The AI Arms Race's Hidden Weapon: Copying Smart to Get Smarter
openrouter:minimax/minimax-m2.7 · prompt: edited
· 2026-05-30T13:56:00 · 0.62¢
· 🔍 view original input ↗
📃 Rewritten passage
The White House has accused China of stealing American artificial intelligence technology on an unprecedented scale, raising the stakes in what officials describe as a modern technological arms race.
In a memo to government departments, Michael Kratsios, director of the Office of Science and Technology Policy, said US intelligence indicated that Chinese entities were conducting "deliberate, industrial-scale campaigns to distil US frontier AI systems." Distillation is a legitimate technique in AI development: it involves training smaller, cheaper models to behave like larger, more powerful ones by learning from their outputs. But Kratsios alleged that Chinese firms had perverted this practice, using "tens of thousands of proxy accounts" to extract data from American AI labs without authorization.
The accusations came just weeks before President Trump was scheduled to meet President Xi Jinping in Beijing, adding diplomatic friction to an already tense relationship. China rejected the claims as "pure slander."
The method gained attention after DeepSeek, a Chinese AI startup, released a powerful model at a fraction of the typical development cost. American AI companies including Anthropic and OpenAI have raised alarms about distillation attacks. In February 2026, Anthropic publicly accused three leading Chinese firms — DeepSeek, Moonshot, and MiniMax — of using distillation techniques on its models. OpenAI had previously said it found evidence that DeepSeek had used outputs from its GPT models to train its own system in early 2025, a potential violation of service terms.
The competitive implications are significant. Frontier AI models require billions of dollars in computing infrastructure and years of specialized research to develop. Distillation allows rivals to build similarly capable products at dramatically lower cost by studying outputs rather than replicating the underlying research process. This concerns US firms because distillation effectively circumvents export controls on advanced American chips — the very mechanism the US has used to slow Chinese AI development.
American AI companies also warn of safety risks. Models trained through distillation may lack the guardrails built into their source systems — safeguards designed to prevent AI from assisting with bioweapon development or malicious cyberattacks. The House Foreign Affairs Committee responded by passing legislation that could add distillation-using entities to the "entity list," an export blacklist that would block US technology companies from doing business with them.
Kratsios acknowledged that distillation serves a legitimate purpose in the AI ecosystem, making powerful technology accessible on smaller devices. But he emphasized that using it to "undermine American research and development" crossed a line. The question now is whether legal and diplomatic pressure can prevent what US officials see as systematic theft — or whether the global AI race will increasingly split into separate, non-compatible ecosystems.
Imagine you spent years perfecting the ultimate study guide, only to discover your rival had been watching over your shoulder the entire time — and built their own version for a tenth of the cost.
📖 What's Going On?
The White House has accused China of conducting "industrial-scale" theft of American artificial intelligence technology. Michael Kratsios, director of the Office of Science and Technology Policy, said US intelligence indicates Chinese entities are deliberately using a technique called distillation to replicate US frontier AI models.
The method exploits a legitimate practice: distillation normally trains smaller, cheaper AI models using outputs from larger ones. But Chinese firms allegedly ran massive campaigns using "tens of thousands of proxy accounts" to extract that data without authorization, allowing them to build competitive products at a fraction of the cost.
The accusations escalate existing tensions before President Trump meets President Xi Jinping in Beijing in coming weeks. China's embassy called the claims "pure slander."
🎯 How To Think About It
Think of frontier AI models like a world-class university. Training one requires billions of dollars in computing power, years of PhD-level research, and vast datasets. Distillation is essentially a student who copies the professor's answer key, studies the graded exams of top students, and then trains new students to perform almost as well — without ever sitting in the lectures or doing the original research.
Here's the catch: the copied version is cheaper to run and deploy, meaning it can reach consumers faster and at lower prices. It's like someone reverse-engineering a Formula 1 car's performance using footage of races, then selling a street-legal version at Toyota prices.
