← Back to articles
Next (newer) →
DeepSeek's $20B Problem: When Your Best People Walk Out the Door
geopolitics · technology · business · April 23, 2026 ✨ Recommended 🎥 Video

The AI Cold War's Newest Weapon: Stealing the Teacher's Answer Key

Admin ★★★★★
No reader ratings yet.
Log in to rate this article
📰 Reading Passage

On April 23, 2026, the White House made an unusually direct accusation: Chinese entities, the administration claimed, are running an industrial-scale campaign to steal from American artificial-intelligence companies. The technique at the heart of the accusation is called distillation, and it works by feeding millions of carefully crafted questions to a powerful model and using its answers to train a smaller, cheaper imitator. According to a memo from Michael Kratsios — the chief science and technology adviser to the president — foreign actors are deploying thousands of fake accounts and so-called jailbreaking tricks to extract knowledge from frontier models built by OpenAI and Anthropic.

The timing of the announcement is its own story. President Donald Trump is scheduled to meet Chinese President Xi Jinping in Beijing within weeks. China's embassy in Washington dismissed the claim as pure slander. Meanwhile, the US Congress passed bills that could blacklist Chinese AI firms suspected of distillation, cutting them off from American technology entirely. Anthropic, in a February 2026 disclosure, said it had documented more than 24,000 fraudulent accounts attached to three Chinese AI labs — DeepSeek, Moonshot, and MiniMax — generating roughly 16 million Claude exchanges between them.

What makes the dispute genuinely difficult is that distillation is a normal, even ordinary, technique inside the AI industry. Companies use it on themselves all the time to make smaller, cheaper versions of their own models. The fight is over who is allowed to do it to whom. Since 2022, the United States has restricted China's access to the most advanced Nvidia chips precisely to slow Beijing's AI progress; distillation is one of the few workarounds that does not require an enormous amount of computing power. A copycat model can approach the capabilities of the original at a fraction of the cost.

The strategic worry runs deeper than industrial espionage. Frontier AI models built in the United States are trained with elaborate safety guardrails — restrictions designed to prevent the model from helping users build weapons or launch cyberattacks. A distilled clone often inherits the underlying capabilities without inheriting the safety layer. That means the cheaper, looser copy may actually be more dangerous than the original, even though it is technically less capable.

If the accusations harden into policy, the consequences ripple outward. Tighter export controls on AI chips and software become more likely. Chinese users could lose direct access to American AI services. Universities and research labs face new restrictions on collaboration. The deeper question being asked in Washington is whether AI leadership can be defended at all when knowledge moves so easily through people, papers, products, and now through models talking to other models. The lead America believes it has, in other words, may be smaller than it thinks.

Source: https://www.ft.com/content/abde4e1e-c69a-4cc4-ad96-d88308314298?syn-25a6b1a6=1

📎 Download Original ⬇ Download Analysis PDF

🎥 Video Overview

📖 Explanation

Imagine cheating on a test by secretly photographing every answer the smartest kid writes — then selling your knockoff version for a tenth of the price. That's what the White House just accused China of doing to American AI.

📖 What's Going On?

On April 23, 2026, the White House accused Chinese entities of running 'industrial-scale' campaigns to steal from American AI labs through a technique called distillation. In a memo, Michael Kratsios — the chief science and technology adviser to the president — said foreign actors are using thousands of fake accounts and 'jailbreaking' tricks to extract knowledge from US frontier AI models like those built by OpenAI and Anthropic. Anthropic, in a separate February 2026 disclosure, said it had documented more than 24,000 such fraudulent accounts attached to three Chinese AI firms.

The accusation lands weeks before Donald Trump is scheduled to meet Xi Jinping in Beijing. China's embassy called the claims 'pure slander.' Meanwhile, Congress passed bills that could blacklist Chinese AI firms suspected of distillation, cutting them off from US technology entirely.

🎯 How To Think About It

Distillation is a normal AI technique — it's how companies make smaller, cheaper models by training them to mimic bigger ones. The fight is over who's allowed to do it to whom.

💡 Key Things To Know

🌟 Why It Matters

If you're considering a career in tech, computer science or international policy, this is the defining conflict of the next two decades. AI capability is now treated like nuclear capability was in 1955 — a strategic asset governments will fight, sanction and spy over. The apps you use, the chips in your phone, even which universities can collaborate with whom: all of it is being reshaped by this rivalry right now.

🔮 The Bigger Picture

The deeper question is whether AI leadership can be defended at all. Knowledge tends to leak — through people, papers, products and now apparently through models talking to each other. If distillation is unstoppable, the US edge built on expensive chips and elite researchers may be smaller than Washington thinks. Watch for: tighter export controls, possible bans on Chinese users accessing US AI services, and a fragmenting global internet where American and Chinese AI ecosystems no longer talk to each other at all.

📚 Key Terms Glossary

Distillation
An AI training method where a smaller model learns by mimicking the outputs of a larger, more powerful one — useful when done legitimately, controversial when done by copying someone else's proprietary model.
Frontier AI
The most advanced AI systems currently in existence, like GPT-class models, that push the limits of what the technology can do.
Jailbreaking
Tricking an AI model into bypassing its built-in restrictions — for example, getting it to reveal information or behaviour its creators tried to block.
Entity list
A US government blacklist of foreign companies and organisations that American firms are largely banned from selling technology to; being added to it can cripple a company's supply chain.
Export controls
Government rules restricting which technologies can be sold abroad — the US uses them to prevent China from acquiring advanced AI chips and equipment.
Proxy accounts
Fake or borrowed user accounts used to disguise who is really accessing a service, often to evade detection or rate limits.
Intellectual property
Legally protected creations of the mind — patents, trade secrets, copyrighted code — that companies and countries treat as valuable assets worth defending.
Safeguards (in AI)
Built-in restrictions that prevent a model from producing dangerous content, such as instructions for weapons or malware.

✏️ Reading Comprehension Quiz

Tip: log in or create a free account to save your score, earn badges, and appear on the leaderboard. Otherwise the quiz works fine without an account.
Question 1
Which choice best states the central idea of the passage?
Question 2
According to the passage, why are American AI companies concerned about distilled models specifically on national security grounds?
Question 3
As used in the passage, the word 'distil' most nearly means
Question 4
As used in the passage, 'frontier' most nearly means
Question 5
Which statement about export controls can most reasonably be inferred from the passage?
Question 6
The passage suggests that DeepSeek's emergence was significant primarily because
Question 7
The author's tone in describing the White House accusations is best described as
Question 8
The author's primary purpose in this article is to
Question 9
Which statement about the relationship between computing power and distillation can most reasonably be inferred?
Question 10
Which choice provides the best evidence for the answer to the previous question?
Next (newer) →
DeepSeek's $20B Problem: When Your Best People Walk Out the Door