You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
What this project is
China currently runs the world's largest mandatory AI disclosure system. The mechanism is that every generative-AI service must (1) submit information about themselves to the regulator, (2) pass a security assessment against binding national standards (right now there are three GB/T standards focused on general safety, data pre-training and fine-tuning, and data annotation), and (3) appear in a public registry (算法备案), published in monthly batches of approximately 500 to 800 filings.
Right now very few people outside China read these batches. The English-language coverage (Carnegie, Concordia AI, Stanford DigiChina) and influential newsletters describe the mechanism, but they don't analyse the filings' content as data.
Where I stand out is that I read this material in the original. I have already analysed the January, March and May 2026 batches and the GB/T security standards for a comparative EU-China analysis, funded by a BlueDot Rapid Grant (publication August to September 2026).
For this project, I am turning that one-time reading of the filings' content into a regular ongoing public monitor.
The specific monthly output will be a public English-language brief that (1) characterises the new batch: the providers' names, which frontier providers appear (the more recent filings include ByteDance/Doubao, Xiaohongshu, Inspur; earlier ones DeepSeek, Alibaba/Qwen, Zhipu, Baidu, Tencent), and what the filings disclose against what was filed but not disclosed by the regulators; (2) tracks changes in the "binding in practice" technical standards and the draft national AI Law; (3) scores the trajectory against a frontier-safety transparency framework (built from Cotra, Heim and Anderljung during the BlueDot AI Governance course): is China trending toward more or less visibility of dangerous capabilities? Alongside the briefs, I will publish a structured English dataset of frontier-relevant filings (provider, model, role, products, use description, filing number), so that researchers in the field who do not read Chinese can work with the registry.
Publication: on LessWrong and my own webpage/ Substack, distributed through my EU policy network (which includes people who have worked with the EU AI Office and researchers at The Future Society, FLI and Concordia AI, several of whom are currently giving feedback on the comparative article I am writing). I will also offer the briefs to existing China-AI outlets.
Theory of impact
Two mechanisms.
First, calibration. The EU's GPAI regime is being interpreted right now: Commission enforcement and penalty powers start from August 2026, the Code of Practice is new, and the systemic-risk compute threshold is under reconsideration after DeepSeek. How strictly the EU calibrates depends partly on what policymakers believe other frontier jurisdictions require. Today that belief is formed from English summaries of formal text, which miss what China's regime actually demands and discloses in practice. Miscalibration is an x-risk problem in both directions: too little regulation tolerates dangerous opacity, too much pushes frontier development to more permissive jurisdictions and the global safety floor is set by whoever is most permissive. A reliable, primary-source picture of the Chinese regime's operational reality lowers the chance of a high-stakes calibration error during the exact window when the error would be made.
Second, early warning. The frontier-governance literature (Cotra's transparency proposals in particular) argues society needs visibility of capability trajectories to see danger coming. China's registry is one of the few places where frontier-adjacent Chinese AI activity leaves a public, monthly, mandatory trace. Nobody is currently watching it systematically in English. A monitor that flags, for example, new frontier-lab filings, changes in what the standards require to be tested, or catastrophic-risk language moving through binding documents (the GB/T standards already instruct providers to treat with caution AI that may deceive humans, self-replicate or self-modify) gives the safety community a low-cost sensor it does not have.
Full proposal: https://app.grantmaking.ai/projects/e6cb7060-5a40-4f8f-a497-51a43ec77a61
The goal is that, twelve months from now, an EU policymaker or a safety researcher who does not read Chinese can answer three questions from public English sources: who is filing with China's registry, what those filings actually disclose, and whether the trend is toward or away from visibility of dangerous capabilities. Today none of those can be answered without reading Chinese.
Scope of what I read each month. The monitor covers both filing tracks: the deep-synthesis algorithm filings (境内深度合成服务算法备案清单) and the generative AI service filings (境内生成式人工智能服务备案信息). The second track carries most of the weight for this project, because those are the services assessed against the GB/T security standards under the Interim Measures, so on that list the gap between a detailed mandatory assessment and a thin public disclosure is actually visible. The dataset marks which track each entry comes from, since the two regimes are legally distinct and conflating them would be an error.
How I achieve it. The method is already built and tested : I read the filings and the standards in the original, with a pipeline handling most pre-processing of hundreds of short filings and all analysis and final translations done by me. I have run this on three batches already, so a monthly cycle is bounded, routine work.
Relationship to my BlueDot-funded article. The three batches I have already analysed (January, March and May 2026) were paid for by the BlueDot Rapid Grant and are not billed again here. They are the baseline the monitor measures against. The twelve cycles funded by this grant cover new batches from the start date forward, plus ongoing standards and draft-law tracking.
Distribution - which is where projects like this usually fail: publication on LessWrong and my own page / Substack, and distributing each brief through my EU policy network rather than relying on organic reach. I will also offer the briefs to existing China-AI outlets.
