The BaseResearchLessWrong
China AI Safety Regulation Monthly Monitor, Issue 1 (September 2026)
In the first issue of a monthly series, Sofia Yablonskaya reports that China's public AI filing lists show five Chinese AI labs' first models but no separate entries for their later general-purpose models, and reviews a non-binding safety framework, a Supreme People's Court opinion on AI liability and a US–China incident-channel agreement.
Why it matters
The same thing, for different readers.
What it means for you
Outsiders cannot see from China's public filing lists how newer, more capable models were reviewed, though the author stresses that does not show they went unreviewed.
For lawmakers
Framework 3.0 names loss-of-control risks but is non-binding, with no metrics or reporting duties; whether it hardens into practice, and what the US–China incident channel covers, is still open.
For researchers
Gives AI-governance researchers a compiled table of Chinese filings by lab and a fixed tracker (disclosures, citations of Framework 3.0, court rulings, compute monitoring) to compare month to month.
The question it raises
How should responsibility for AI harm be shared as systems grow more autonomous? The court opinion ties it to autonomy and risk disclosure; the author calls the reading unclear.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“The published lists contain no separate entries for successive general-purpose model generations at the five frontier AI labs examined here. This finding is primarily about the lack of public visibility, but it cannot be inferred that the new models were not reviewed.”
Found word for word in the source
“The framework describes the possibility that models could strategically lower their performance in evaluations, conceal actions, or exploit flaws to break isolation restrictions and access external systems.”
Found word for word in the source
“For frontier AI governance, recognition of the risks is distinct from evidence that labs are required to manage them. Framework 3.0 provides the former. Whether its provisions acquire practical force through regulators, standards, courts or company practice is the question for the implementation tracker.”
Found word for word in the source
“The opinion focuses on preventive measures and technical feasibility, and also provides conditional protection for some open-source code providers, but it does not itself establish duties related to the loss-of-control risks mentioned in Framework 3.0.”
Found word for word in the source
“For the incident channel to be useful, the next round in November would need to clarify terms such as “incident” and “frontier model”: Will the channel cover internal models, or only commercially released “services”, to which most Chinese AI regulation is currently limited?”
Found word for word in the source
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- One author compiled all 31 public filing lists of China's Cyberspace Administration (13 generative-AI lists with 1,843 entries, covering approvals to 31 August 2026, and 18 deep-synthesis lists with 7,764 entries) and matched five Chinese AI labs (DeepSeek, Moonshot AI, Zhipu, MiniMax, StepFun) by legal entity against their public model releases. The rest of the issue reads official documents and announcements.
- What it found
- Each of the five labs appears in the generative-AI registry once or a few times, at launch or for a new product; none of their later general-purpose generations (for example DeepSeek V3, R1 and V4, Kimi K3, GLM-5.x) produced a new public entry. Only ModelBest, included for comparison, filed for each new model family. The author stresses this concerns public visibility and does not show that newer models went unreviewed.
- What it does not show
- The author calls both possible explanations (the term 'service' may cover all model versions; labs may tell the regulator of upgrades directly) hypothetical. Large firms (Alibaba, ByteDance, Baidu, Tencent) were left out, change notes in the deep-synthesis lists may record some upgrades, and the tracker sections are limited by publicly available information.
- Status
- A report
The record
- Questions
- Who can demand evidence, delay a release or stop an unsafe deployment? · Can AI systems act beyond what people intended or allowed?
- People
- Sofia Yablonskaya (Wrote it, Head of Legal at an AI company; incoming Winter Fellow at GovAI (2027)) · Xi Jinping (Mentioned) · Donald Trump (Mentioned)
- What we read
- We read the whole source.
- Found by
- The Radar, run of October 14, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “China AI Safety Regulation Monthly Monitor, Issue 1 (September 2026)”, October 13, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-13-china-ai-safety-regulation-monitor
Our notesFull LessWrong crosspost read; the original is on the author's Substack (sofiayab.substack.com), not read. Author bio is from the post: Head of Legal at an unnamed AI company, incoming GovAI Winter Fellow, and the views are stated to be the author's own. The byline is 'Sofia Yablonskaya'; a footnote citing the author's own 2020 guest post gives 'Safia Yablonskaia', and the contact address uses 'yablonskaia'. Section 2.2 says, citing Concordia AI, that the framework appears to engage with the OpenAI–Hugging Face incident in July; not linked to that story here. Filing counts, dates, the Framework 3.0 and court-opinion contents and the summit outcomes are the author's account of Chinese official sources; none were checked here. Section 5.3 also mentions Dario Amodei's call to slow down and a state-media-affiliated account on Anthropic; this entry does not use it, so conflict is false. Footnote numbers appear inline in the saved text, so quotes avoid them. No instructions to an AI found.