Product1 publisher3 min readPublished
JoyIn says it has started suing OpenAI over the distillation of its Aether model
American labs have spent months accusing Chinese developers of illicit distillation. A Chinese robotics startup has now made the same charge against OpenAI. For buyers, the question is whose model answered the call.
The Product Desk · Product desk
What happened
- Chinese robotics startup JoyIn, developer of a model called Aether, accused OpenAI on Thursday of distilling its systems and copying its website's cosmic design for the GPT-6 Astra launch and branding.
- In an open letter written in Chinese, JoyIn chief executive Guo Renjie said the company had started the process of filing a lawsuit, in a translation Gizmodo credited to CNBC.
- CISA said on Tuesday that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai are extracting US model capabilities through industrial-scale knowledge distillation campaigns.
- Anthropic reported on Thursday that Moonshot's Kimi customers were shown responses generated by Claude while believing the text came from Kimi.
- Gizmodo said Guo's letter is the first known case of a Chinese developer accusing an American company of distillation.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Any product built on one of the six firms CISA named now has a US government allegation attached to a supplier, and support teams will field questions about whose model answered.
- decision Renewal talks now turn on whether a vendor will name the serving model in the contract and log it per request. Asking costs a buyer nothing, and a refusal tells the buyer something too.
- contradiction OpenAI presses distillation claims against Chinese rivals while defending publisher suits by invoking fair use.
- precedent The charge now runs from a Chinese startup toward an American lab. Indemnity clauses have to contemplate plaintiffs in either country.
The place this lands is a vendor review. The buyer has a line item for model access at a fraction of frontier prices and a product that passes those tokens to paying customers, and the dashboard has no field that proves which model wrote them. Anthropic's account of Kimi, if it holds, describes that gap being used: paying customers who believed they were buying one model's output and got another's [7].
CISA's statement, published Tuesday, names six firms [15]. The agency said they are conducting "systematic extraction of proprietary functionalities and capabilities of U.S. AI companies' models through industrial-scale knowledge distillation campaigns" [5], and said this was likely being carried out with the knowledge and blessing of the Chinese government [6].
Distillation itself is routine. A larger teacher model helps train a smaller student, and developers do it constantly [4]. It becomes an accusation only where a contract or a terms-of-service clause says so. The same practice therefore supports a US agency statement in one direction and an open letter in the other. Gizmodo said it had not independently verified the JoyIn report [3], and the account does not include a court or a damages figure.
OpenAI has been here often. Artists and news publishers have sued over training data since ChatGPT launched a little under four years ago, and the company has invoked fair use [9]. Apple sued in July, claiming OpenAI stole company secrets to get its hardware business off the ground, and OpenAI denied wrongdoing [10]. After an NYU mathematics professor asked this week whether his Codex work had been used to train the model OpenAI used on the Navier-Stokes Millennium Problem [11], the company told the Australian Broadcasting Corporation it "can say categorically that it is impossible for Dr. Buckmaster's Codex prompts over the last two months to have influenced the system in any way, including training" [12]. Both OpenAI and Anthropic are preparing IPOs that are expected to be record-breaking [13].
For a buyer, one invoice hides two questions. The first is whether the model named on the invoice is the model answering the call, and a vendor can settle that with a contract term and a logged model identifier per request. The second is whether that model was trained on a competitor's output. A customer has no way to check that from the outside, so the answer sits in whether the vendor's indemnity covers the claim.
Teams that adopt a cheap API tend to treat it as a pricing decision, because price is the number they can see. Provenance and indemnity decide who pays when the owner of the teacher model sues the owner of the student, and US labs including OpenAI and Anthropic have been making that claim repeatedly [14].
What to watch
- Whether JoyIn files in an identifiable court, and what damages it claims; the open letter is the only document in the record so far.
- Whether Moonshot, Alibaba or the other firms CISA named respond, and whether Moonshot disputes Anthropic's account of Kimi serving Claude output.
- Whether model provenance language appears in API terms, or in the risk factors OpenAI and Anthropic have to write for their IPOs.