From Abstract Idea to Technical Solution

Executive Summary: Artificial-intelligence inventions are patentable in both the United States and China, but neither jurisdiction protects “AI” as an abstract label or rewards generic use of a known model in a new field. U.S. practice applies the Alice/Mayo eligibility framework, with current USPTO guidance emphasizing the claim as a whole, a specific improvement to technology, and evidence connecting that improvement to the claimed mechanism; properly supported declarations may assist during examination but cannot cure missing disclosure. China reaches many comparable results through its technical-solution requirement, mental-activity exclusion, inventive-step analysis, and heightened expectations for explaining how algorithmic features interact with technical features, while also imposing an express ethics and data-governance screen. Both systems require a qualifying natural-person inventor and sufficient disclosure of the architecture, training process, inputs, outputs, and causal mechanism producing the asserted technical result. The most durable cross-border strategy is therefore to document human conception, file before public disclosure, and claim the concrete mechanism that solves an identified technical problem rather than merely reciting an AI-generated result.

I.                   Introduction

“Is artificial intelligence patentable?” compresses several questions that United States and Chinese patent law answer through different doctrinal gates. Neither jurisdiction grants patents on “AI” as an abstraction. Each instead asks whether a claimed AI-related invention has a qualifying human inventor, falls within protectable subject matter, is novel and nonobvious or inventive, and is disclosed with enough technical detail to be reproduced. The comparison therefore must distinguish an innovation in an AI model or training technique from an invention that merely uses AI in a larger system, and both from an invention developed with AI assistance.

Consider a claim reciting a computer-implemented method that receives domain data, trains a machine-learning model, and uses the model to generate an optimized output. That formulation says almost nothing about the asserted technological advance. It does not identify what is different about the model, how the training process addresses a technical constraint, why the inputs and outputs have technical meaning, or what the output causes a machine or process to do. Recent United States decisions make such field-of-use claiming particularly vulnerable under 35 U.S.C. § 101. Chinese practice may reach a similar result through the exclusion of rules and methods for mental activities, the requirement for a “technical solution,” or inventive step.

The central contrast is methodological rather than simply pro- or anti-patent. In the United States, eligibility under § 101 remains a powerful threshold filter, and recent Federal Circuit decisions scrutinize whether a claim merely applies generic machine learning to a new data environment or instead improves model, computer, or other technological operation. China performs comparable screening through its technical-solution requirement and mental-activity exclusion, then asks at inventive step whether algorithmic features functionally support and interact with technical features. Its examination rules effective January 1, 2026, additionally make AI ethics, data acquisition, and detailed model disclosure more prominent parts of examination.

Yet the systems converge on a practical proposition: the strongest application identifies the technical bottleneck, claims the mechanism that solves it, and discloses enough of the architecture, training process, and input-output relationship to reproduce the asserted result. A label such as “artificial intelligence,” a new dataset, or a new industry is not a substitute for a qualifying natural-person inventor or a technically specific contribution.

II.                One Technology, Four Questions

“AI invention” can describe at least four different circumstances. First are inventions in AI itself: model architectures, continual-learning mechanisms, training and inference methods, compression, distributed execution, and specialized accelerators. Second are AI-enabled applications, in which a known or modified model is used in manufacturing, communications, transportation, finance, medicine, or another field. Third are AI-assisted inventions, in which a natural person uses an AI system as a research, design, or problem-solving tool. Fourth are outputs produced with AI for which the evidence may not establish that any natural person satisfies the applicable inventorship standard.

These categories engage different rules. The first two primarily raise subject-matter eligibility, novelty, nonobviousness or inventive step, and disclosure. The latter two make inventorship central, but AI use does not create a separate United States inventorship standard. Under United States law, the ordinary conception inquiry asks whether a natural person formed a definite and permanent idea of the complete and operative claimed invention. In China, a natural person must make a creative contribution to the invention’s substantive features. If no natural person satisfies the governing standard, naming the model owner, operator, prompt author, or another convenient person does not supply inventorship.

Eligibility also must be separated from patentability as a whole. A claim can pass § 101 and remain obvious, as the USPTO’s leading favorable AI eligibility decision illustrates. Conversely, a technically sophisticated idea can be novel but excluded because the claim defines only an abstract result. Disclosure is an independent gate: even an eligible and inventive model cannot support a claim broader than the specification enables or possesses.

