim would be extraordinary.

The bottom line: Existing US law doesn't provide a clear, effective remedy for distillation-based competition. This is part of why the dispute is escalating — and why it may ultimately require new legislation or international agreements rather than litigation under existing law.

What the Dispute Means for the AI Industry

1. The API model has a vulnerability. The distillation dispute exposes a structural vulnerability in the 'sell API access to your AI model' business model. If API access allows competitors to extract your model's capabilities and build competing products, the API business model is self-undermining — every customer is potentially a future competitor. AI companies may respond by: restricting API access more tightly, raising API prices to capture more of the distillation value, implementing technical measures to detect and prevent distillation, or shifting toward on-device and enterprise deployment models where the model is less accessible.

2. Technical anti-distillation measures are coming. AI companies are already working on technical measures to prevent or detect distillation: output watermarking (embedding detectable patterns in model outputs), query pattern analysis (detecting the large-scale, systematic querying characteristic of distillation), and output degradation for suspected distillation queries. These measures will create an ongoing arms race between distillation techniques and anti-distillation defenses.

3. Open-weight models change the calculus. Open-weight models (Llama, DeepSeek, Qwen, Mistral) don't have API terms of service — anyone can download and use them for any purpose, including distillation. The growth of capable open-weight models means distillation is becoming democratized: you don't need API access to a proprietary frontier model to practice distillation; you can use the best available open-weight models. This undermines the API-based anti-distillation approach and may accelerate the shift toward open-weight models as the distillation foundation.

4. The global AI ecosystem may fragment. If the US restricts distillation (through law or technical measures), and China permits it, the global AI landscape fragments: Chinese AI models improve rapidly through distillation while US models are protected but develop more slowly without the competitive pressure. If both countries restrict it, AI development slows globally but incumbents' positions are protected. If neither restricts it, competition intensifies and costs fall, but the economic incentive to invest in frontier research may weaken. These are genuinely difficult tradeoffs with no obvious right answer.

5. Businesses benefit from the competition — for now. The distillation-enabled competition between US and Chinese AI models has been a major driver of falling AI costs and improving capabilities. Chinese open-weight models, many of which benefited from distillation techniques, have given businesses access to capable AI at dramatically lower costs than US proprietary alternatives. If distillation is restricted, some of that competitive pressure diminishes — and AI costs may stop falling as fast. Businesses should enjoy the current competitive environment while recognizing it may not last.