Distillation is a way to train one AI model on the answers of another. OpenAI says attackers’ success at pulling its models’ hidden reasoning fell from 54.0% to 0.2% after new safeguards, as we reported on 21 November.
This explainer covers what distillation is, why labs use it on their own models, why OpenAI calls some of it “adversarial”, what it says it does about it and what is still unproven.
We track what AI models can do, and how they compare, on our AI model benchmarks page.
Business Pill: The Model and the Harness
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss distillation or OpenAI’s post.
The key insight: Distillation trains a model on another model’s answers; OpenAI says its updated safeguards cut reasoning-extraction success from 54.0% to 0.2% in its own tests and that its weekly enforcement against distillation campaigns ranges from thousands to tens of thousands of accounts.
What Is Distillation?
In a 2015 paper, Geoffrey Hinton, Oriol Vinyals and Jeff Dean described distilling the knowledge of several large models into a single model that is “much easier to deploy”.
The idea is simple. A big model, the teacher, produces answers. A second model, the student, is trained to copy them. OpenAI’s own definition: “one model learns from data produced by another”, from its answers, its reasoning or its judgments about which answers are better.

Why Do Labs Distil Their Own Models?
Smaller models are cheaper and faster to run. Distillation lets a lab pass some of a big model’s skill to a small one.
DeepSeek’s R1 paper is one example: it says the reasoning patterns of its large models “can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models”.
What Is Adversarial Distillation?
The same technique becomes a problem when one company does it to a rival without permission. OpenAI calls this adversarial distillation: an actor “may use a frontier model’s outputs without authorization to train another model with similar capabilities”.
OpenAI’s safety argument is that the copy can learn the skills without the guardrails. A model trained this way could help with cyberattacks “that the original model would have refused to assist with”, it says.
Why Hidden Reasoning Matters
Reasoning models think in steps before they answer. OpenAI calls those steps protected reasoning and says they “can contain information the final answer leaves out, making them especially valuable training data for another model”.
In an internal study, after supervised fine-tuning, a student trained on the exact reasoning generally outperformed one trained on what users normally see (final answers, reasoning summaries and tool interactions); after a further round of reinforcement learning the results were mixed, OpenAI says. That is why it now tries to keep the full reasoning private.
What OpenAI Says It Does
OpenAI says it trains its models to refuse extraction attempts, uses classifiers to catch leaks, and ties reasoning to the conversation that produced it. In its red-team tests, extracting reasoning from an earlier request succeeded 54.0% of the time on GPT-6 Astra and 0.2% on GPT-6.2 Astra, 270 times lower by our arithmetic.
It also says its weekly enforcement actions against distillation campaigns “have ranged in scale from thousands to tens of thousands of accounts”, with the rate rising this past summer.
OpenAI says it briefed Congress in February on activity indicating that DeepSeek “was attempting to distill U.S. frontier models”, and in September attributed a core part of one campaign to individuals associated with Moonshot AI, the developer of Kimi. Those are OpenAI’s claims; we have not seen public responses from DeepSeek or Moonshot to this post.
The Open-Model Angle
The debate reaches US companies that use Chinese open models. Cursor has told lawmakers it is considering phasing out Chinese open-weight models, and Reuters reported that the administration is weighing a disclosure rule for companies that use them.
A separate test this week suggested requests sent through some AI routers reached other companies: test links sent via OpenRouter were later opened, as we reported. That is a data-handling question, not a distillation finding.
The Structural Read
As we read it, the fight has moved from final answers to the reasoning behind them, because that is the part a rival can learn most from.
As we read it, the evidence so far is OpenAI’s own: the numbers are its tests and its enforcement counts, and the named cases rest on its attributions.
OpenAI, ‘Combatting adversarial distillation’
“Through adversarial distillation, a second model can learn advanced frontier capabilities without inheriting the safeguards that govern how the original model can use them.”
Three Implications
FOR AI USERS OpenAI calls distillation common; its objection is unauthorized use of its outputs.
FOR AI LABS OpenAI now keeps its models’ full reasoning private, it says.
FOR POLICYMAKERS OpenAI says it supports laws to sanction foreign actors that extract capabilities.
What Is Not Established
OpenAI’s figures come from its own tests and its own enforcement. It has not published account-level evidence in this post, and it says it plans to “share more soon about recent campaigns”.
OpenAI itself calls distillation a common machine-learning technique; its objection is to using a frontier model’s outputs without authorization. Whether that is happening at scale, and by whom, rests for now on OpenAI’s account.
The Bottom Line
Distillation, described by Hinton, Vinyals and Dean in 2015, trains a small model on a big one’s answers. OpenAI says it now blocks most attempts to copy its models’ hidden reasoning, with success falling from 54.0% to 0.2% in its tests, and that it acts against thousands to tens of thousands of accounts a week.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The definition, the safety argument, the 54.0% and 0.2% rates, the enforcement range and the DeepSeek and Moonshot attributions are from OpenAI’s 20 November post and are OpenAI’s claims. The origin of the term is the 2015 Hinton, Vinyals and Dean paper; the example of a lab distilling its own models is from DeepSeek’s R1 paper. The 270 times comparison is our arithmetic.
Sources: OpenAI, ‘Combatting adversarial distillation’, 20 Nov 2026 · Hinton, Vinyals and Dean, ‘Distilling the Knowledge in a Neural Network’, arXiv:1503.02531, 9 Mar 2015 · DeepSeek-AI, ‘DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning’, arXiv:2501.12948, 22 Jan 2025
Frequently Asked Questions
What is AI model distillation?
Distillation trains one AI model, the student, on the outputs of another, the teacher. The idea was described by Geoffrey Hinton, Oriol Vinyals and Jeff Dean in a 2015 paper on distilling knowledge into a single model that is easier to deploy.
Is all distillation a problem?
No. OpenAI calls distillation a common machine-learning technique, and labs use it on their own models; DeepSeek's R1 paper says its large models' reasoning patterns can guide smaller models. OpenAI's objection is to adversarial distillation: using another company's model outputs without authorization to train a rival model.
What does OpenAI say it does about it?
OpenAI says it trains models to refuse extraction attempts, uses classifiers to catch leaks and ties reasoning to the conversation that produced it. In its own tests, extraction success fell from 54.0% to 0.2%, and its weekly enforcement actions have ranged from thousands to tens of thousands of accounts.
Which companies has OpenAI named?
OpenAI says it briefed Congress in February on activity indicating DeepSeek was attempting to distill US frontier models, and in September attributed a core part of one campaign to individuals associated with Moonshot AI. These are OpenAI's claims.