- The legitimate AI ecosystem relies on distillation to make powerful AI accessible — it's how your phone can run complex tasks without a supercomputer
- The US already restricts advanced AI chips from reaching China through export controls; distillation attacks are seen as a way to circumvent those limits
💡 Key Things To Know
- Anthropic accused three Chinese firms — DeepSeek, Moonshot, and MiniMax — of distillation attacks in February 2026
- OpenAI previously said DeepSeek used GPT model outputs to train its own model in early 2025, violating its terms of service
- Distilled models can lack crucial safety guardrails — the safeguards that prevent AI from, say, assisting in bioweapon development
- The House Foreign Affairs Committee passed bills that could add distillation-using entities to the "entity list" — an export blacklist blocking US tech sales to those groups
- Most people assume AI competition is about compute power and algorithms; this story reveals it's equally about who can access and legally use that knowledge
🌟 Why It Matters
If you're interested in tech careers, this illustrates a fundamental shift in how innovation gets guarded. Traditional intellectual property law assumes someone can observe a product and reverse-engineer it; AI changes that because the 'product' is a process that runs continuously and generates outputs that can themselves be used to train competitors. Every tech company, university lab, and startup is now asking: how do you protect your research when the output itself becomes the vulnerability?
For students watching US-China relations, this isn't abstract geopolitics — it's the template for how great-power competition will actually play out in the 21st century. The countries that set the rules for AI governance will shape global power for decades.
🔮 The Bigger Picture
This dispute sits at the intersection of three explosive trends: US-China strategic rivalry, the global race to lead in AI, and debates about who should govern transformative technology. If the US successfully bans distillation-based Chinese AI firms from accessing American models, it may accelerate a 'splinternet' for AI — separate Chinese and Western ecosystems that don't share data, standards, or safety norms.
Watch for whether other governments (EU, India) adopt similar restrictions, and whether American AI firms themselves face pressure to limit open-source releases that make distillation easier. The next phase of the AI arms race won't be won by the fastest model — it'll be won by whoever controls who can learn from whom.
📖 Glossary (6)
- Distillation — A machine learning technique where a smaller AI model is trained to mimic the behavior of a larger, more powerful one. The smaller model learns from the larger model's outputs, achieving similar performance at lower computational cost.
- Frontier AI — The most advanced, cutting-edge AI systems, typically requiring hundreds of millions to billions of dollars to develop. These represent the current limits of what AI can do.
- Entity List — A US Commerce Department blacklist of foreign individuals and companies restricted from purchasing American technology. Being added means US firms cannot legally sell to that entity without special government approval.
- Export controls — Government restrictions on what technology can be sold to foreign countries, typically justified on national security grounds. The US has used these to limit China's access to advanced AI chips.
- Proxy accounts — Fake or borrowed user accounts used to access a service while concealing the true identity of the operator. Chinese entities allegedly used tens of thousands of these to extract AI outputs without detection.
- AI safety guardrails — Built-in restrictions in advanced AI systems that prevent them from generating harmful content — such as instructions for weapons, biotoxins, or cyberattacks. Distilled models may lack these safeguards if they weren't explicitly trained into the smaller copy.
📝 Quiz (10) — with answers
Q. The passage primarily argues that China is conducting industrial-scale theft of American AI technology through which method?
A. Illicit acquisition of advanced AI chips
B. Large-scale hacking of American computer systems
C. Training AI models using outputs from US systems ✓
D. Recruiting American AI researchers to work in China
Answer: C — The passage is explicit: Kratsios said foreign entities 'are engaged in deliberate, industrial-scale campaigns to distil US frontier AI systems.' Distillation — training smaller models on the outputs of larger ones — is the specific technique described. Options A and B are plausible real-world concerns but aren't what this article argues is happening. Option D isn't mentioned.
Q. According to the passage, why does distillation pose a competitive threat to American AI companies?
A. It allows Chinese firms to develop models with superior safety standards
B. It enables Chinese firms to create competitive AI at significantly lower cost ✓
C. It gives Chinese firms access to proprietary chip manufacturing techniques
D. It lets Chinese firms circumvent patent registration requirements
Answer: B — Kratsios states that while distilled models don't match original models in performance, 'they can benefit foreign groups because of the significantly lower cost.' This cost advantage is the central competitive threat described. Option A is wrong because the article explicitly says distilled models 'lack the safeguards' of originals. Options C and D aren't mentioned in the passage.
Q. The passage suggests that American AI companies view distillation attacks as threatening their competitive advantage because:
A. Chinese companies can replicate their models exactly
B. Export controls on advanced chips can be bypassed through distillation ✓
C. Distilled Chinese models will outperform US models within two years
D. US companies will be legally required to share their models with Chinese firms
Answer: B — The article states that American AI companies 'argue enables foreign labs to close the competitive advantage that the US enjoys because of export controls on advanced American chips.' This directly connects distillation to circumventing chip restrictions. Option A contradicts the passage, which says distilled models 'did not match the performance of the original models.' Options C and D aren't supported by the text.