III.             Two Legal Architectures

United States utility-patent law applies the same basic statutes to AI that it applies to other technologies. Section 101 defines statutory subject matter and is limited by the judicial exceptions for abstract ideas, laws of nature, and natural phenomena. Sections 102 and 103 address novelty and nonobviousness; § 112 governs written description, enablement, and definiteness; § 100(f) defines an inventor as the “individual” who invented or discovered the claimed subject matter; and § 115 governs the inventor’s oath or declaration. 35 U.S.C. §§ 100(f), 101–103, 112, 115; see Thaler v. Vidal, 43 F.4th 1207, 1211–13 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023).

Under Alice Corp. Pty. Ltd. v. CLS Bank International, a court first asks whether the claim is directed to a judicial exception and, if so, whether its elements, individually and as an ordered combination, supply an inventive concept sufficient to transform the claim. 573 U.S. 208, 216–18 (2014); see also Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66, 71–73 (2012).

The USPTO implements that framework through Step 2A, including separate inquiries into whether a claim recites a judicial exception and whether the claim as a whole integrates it into a practical application, followed when necessary by Step 2B. The July 2024 AI eligibility update elaborates those inquiries for AI-related claims and announces AI-focused Examples 47–49. 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence, 89 Fed. Reg. 58,128, 58,131–38 (July 17, 2024) [hereinafter 2024 AI SME Update]. It is examination policy rather than substantive rulemaking: it does not have the force of law, create an enforceable right, or bind an Article III court applying Supreme Court and Federal Circuit precedent. Id. at 58,131–32. Later examiner memoranda refine Office practice within that framework; they do not replace the governing judicial law.

China begins with a different statutory vocabulary. An invention must be a “new technical solution relating to a product, a process, or improvement thereof.” Patent Law of the People’s Republic of China (promulgated by the Standing Comm. Nat’l People’s Cong., Mar. 12, 1984, rev. Oct. 17, 2020, effective June 1, 2021), art. 2(2) (China) [hereinafter PRC Patent Law]. Article 25 excludes rules and methods for mental activities. Article 22 separately requires novelty, inventiveness, and practical applicability, while Article 26 requires an enabling specification and claims that are clear, concise, and supported. Id. arts. 22, 25(1)(2), 26(3)–(4).

CNIPA’s Patent Examination Guidelines instruct examiners not to sever algorithmic or business-rule features mechanically from technical features. The claim is evaluated as a whole by examining its technical means, the problem it solves, and the effect it obtains. China Nat’l Intell. Prop. Admin., Patent Examination Guidelines pt. II, ch. 9, §§ 6.1.2–6.1.4 (2023, as amended 2025) [hereinafter CNIPA Guidelines]. The 2025 amendments, effective January 1, 2026, strengthened that framework by adding express provisions concerning AI ethics, truthful natural-person inventorship, AI-specific sufficiency, and new examination examples. China Nat’l Intell. Prop. Admin., Decision Amending the Guidelines for Patent Examination, CNIPA Order No. 84, pts. I, VII (Nov. 10, 2025) (effective Jan. 1, 2026) [hereinafter CNIPA Order No. 84]. Those guidelines govern administrative examination; they should not be equated with legislation or binding judicial precedent.

Table 1. Core Comparison

Question	United States	China	Drafting Consequence
Threshold subject matter	Section 101 and the Alice/Mayo abstract-idea framework	Article 25 mental-activity exclusion plus Article 2(2) technical-solution requirement	Do not rely on an “AI” label or generic computer wrapper.
Known AI in a new field	Vulnerable when the claim merely applies established machine learning to a new data environment	May fail technical character or inventive step if the model and technical mechanism remain unchanged	Claim the field-driven technical adaptation, not merely the field.
Improvement to AI itself	Potentially eligible when the claim reflects a specific improvement to model or computer operation	Potentially technical and inventive when algorithmic and technical features interact to solve a technical problem	Put the improvement-producing mechanism in the claim and specification.
Inventorship	Ordinary natural-person conception; no separate AI standard, and AI remains a tool	Natural person making a creative contribution to substantive features	Preserve claim-specific evidence of human conception in the United States and creative contribution in China.
Disclosure	Section 112, calibrated to claim breadth and predictability	Article 26 plus express 2026 requirements for architecture, training, and input-output relationships	Draft the original priority specification to the stricter common denominator.
Ethics	No comparable general AI morality screen in patent examination	Article 5 examines illegality, social morality, and public interest	Audit data acquisition and automated decision rules before filing in China.