On adjusting scope. The committed deliverables are the twelve monthly briefs, the dataset and the two synthesis reports. Within that, I keep discretion over what each brief covers: if I find something in the filings, the standards or the draft law that is more useful to this audience than what I planned, I will follow it and say in the brief why. If a planned element turns out to be uninformative, I will say that rather than pad the output. Anything that would reduce the committed deliverables I would raise with you (grantmaking.ai) first.
$19,650 over twelve months.
$10,800: twelve monthly cycles at $900. Completion stipend for reading and analysis time. Each batch is 500 to 800 filings, plus standards and draft-law tracking. As a practising lawyer my binding constraint is protected time, and a committed public deliverable defends this work against billable hours.
$3,200: two synthesis reports (month 6 and month 12): trajectory analysis across batches, policymaker-facing.
$2,650: the structured public dataset of frontier-relevant filings, maintained monthly: schema design, translation quality control, publication.
$2,000: distribution. A simple public site for the monitor and the dataset, and time to route each brief to EU policy readers rather than relying on organic reach.
$1,000: gated access to the full texts of GB/T standards referenced normatively by the security standards, and API costs for the translation-assist and dataset structuring pipeline. All analysis and final translations are mine; the pipeline handles bulk pre-processing.
This is the twelve-month version of my proposal scaled to the amount awarded. I removed the contingency line and simplified the dataset, and kept all twelve cycles and both synthesis reports, because the value of a monitor is in the unbroken series.
Figures are gross; this is taxable to me as service income in Poland, so net amounts are about 15% lower.
The $19,650 minimum is the awarded amount and funds the full twelve months as described above. The $22,000 maximum is my original ask: additional funding restores the fuller public dataset and a contingency for unusually large batches.
Just me. I am a lawyer, currently Head of Legal at a US-based AI company whose product runs on neural-network image processing. I have a law degree from Fudan University (taught in Mandarin) and an MSc from LSE in comparative China policy. I read regulatory and technical Chinese at working pace.
The relevant track record is that the method here is proven and tested. I have worked through the January, March and May 2026 registry batches and the GB/T security standards in the original, and I am writing the resulting comparative EU-China analysis now under a BlueDot Impact Rapid Grant, publication due August to September 2026. The question of whether one person can actually read and characterise these batches and say something non-obvious about them has already been tested once.
I completed BlueDot Impact's AI Governance course in May 2026, where I developed the transparency-and-reporting framework this monitor scores against. In my legal work I run a Unified Patent Court case in Germany essentially solo, including the technical investigation, the evidence and the court documents, which is the closest evidence I can give that I ship long technical-legal work to deadline without a team behind me.
The most likely failure is time, not analysis. I am a practising lawyer with a live UPC case, and the risk is that the litigation eats two or three monthly cycles and the monitor becomes irregular. Irregular practically means useless here, because the value is in the regular series. Mitigation: the monthly unit is deliberately small and the reading method is already routine, so a cycle is bounded work. If I fall behind I will publish a combined two-month brief rather than skip one silently, and say so publicly.
Second, the registry could change: the publishing authority could alter the format, disclose less, stop publishing batches, or move them somewhere less accessible. Then part of the object disappears. That is itself a finding worth publishing, and the standards and draft-law tracking would continue, but the dataset would shrink.
Third, the batches could turn out to be less informative month over month than they look. The filings are short, and it is possible that after several cycles the marginal new information from a single batch is small. That would mean the monthly cadence is wrong for the registry component, and I would say so openly rather than dilute the briefs and fill them in just for the sake of filling them in. My response would be to shift the weight of each brief: less batch-by-batch characterisation, more on the rest of the Chinese regulatory picture (movement in the GB/T standards and the technical documents underneath them, the draft national AI Law, new measures and enforcement activity), with the registry appearing as processed cumulative data (trends across batches, cumulative provider counts, changes in what is disclosed over time) rather than as a monthly description. The dataset keeps accumulating either way, and the cumulative view is where its value lies.
Fourth, nobody reads it. This is the failure I can most directly work against, which is why distribution is part of this fund. If the briefs are published but unread, the money still produced a public dataset and an archive of primary-source analysis, which is real but much smaller than intended.
Worst realistic case: a partial series of briefs and a partial dataset exist publicly, and the residual value is that researchers who do not read Chinese can see what a systematic monthly reading of the registry looks like and decide whether to pick it up themselves.
$2,350: a Rapid Grant from BlueDot Impact, for the comparative EU-China GPAI transparency analysis that this project builds on. That is my only philanthropic funding so far. This grantmaking.ai grant of $19,650 would be the second.
My other income is from legal work (employment as Head of Legal, plus consulting). It is unrelated to this project, and it is the constraint that the stipend line is meant to buy time away from.