A patent examiner in Beijing opens a file for an invention that was conceived, trained, and tuned largely by software. The applicant wants the algorithm listed as a co-inventor. The question is no longer hypothetical for Chinese filers.
In late 2025 the State Intellectual Property Office rewrote part of its rulebook for how such applications are judged. The change took effect at the start of 2026 and is now shaping how examiners treat machine-learning and big-data claims.
Key takeaways
- China's patent guide amendment (Order 84) took effect on 1 January 2026, the first to theme a section on AI and data.
- Inventors must be natural persons; an AI system cannot be named as an inventor on a Chinese patent.
- Applications touching AI must meet Article 5(1): lawful data collection, labeling, and rule-setting under public-interest tests.
- Model construction and training must be disclosed in enough detail to address the "black box" problem.
- Inventive-step review now accounts for how algorithms improve user experience across specific scenarios.
A rulebook catches up with the models
The amendment to the Patent Examination Guide (专利审查指南) was promulgated on 10 November 2025 as State Intellectual Property Office Order No. 84, after an 18 September 2025 bureau meeting, and came into force on 1 January 2026 (gov.cn; State IP Office Order No. 84). For the first time, the guide organizes a section around the theme of "artificial intelligence (人工智能) and big data" rather than treating them as edge cases.
The move did not appear from nowhere. China's patent office had already adjusted the guide for algorithm-related claims in 2019 and for AI and big-data patentable subject matter in 2023 (CNIPA, November 2025 briefing). The 2025 revision is the third pass, and the most explicit: it tells examiners what to do when a claim leans on a model rather than a mechanical device, and it does so in a standalone chapter rather than a footnote.
The inventor cannot be a machine
One clause is blunt. The guide now states that an inventor must be a natural person, and that the request form may not list a unit, a collective, or an AI name such as "artificial intelligence X" (State IP Office Order No. 84). This settles a debate that has surfaced in courts elsewhere: a model can contribute to an invention, but it cannot hold one.
The practical effect is narrow but symbolically clear. Teams that build with generative AI (生成式人工智能) still file in the names of the humans who made the creative contribution. What changes is the surrounding scrutiny of how those humans describe what the machine did, and whether the human contribution is real enough to support the claim.
What applicants must now show
The revision adds a review standard drawn from Article 5(1) of the Patent Law. For AI or big-data applications, if data collection, label management, rule-setting, or recommendation decisions contain content that violates law, social morality, or public interest, the invention is unpatentable (CNIPA, 4 December 2025 interpretation). Two worked examples were added: a mall mattress-sales system that secretly filmed customers, and an autonomous-vehicle emergency-decision model.
The guide also tackles the "black box" problem directly. Where a model's logic is hidden between input and output, the specification must now explain model construction and training well enough that the technical solution is fully disclosed (CNIPA interpretation). Examiners may look beyond the claims to the description when needed, a shift from the older practice of judging mainly the claimed text.
Applicants should expect to document:
- how training data was collected and labeled, and whether consent or law was respected;
- why a model architecture or training step is technically necessary, not merely effective;
- how the algorithm improves a concrete scenario rather than "intelligence" in the abstract.
A signal for AI firms, not a crackdown
The creative-step rules were also refined. Examiners are told to consider how AI contributes to user experience in specific applications, using examples rather than a blanket test (CNIPA). The office has separately published a guidance document on AI-related invention patent applications, which sits alongside the guide as a softer, interpretive companion rather than a binding rule (CNIPA).
Read together, the changes point less to restriction than to normalization. China is treating AI as a routine, if sensitive, field of invention — one that needs clear disclosure and an ethical floor, but that remains firmly patentable when the human contribution is real. For foreign filers using the Paris Convention or the Patent Cooperation Treaty route into China, the same examination text now applies, so the disclosure bar is no longer a grey area.
Honest limitations
The amendment is in effect as binding examination practice, but it interprets existing patent law rather than creating new rights or bans. The added ethical floor under Article 5(1) is framed through examples and remains subject to examiner judgement, so outcomes will vary case by case. It is unclear how aggressively the "black box" disclosure rule will be applied to trade-secret-protected model weights. The separate AI patent guidance document is interpretive, not a statute, and the 2019 and 2023 changes remain part of the same evolving guide.
What readers can do now
- Patent applicants using AI in invention: name only human inventors and document data sourcing and training steps now.
- IP counsel: review pending AI or big-data claims for Article 5(1) exposure before substantive examination begins.
- Researchers: treat model architecture and training disclosure as part of the publication record, not only the paper.