IV.            Patent-Eligible Subject Matter

A.                 The United States: From Generic Machine Learning to a Claimed Improvement

For AI claims, the principal abstract-idea groupings are mathematical concepts, mental processes, and certain methods of organizing human activity. The 2024 AI SME Update instructs examiners to identify the particular limitation said to recite an abstract idea and determine whether it falls within an enumerated grouping. A claim does not recite a mathematical concept merely because it is based on or involves mathematics, and a limitation that cannot practically be performed in the human mind does not fall within the mental-process grouping. 2024 AI SME Update, 89 Fed. Reg. at 58,134–36. The August 2025 examiner reminder reiterates those limits and distinguishes a claim that “recites” a judicial exception from one that merely “involves” one. U.S. Pat. & Trademark Off., Reminders on Evaluating Subject Matter Eligibility of Claims Under 35 U.S.C. § 101 1–3 (Aug. 4, 2025) [hereinafter 2025 Eligibility Reminder]. Thus, a limitation calling generally for training a neural network may involve mathematics without setting forth a mathematical relationship, while a limitation expressly requiring backpropagation and gradient descent recites mathematical calculations and proceeds to the remaining eligibility analysis.

That distinction is not a drafting safe harbor for generic functional claiming. The Office first applies the claim’s broadest reasonable interpretation, and eligibility does not displace novelty, nonobviousness, or disclosure. At Step 2A, Prong Two, examiners must evaluate the claim as a whole rather than isolate the exception from interacting limitations. A claim focused on a particular technological solution may integrate an exception into a practical application, while a desired result, generic computer implementation, or field-of-use label does not. The specification need not use the word “improvement,” and the claim need not recite that conclusion expressly, but the improvement must be apparent to a skilled person from the disclosure and reflected in the claim. 2024 AI SME Update, 89 Fed. Reg. at 58,136–38; 2025 Eligibility Reminder, supra, at 3–5; see Enfish, LLC v. Microsoft Corp., 822 F.3d 1327, 1335–36 (Fed. Cir. 2016); McRO, Inc. v. Bandai Namco Games America Inc., 837 F.3d 1299, 1313–16 (Fed. Cir. 2016).

Recentive Analytics, Inc. v. Fox Corp. supplies the clearest current appellate rule for generic AI applications. The patents applied machine learning to generate broadcast schedules and network maps. Their specifications permitted “any suitable machine learning technology,” and the claimed training, updating, and real-time adjustment were ordinary features of machine learning rather than a new technical implementation. The Federal Circuit held, as a question of first impression, that claims doing no more than applying established machine-learning methods to a new data environment were ineligible. 134 F.4th 1205, 1211–14 (Fed. Cir. 2025), cert. denied, 146 S. Ct. 891 (2025). The holding is important but narrow: it does not make every field-specific AI application abstract. It condemns a claim whose asserted advance is the field of use rather than an improvement to machine learning or another technology.

Two nonprecedential 2026 decisions illustrate the reach of Recentive. In Rensselaer Polytechnic Institute v. Amazon.com, Inc., the court treated known case-based reasoning applied to natural-language processing as generic AI used in a new environment. The claims used result-oriented language and generic databases without requiring a specific implementation that improved those functions. Nos. 2024-1725, 2024-1739, 2026 WL 506661, at *3–5 (Fed. Cir. Feb. 24, 2026) (nonprecedential). In Dental Monitoring SAS v. Align Technology, Inc., claims collected dental-arch images, used a “deep learning device” to analyze them, and presented results. Training the device on more than 1,000 dental images did not create a technological solution because domain-specific training data were incident to machine learning and the claim did not recite how the asserted improvement was achieved. No. 2024-2270, slip op. at 9–13 (Fed. Cir. July 7, 2026) (nonprecedential). These decisions are illustrative applications of Recentive, not independent binding rules.

The USPTO’s precedential decision in Ex parte Desjardins provides the useful contrast. The claims trained one model sequentially on different tasks while using parameter-importance measures and a penalty term to protect performance on previously learned tasks. The specification connected that mechanism to reduced storage, lower system complexity, and mitigation of catastrophic forgetting. The Appeals Review Panel concluded that the claim, considered as a whole, improved how the model operated and integrated the mathematical concept into a practical application. Appeal No. 2024-000567, at 7–10 (P.T.A.B. Appeals Review Panel Sept. 26, 2025) (designated precedential Nov. 4, 2025). The favorable § 101 ruling did not establish patentability: the panel left the § 103 rejection undisturbed. Id. at 10.

