Short answer: On January 28, 2025, David Sacks alleged that DeepSeek had “distilled the knowledge” from OpenAI models. OpenAI separately said it had seen evidence of distillation attempts by China-based groups, while Microsoft was reportedly investigating suspicious API activity potentially linked to DeepSeek. But the public record described in those reports did not include API logs, the alleged training dataset, or an independent technical finding proving that DeepSeek trained its R1 model on OpenAI outputs.
What David Sacks claimed
Sacks made the claim during a Fox News interview on January 28, 2025. At the time, he was serving as the Trump administration’s AI and cryptocurrency adviser, a position often described in news coverage as the administration’s “AI czar.”
As an Amazon Associate I earn from qualifying purchases.
He said there was “substantial evidence” that DeepSeek had “distilled the knowledge” from OpenAI’s models. Sacks explained distillation as a process in which a smaller or newer AI model repeatedly queries a more capable model and learns from the answers. He said that, if this occurred, it could amount to intellectual-property theft, but he did not publicly present the underlying evidence during the interview.
That distinction matters. Sacks’s statement was a serious allegation by a senior government adviser, not a publicly documented court finding or a released forensic report.
#1 Best Overall
TechCrunch’s contemporaneous report, the Associated Press, and The Guardian all reported the allegation and its context.
What model distillation means
In machine learning, distillation is a way of transferring useful behavior from a larger “teacher” model to a smaller “student” model.
The teacher may generate:
- Answers to questions
- Step-by-step explanations or reasoning traces
- Rankings of alternative answers
- Demonstrations of how to solve particular tasks
- Specialized examples for coding, mathematics, writing, or other capabilities
The student is then trained or fine-tuned on those outputs. It may learn some of the teacher’s response patterns, problem-solving behavior, or task-specific capabilities without receiving the teacher’s model weights or source code.
Distillation is a routine machine-learning technique. It is not inherently illegal or unethical. Its permissibility depends on factors such as the provider’s contract, the model’s license, how the outputs were obtained, what material the outputs contain, and the law that applies.
A simple example is a small model learning from thousands of carefully generated mathematics solutions produced by a larger model. The smaller model is not necessarily a copy of the larger one, but the generated examples can make it better at similar problems.
Why OpenAI regarded the alleged conduct as prohibited
OpenAI said that using its model outputs to develop competing models would violate its terms of service. In practical terms, that would present a contractual issue between OpenAI and the account holder or organization that accessed the service.
That is separate from several other legal questions:
- Copyright: Whether particular outputs contain protected expression, whether copying occurred, and whether a defense or license applies.
- Trade secrets: Whether confidential information or protected secrets were acquired or used improperly.
- Unauthorized access: Whether someone circumvented access controls or violated applicable computer-access laws.
- Model imitation: Whether similar behavior reflects unlawful copying rather than common training data, ordinary benchmarking, or independent development.
Consequently, “DeepSeek used OpenAI outputs” and “DeepSeek committed copyright infringement” would not be interchangeable conclusions. Even a proven terms-of-service violation would not automatically resolve every copyright or trade-secret question.
OpenAI’s public position, as reported by Axios, was that it had seen evidence of China-based companies attempting to replicate advanced U.S. models through distillation. OpenAI said it was taking countermeasures and working with the U.S. government. The publicly reported statement, however, did not release a complete case file proving that DeepSeek itself had used OpenAI outputs to train R1.
What Microsoft was reportedly investigating
Reporting in January 2025 said Microsoft security researchers had observed people believed to be associated with DeepSeek obtaining substantial quantities of data through the OpenAI API. Microsoft and OpenAI were reportedly examining whether the activity involved unauthorized model distillation.
The reports described an investigation, not a final public determination. The individuals were not publicly identified, and the alleged connection to DeepSeek was not fully documented for outside review.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Large-scale API use can be suspicious, especially when it involves unusual volumes or patterns. But API activity alone does not establish:
- Who controlled the accounts
- Whether the outputs were collected for training, testing, or another purpose
- Whether the data reached DeepSeek’s model-development team
- Whether the data was used to train R1
- Whether any conduct violated a law rather than only a contract or platform rule
Bloomberg reporting carried by BNN Bloomberg and The Economic Times described the reported inquiry and potential account restrictions, but not a publicly adjudicated finding that DeepSeek used the data to train a named model.
What DeepSeek says about R1
DeepSeek’s public documentation says that R1 was built on the DeepSeek-V3-Base model and that its development involved reinforcement learning and supervised fine-tuning stages.
DeepSeek also openly documents a distillation pipeline. Its R1 repository describes smaller “R1-Distill” models based on Qwen and Llama checkpoints. The documentation says those models were fine-tuned using reasoning data generated or curated from DeepSeek-R1.
That is an important fact, but it is not an admission of the allegation made by Sacks. DeepSeek’s documented pathway is:
DeepSeek-R1 → smaller Qwen- or Llama-based models
The disputed allegation is:
OpenAI model outputs → DeepSeek model training
Those are different teacher-student relationships. DeepSeek’s admission that it used distillation does not identify OpenAI as the teacher model.
The cited R1 materials also do not say that OpenAI outputs were used to train R1. That is not conclusive proof that no such outputs were used; it is simply the public training account DeepSeek supplied.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
DeepSeek’s V3 documentation reports 14.8 trillion pre-training tokens and 2.664 million H800 GPU-hours. Such figures describe specified training work, not necessarily every research, engineering, experimentation, data, infrastructure, salary, or earlier-model cost. Low or comparatively efficient reported costs therefore cannot prove that another company’s outputs were used.