Q. As used in the passage, the word "leverage" most nearly means:
A. To pry open something using force
B. To use something to gain an advantage ✓
C. To provide financial support
D. To threaten or coerce someone
Answer: B — The passage states Chinese campaigns were 'leveraging tens of thousands of proxy accounts.' In context, this means they were using those accounts as a tool to accomplish their goal. The passage meaning is instrumental/functional use. The common meaning of 'leverage' (financial leverage, borrowing money) is irrelevant here. Options A, C, and D don't fit the sentence structure.
Q. As used in the passage, the word "surreptitious" most nearly means:
A. Extremely large in scale
B. Carried out in a stealthy, hidden manner ✓
C. Approved by government authorities
D. Focused on achieving commercial profits
Answer: B — Kratsios references models created by 'surreptitious, unauthorised distillation campaigns.' The word modifies 'campaigns' in the context of being unauthorized and hidden from detection — the passage uses it to emphasize the covert nature of the activity. Option A ('large scale') is the opposite of what 'surreptitious' conveys. Options C and D contradict the passage's framing of unauthorized activity.
Q. Which statement about the House Foreign Affairs Committee can most reasonably be inferred from the passage?
A. The Committee believes current export controls on AI chips are insufficient ✓
B. The Committee has voted to ban all Chinese AI companies from operating in the US
C. The Committee opposes the White House's stance on Chinese AI theft
D. The Committee's bills apply equally to all countries attempting AI distillation
Answer: A — The Committee passed bills designed to make it 'harder for China to catch up in the AI race' specifically by targeting distillation. This implies existing export controls (focused on chips) are seen as inadequate to stop China. The passage doesn't say the bills ban all Chinese companies (B), doesn't show Committee-White House disagreement (C), and the bills appear China-specific (D isn't supported).
Q. The passage suggests that the safety concerns about distilled AI models stem primarily from:
A. Faster development timelines that skip testing phases
B. The absence of safety guardrails in the smaller distilled model ✓
C. Intentional removal of safeguards by Chinese companies
D. The use of older, less sophisticated training data
Answer: B — The passage states that American AI companies are concerned because distilled models 'pose national security risks because they lack the safeguards that, for example, prevent the development of bioweapons.' This is a capability gap, not intentional removal (C) or testing shortcuts (A). Option D isn't mentioned. The core concern is what's missing from the distilled copy, not what's been deliberately stripped.
Q. The author's tone in the 'What's Going On?' section is best described as:
A. Indifferent and detached
B. Alarmed and urgent
C. Analytical and explanatory ✓
D. Skeptical of the US accusations
Answer: C — The section presents facts neutrally — who said what, what distillation is, what the Chinese embassy responded — without emotional language or alarm. Words like 'industrial-scale' and 'deliberate' come from Kratsios's memo, not the author's commentary. Option B overstates the author's framing; options A and D are clearly wrong given the factual, balanced presentation.
Q. Based on the passage, which statement best explains why the US government is sharing intelligence about distillation attacks with American AI companies?
A. To help those companies pursue legal action against Chinese firms
B. To enable companies to better protect their AI systems from unauthorized extraction ✓
C. To convince companies to slow down their AI development to prevent theft
D. To encourage companies to openly share their models with government agencies
Answer: B — Kratsios said the administration would 'share information with American AI companies about attempts by foreign actors to conduct unauthorised, industrial-scale distillation' and 'help them co-ordinate against attacks.' The purpose is defensive: helping companies protect themselves. Option A (legal action) isn't mentioned; Options C and D contradict the passage's framing of competitive AI development.
Q. Which choice provides the BEST evidence for the answer to the previous question?
A. "The issue gained attention after China's DeepSeek was accused of using distillation"
B. "The US would explore measures 'to hold foreign actors accountable for industrial-scale distillation campaigns'"
C. "The administration would share information with American AI companies about 'attempts by foreign actors to conduct unauthorised, industrial-scale distillation'" ✓
D. "China's embassy called the accusations 'pure slander'"
Answer: C — This quote directly supports the answer to Q9: sharing information with companies about unauthorized distillation attempts is precisely the defensive coordination the passage describes. Option A is background context; Option B describes accountability measures, not defensive sharing; Option D is China's response and doesn't address the sharing of intelligence. Only C matches both the mechanism (sharing information) and the purpose (coordination against attacks) described in the passage.
💬 Suggested questions
- What is 'distillation' in AI, and why is it controversial?
- How could distillation allow China to bypass US chip export controls?
- What safety guardrails might be missing from a distilled AI model compared to the original?
Raw JSON