For examination, uncertainty alone does not justify a § 101 rejection. The August 2025 reminder instructs that a close call should be rejected only when ineligibility is more likely than not, while compact prosecution still requires examination under §§ 102, 103, and 112. 2025 Eligibility Reminder, supra, at 5. That internal examination instruction does not alter the standard a court applies.

The emerging line is therefore not simply “AI core technology good, AI application bad.” A field application can be eligible if it requires a specific technical implementation that improves a machine, process, signal, network, or other technology. Conversely, a claim nominally directed to training can remain abstract if it recites only mathematical objectives and generic execution. The decisive drafting move is to claim the operational mechanism that produces the technical result.

B.                 China: Mental Activity and the Technical-Solution Requirement

China ordinarily applies two related subject-matter screens. First, if a claim consists only of an abstract algorithm, mathematical rule, or business rule, it is excluded as a rule or method for mental activities under Article 25. A claim containing a genuine technical limitation, rather than placing a computer only in the title, may pass that screen. It must still satisfy Article 2(2): considered as a whole, the claim must use technical means conforming to natural laws to solve a technical problem and obtain a technical effect. PRC Patent Law arts. 2(2), 25(1)(2); CNIPA Guidelines pt. II, ch. 9, §§ 6.1.2–6.1.3.

Adding a processor, memory, or AI chip is not enough when those components merely store and execute an abstract algorithm. CNIPA’s AI guidance gives the example of a computer system for training a neural network in which the processor and memory are only carriers for instructions; improving generic training efficiency, without a technical association to computer operation, does not necessarily solve a technical problem or produce a technical effect. China Nat’l Intell. Prop. Admin., Guidelines for Patent Applications for Artificial Intelligence-Related Inventions (Trial) pt. III (Dec. 31, 2024) [hereinafter CNIPA AI Guidelines].

Three routes are especially important. First, an algorithm may process data having definite technical meaning, such as images, audio, communications signals, or sensor measurements, and produce a result tied to a technical process. Second, an algorithm may have a specific technical association with internal computer structure and improve storage, transmission volume, processing speed, processor utilization, or another aspect of computer performance. CNIPA has treated a training method that chooses between single- and multi-processor execution based on training-data size as technical because the selection improved hardware execution. Third, big-data processing may be technical when it discovers an internal relationship governed by natural laws and improves the reliability or accuracy of analysis in a specific field. By contrast, a financial-price prediction based on historical prices follows economic, not natural, relationships and does not become a technical solution merely because a neural network performs the prediction. CNIPA Guidelines pt. II, ch. 9, § 6.1.3; China Nat’l Intell. Prop. Admin., Interpretation of the 2023 Amendments to the Patent Examination Guidelines: Computer-Program Inventions pts. II.2–II.3 (Jan. 18, 2024).

The Chinese threshold can therefore diverge from the U.S. outcome. Processing technical images or sensor data may more readily supply technical character in China, while a U.S. claim that merely collects, analyzes, and displays that information may still fail § 101. Dental Monitoring illustrates the risk. But passing Article 2 does not answer inventive step. A conventional model attached to a technically meaningful data stream may be protectable subject matter yet remain an obvious transfer of known technology.

V.               Novelty in an AI-Saturated Prior-Art Environment

Neither country has displaced ordinary novelty law simply because a reference or claimed invention involves AI. In the United States, § 102 encompasses patents, printed publications, public uses, sales, and other public availability. 35 U.S.C. § 102(a). China defines prior art as technology known to the public in China or abroad before the filing date. PRC Patent Law art. 22. In practice, an AI search must extend beyond patent classifications to papers, preprints, model cards, technical documentation, source-code repositories, datasets, conference presentations, standards submissions, and benchmark disclosures.

Machine authorship does not automatically immunize a public disclosure from prior-art treatment. A publicly accessible AI-generated document could potentially qualify under ordinary principles, but its legal effect still depends on the statutory category, timing, accessibility, what it actually teaches, and whether the disclosure is enabling. The USPTO’s 2024 request for comments identified unresolved questions about AI-generated publications, hallucinated disclosures, enablement, and whether routine AI use affects the level of ordinary skill. Request for Comments Regarding the Impact of the Proliferation of Artificial Intelligence on Prior Art, the Knowledge of a Person Having Ordinary Skill in the Art, and Determinations of Patentability Made in View of the Foregoing, 89 Fed. Reg. 34,217, 34,218–20 (Apr. 30, 2024). No final AI-specific rule had resolved those issues as of July 19, 2026.