What evidence would establish the allegation?
The phrase “substantial evidence” does not tell readers what type of evidence exists or how independently it can be checked. The strongest evidence would include some combination of:
Rank #4
- Direct technical records: API logs, account identities, prompts, timestamps, output volumes, or a dataset showing that OpenAI responses entered training.
- Training documentation: Data-mixture records, internal pipelines, or reproducible evidence linking specific outputs to a model’s training.
- Independent forensic analysis: A technical report demonstrating distinctive, non-public similarities that are unlikely to result from common source material or ordinary model behavior.
- First-party admissions: A statement from DeepSeek or another directly involved organization.
- Legal or regulatory findings: A court, regulator, or independent investigator establishing what happened and whether it violated applicable law.
Public reporting in the dossier did not disclose a complete package of that kind. In particular, it did not make public the alleged API logs, account identities, prompt volumes, training dataset, model-comparison methodology, or a final independent finding linking specific OpenAI outputs to R1.
Why similar performance is not proof
DeepSeek’s January 2025 release of R1 drew intense attention because it appeared to offer competitive reasoning performance with substantially lower reported development costs than leading U.S. systems. That market reaction increased scrutiny of how DeepSeek achieved its results.
Recommended Free Tools
But capability similarity is not the same as provenance. Two models can produce similar answers because they:
- Were trained on overlapping public information
- Use similar model architectures or reinforcement-learning methods
- Were evaluated on widely circulated benchmarks
- Learned common solutions to common problems
- Were independently optimized for the same tasks
Likewise, a company might query a model for benchmarking, product testing, or prompt development without using the resulting data as training material. A third party could also collect outputs and provide them to a developer later. These possibilities make account activity relevant, but not automatically conclusive.
The expression “OpenAI models” is also imprecise. It could refer to an API-accessible model, a ChatGPT product, or a particular model family. Establishing what was allegedly used would require identifying the model, the outputs, the collection method, and the role those outputs played in training.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What OpenAI’s later memo adds
In a February 12, 2026 memo to the U.S. House Select Committee on Strategic Competition with the Chinese Communist Party, OpenAI continued to discuss DeepSeek in the context of distillation and model replication. The memo shows that OpenAI maintained the allegation as a strategic and national-security concern.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallIt remains an account from an interested party, however. The memo does not by itself substitute for independently disclosed logs, training records, forensic analysis, or a formal adjudication. It is evidence of OpenAI’s continuing position, not proof that resolves the underlying factual dispute.
The memo is available as a PDF from OpenAI.
Why the dispute matters beyond DeepSeek
AI-company economics
Frontier models require substantial computing, data, research, and engineering investment. If a company can cheaply collect outputs from a powerful model and use them to improve a competing system, the original provider may lose part of the economic value of that investment. The dispute therefore concerns not only national competition but also how AI companies protect the products they sell through APIs.
Closed APIs versus open weights
Closed APIs restrict direct access to model weights and can give providers visibility into usage patterns. That may help detect systematic extraction, although it does not make collection impossible. Open-weight models are easier to inspect, modify, and reuse, but their openness also makes downstream distillation and derivative development easier.
Licensing and access rules
The controversy could influence how model licenses define permitted training, whether API terms expressly address competitive distillation, how providers impose volume or account controls, and what disclosures companies provide when they suspend access.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →U.S.–China technology policy
The allegation also became part of a wider debate over whether restrictions on advanced chips and computing infrastructure are sufficient when model capabilities can potentially be transferred through outputs. Policymakers may consider whether model distillation should be addressed separately from semiconductor export controls, and how to distinguish legitimate research from unauthorized extraction.
What is established—and what is not
| Question | Publicly supported answer |
|---|---|
| Did David Sacks make the allegation? | Yes. He made the claim on Fox News on January 28, 2025. |
| Did OpenAI report seeing distillation attempts? | Yes. OpenAI said it had seen evidence involving China-based groups. |
| Was Microsoft reportedly investigating related API activity? | Yes. Reports described an investigation into suspicious activity potentially linked to DeepSeek. |
| Did DeepSeek document using distillation? | Yes. It described using R1-generated reasoning data to fine-tune smaller Qwen- and Llama-based models. |
| Did DeepSeek publicly admit distilling OpenAI models? | Not in the cited documentation. |
| Did the public reporting establish that R1 was trained on OpenAI outputs? | Not on the publicly disclosed record described here. |
| Has distillation itself been shown to be illegal? | No. Its legality depends on the contract, license, data, conduct, and applicable law. |
Final assessment
Sacks’s statement was a consequential allegation, and it was followed by OpenAI’s claims and reports of suspicious API activity that made the issue more than mere online speculation. DeepSeek also openly used distillation in its own documented work, but its published materials describe distilling R1 into smaller models—not distilling OpenAI models into R1.
On the public record covered here, the most accurate conclusion is narrower than “DeepSeek stole ChatGPT.” Sacks and OpenAI alleged or strongly suggested that DeepSeek-linked actors may have used OpenAI outputs for distillation. Public reports described an investigation and circumstantial evidence, but did not disclose a complete, independently verifiable finding that DeepSeek trained R1 on OpenAI’s outputs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors
