What is AI model distillation, and why is it a geopolitical fight?

The technique that built cheap AI, and started a political storm

Earlier this year, a word that had previously lived almost entirely inside AI research papers started turning up in Congressional hearings, government reports, and geopolitical briefings. That word is distillation. Distillation has long been a topic for AI wonks, but it has become a hot-button issue as technologists and lawmakers debate how it should be regulated. To understand why, you first need to understand what it actually is, because the concept is genuinely useful and the controversy around it is genuinely messy.

What distillation means in plain English

Building a frontier AI model from scratch is extraordinarily expensive. Training runs for the biggest systems cost hundreds of millions of pounds in compute time, consume vast quantities of electricity, and take months. The result is a powerful but unwieldy thing: too large to run on a laptop, too expensive to serve to millions of users cheaply.

Distillation is a way around that problem. It trains a smaller "student" model to mimic a larger "teacher." The process transfers probability distributions, intermediate features, and relational patterns, replicating the teacher's behaviour while reducing size and complexity.

A reasonable analogy: imagine a junior doctor who has never read a textbook being trained not from scratch, but by shadowing an experienced consultant for a year. They absorb the consultant's reasoning and judgement. They do not need to repeat all the original reading and training the consultant did. They come out smaller, cheaper to employ, and surprisingly capable given how little raw schooling they received.

Knowledge distillation as a formal technique was first proposed by Geoffrey Hinton and others in 2015. Back then it was mostly applied to image recognition models. The core idea: instead of training a small model on raw labelled data, you train it on the outputs of a big model. Those outputs carry much richer information than a simple yes/no label. A teacher model does not just say "this is a cat." It says "this is 94% likely to be a cat, 4% a fox, 1% a dog." That probability distribution is a kind of compressed wisdom, and the student learns from it far more efficiently.

Why it matters for modern AI

In distillation, a large, complex AI model (the teacher) transfers its knowledge to a smaller AI model (the student). The goal is to keep most of the intelligence while reducing size, cost, and computing requirements. That matters for several practical reasons.

• Running AI on devices. The method powers efficient models used in phones, private networks, and enterprise software. The AI assistant on your phone almost certainly runs a distilled model, not the giant version sitting in a data centre.
• Cutting inference costs. Distillation speeds up inference and reduces compute without a large hit to accuracy. If you are serving millions of queries a day, that saving adds up fast.
• Turbocharging new models. Since early 2025, distillation has greatly increased the efficiency with which some models can be fine-tuned. Rather than teaching a student model everything from raw data, you let an already-smart teacher do most of the heavy lifting.

A real example makes this concrete. DeepSeek developed a large model called DeepSeek-R1, which excels at chain-of-thought reasoning. R1 produced reasoning outputs that were then used to fine-tune smaller student models, including DeepSeek-V3. DeepSeek-V3 maintains much of R1's mathematical, coding, and document-analysis capabilities and was reportedly fine-tuned for approximately $10,000. That is likely orders of magnitude lower than the cost of fine-tuning similarly capable, larger models.

Where the controversy comes in

Distillation from your own teacher model is entirely legitimate and widely practised. The controversy arises when a company is accused of distilling from someone else's model without permission.

At a high level, distillation refers to the use of answers from a chatbot or work product from an advanced AI model to train another model. The practice is controversial because, depending on how it is used, it can allow a model developer to create a competitive offering by simply using the output from companies that have invested many millions or billions of dollars developing the most sophisticated training technology.

The specific flashpoint was DeepSeek. When the Chinese lab released its R1 model in January 2025, it stunned the industry by matching top American models at a fraction of the reported cost. It confronted the AI industry with a question that has since escalated into a geopolitical dispute: how did a Chinese lab achieve frontier-level reasoning capabilities at roughly one-tenth the compute cost of its American rivals?

OpenAI's answer was distillation. OpenAI investigated and found accounts linked to DeepSeek pulling massive amounts of data through their API. Microsoft blocked the accounts. OpenAI claims to have detected activity consistent with persistent attempts to distill its models, even through obfuscated methods, with accounts associated with DeepSeek employees allegedly developing techniques to bypass access restrictions using third-party routers and other methods designed to hide the origin of requests.

OpenAI told US lawmakers that rival DeepSeek secretly copied its frontier systems. The accusation centres on distillation performed through automated scraping and network obfuscation. DeepSeek has not publicly confirmed this allegation.

The legal tangle

Here is where things get genuinely complicated, and the answer is not what you might expect.

IP law traditionally protects creative works as copyrights, but the application of copyright law to AI-generated outputs is not straightforward. OpenAI's own terms of service create an awkward problem for any legal action. In OpenAI's own terms of use, it specifically states that users retain ownership of their inputs and own the outputs. OpenAI assigns all its rights in any output to the user. Therefore, even if OpenAI can present enough evidence to show that DeepSeek extracted data from its models, OpenAI likely does not have copyrights over that data.

What OpenAI can argue is a terms-of-service violation. Companies such as OpenAI, Anthropic, Mistral, and xAI include strict anti-competitive distillation clauses in their terms of use, prohibiting users from using their services or outputs to develop competing models. But enforcing terms of service against a Chinese company operating mainly outside US jurisdiction is a different matter from enforcing copyright.

Federal authorities indicated they were investigating potential trade secret and computer fraud violations, but enforcement against a Chinese firm operating primarily outside US jurisdiction presents serious practical obstacles.

There is also a pointed irony that critics have not let pass. OpenAI is simultaneously fighting The New York Times in court, arguing it had every right to train ChatGPT on copyrighted news articles without permission. The company that argues training on others' content is fair use is now arguing that training on its content is theft. The legal and moral logic does not sit entirely comfortably in either direction.

If this one was useful, there's plenty more on the site. Pieces on how AI works, plus coverage of AI news, the downsides included. All free to read, no account needed.

What this means for the AI industry

The distillation dispute is not just a legal curiosity. It shapes how companies build models, how governments think about AI exports, and who can actually afford to compete at the frontier.

Congressional scrutiny signals that AI distillation by foreign adversaries will likely become a formal export control issue, not just a civil IP matter between private companies. That is a significant shift. It would mean distillation techniques, or access to the models that enable them, could be regulated the way certain semiconductor exports already are.

There is also a broader tension inside the AI industry itself. Since distillation requires a pre-existing teacher model, it cannot directly be used to advance state-of-the-art model capabilities. However, it can speed up the proliferation of advanced AI capabilities, even from closed-source models. In other words, distillation does not push the frontier outward, but it does let more actors reach the frontier without paying to discover it themselves.

Some people think that democratises AI in a good way. Others think it eliminates any incentive to invest in the expensive foundational research in the first place. Without IP protection, why invest billions in research if competitors can copy your work for a fraction of the cost? That question does not have a clean answer yet, and the law has not caught up with it.

The policy debate is unresolved as of mid-2026. What is clear is that distillation, a dry-sounding academic technique invented a decade ago to make image classifiers run on slower hardware, now sits at the intersection of intellectual property law, national security, and the economics of the entire AI industry. That is quite a journey for a word that most people had never heard of twelve months ago.