This article assembles the whole sequence in one place. Each entry links to our fuller treatment, and each is sourced there.
Background: 2025
The pattern was set when DeepSeek released a model competitive with US frontier systems at implausibly low reported cost, and OpenAI suggested its outputs had been used as training data. The dispute was conducted publicly and produced no resolution, but it established the template every subsequent episode followed.
February 2026: it becomes a lobbying position
OpenAI took the claim to US lawmakers, describing efforts to free-ride on capabilities developed by US frontier labs and saying it had detected new obfuscated methods designed to evade its defences.
Twelve days later Anthropic named three laboratories: DeepSeek, Moonshot AI and MiniMax, citing more than 16 million API exchanges through over 24,000 fraudulent accounts. This was the first time hard numbers entered the argument, and they were observations of the provider's own infrastructure rather than inferences about anyone's model.
March 2026: policy looks for a lever
Proposals converged on Entity List designation as the most practical instrument, while analysis noted that export controls do not fit a transfer with no item in it.
April to June 2026: escalation
A congressional hearing in April moved the framing to national competitiveness.
In June, Anthropic wrote to the Senate Banking Committee alleging that operators connected to Alibaba's Qwen lab had run roughly 28.8 million exchanges through around 25,000 fraudulent accounts between 22 April and 5 June. These figures are frequently conflated with February's; they are a separate disclosure about different parties.
July 2026: the government joins
Moonshot released Kimi K3 on 16 July, fifteen days after Anthropic's Fable 5 returned to public availability.
On 22 July the Director of the White House Office of Science and Technology Policy stated that Moonshot had distilled Fable to build K3, and separately alleged access to restricted Nvidia hardware via Thailand. The Treasury Secretary raised sanctions and Entity List restrictions.
Objections followed immediately. The fifteen-day window is difficult to reconcile with distilling a frontier model, and neither Anthropic nor the White House published evidence directly connecting the alleged activity to Fable 5 or to K3. Researchers characterised the claims as political, noting that model outputs are not copyrighted.
On 26 July Moonshot published the K3 weights openly, placing the disputed model permanently beyond recall.
The same week, in the other direction
On 21 July, one day before the White House statement, a federal judge gave final approval to the largest copyright settlement in US history, $1.5 billion, against Anthropic, for downloading and retaining seven million pirated books. The court held that training on the books was fair use; liability attached to acquisition.
What was actually established
- Established: large-scale querying occurred, evidenced by provider infrastructure logs
- Established: fraudulent account creation at scale, which is the strongest legal hook available
- Not established: that any specific released model was built from that activity
- Not established: any legal violation, by any tribunal
- Not imposed: any enforcement action whatsoever
What the episode demonstrates
Three things, none of which required the accusations to be true or false.
Attribution is unsolved, and every technical claim ran into it. See The attribution problem.
The legal foundation is thinner than the rhetoric. Outputs are not copyrighted, so the claims rest on contract and fraud, which do not naturally support sanctions.
The structural concern is real regardless. Capability transfer without compute is a genuine gap in the control regime, and it is the one part of the argument that survives every objection to the specific accusations.
Common questions
What are the AI distillation accusations of 2026?
A sequence of allegations by US frontier labs and later the US government that Chinese laboratories extracted capability from American models by querying them at scale. The principal disclosures involved DeepSeek, Moonshot AI, MiniMax and Alibaba's Qwen lab, and escalated to threatened sanctions in July 2026.
Was anything ever proven?
No enforcement action has been imposed, and no evidence directly connecting alleged extraction activity to any specific released model has been published. The strongest material produced was infrastructure evidence showing that large-scale querying occurred.