CNIPA likewise has cautioned against treating an LLM response as proof of the prior art or the skilled person’s knowledge. In a 2026 reexamination, an applicant relied on an LLM-generated answer to argue that the prior art lacked a technical teaching. CNIPA concluded that an answer dependent on data sources, model, training accuracy, and prompt could serve as a reference but could not directly establish the technical teaching or the legal conclusion. Reexamination Decision No. 1927206, Application No. 202010182089.3 (CNIPA 2026) (published July 1, 2026). That decision concerns evidence, not whether an underlying source located or reproduced by an AI system can qualify as prior art.

Public-disclosure timing creates a major portfolio trap. U.S. law provides a one-year exception principally for inventor-originated or derived disclosures. 35 U.S.C. § 102(b)(1). China has no comparable general grace period. Its six-month protection applies only to enumerated circumstances, including specified emergency or public-interest disclosures, recognized exhibitions or conferences, and unauthorized disclosure. PRC Patent Law art. 24. A repository release, model card, customer demonstration, or ordinary conference submission that may fall within a U.S. exception can still destroy Chinese novelty. The reliable cross-border rule is to file before public benchmarking, publication, deployment, or demonstration.

VI.            Nonobviousness and Inventive Step

Section 103 does not ask whether the invention was created with or without AI; it expressly provides that patentability “shall not be negated by the manner in which the invention was made.” 35 U.S.C. § 103. The inquiry remains whether the differences from the prior art would have been obvious to a person having ordinary skill, considering motivation, predictable use, and reasonable expectation of success. KSR International Co. v. Teleflex Inc., 550 U.S. 398, 415–22 (2007). A new dataset, target label, customer segment, or desired output may be a reason to apply known machine learning, not evidence of a nonobvious implementation. Unexpected results and objective indicia are most persuasive when tied to the mechanism actually recited in the claim.

China evaluates inventive step under Article 22, commonly using a three-step analysis: identify the closest prior art, determine the distinguishing features and the technical problem actually solved, and ask whether the prior art supplied a teaching or motivation to reach the claimed solution. For mixed claims, algorithmic features contribute to inventiveness when they functionally support and interact with technical features—meaning they work together as technical means to solve the technical problem and produce the corresponding effect. CNIPA Guidelines pt. II, ch. 9, § 6.1.4. The 2025 amendments further emphasize that a feature said to make the inventive contribution must appear in the claim; a technical effect described only in the specification does not enter the inventive-step analysis if the claim omits the feature that causes it. CNIPA Order No. 84, pt. VI.

The new examples draw a clean line. Example 18 compares prior-art fruit counting with a claim using conventional deep learning to count ships. Although the objects differ in appearance, size, and environment, the claim did not require a changed training mode, model hierarchy, labeling method, or other substantive adjustment. Changing what the images depicted did not make the solution inventive. CNIPA Order No. 84, pt. VII, ex. 18. Example 19 reaches the opposite result for grading scrap steel. Overlapping and irregular scrap required extraction of color, edge, texture, and inter-feature relationships. The claim changed the number and arrangement of convolution and pooling paths and layers to address that field-specific technical problem and improve grading accuracy. Because the prior art did not teach those adjustments, the algorithmic features contributed to inventiveness. Id. ex. 19.

Recentive and Example 18 thus converge on a practical result through different legal routes. Neither system rewards a claim merely because a known model is deployed in a new environment. Architecture, parameters, preprocessing, feature selection, training, or execution changes must be more than implementation detail when they are the asserted advance; they should be the center of the claim.

VII.         Disclosure and the AI “Black Box”

United States § 112 requires the specification to demonstrate possession of the claimed invention and enable a skilled person to make and use its full scope without undue experimentation. 35 U.S.C. § 112(a); Ariad Pharmaceuticals, Inc. v. Eli Lilly & Co., 598 F.3d 1336, 1351–52 (Fed. Cir. 2010) (en banc); Amgen Inc. v. Sanofi, 598 U.S. 594, 610–14 (2023). AI does not create a special statutory standard, but result-oriented and “black box” claims sharpen ordinary problems. A specification that says only to select any suitable model, train it on relevant data, and generate an improved result may not show possession or enable the breadth of a claim covering materially different architectures, data regimes, and training techniques.

The necessary detail depends on the contribution. When the advance lies in model operation, the specification should address the material architecture, modules, layers, connections, objective or loss function, training sequence, parameter updates, and representative alternatives. When the advance lies in an application, it should explain the technical meaning and preprocessing of the inputs, relevant characteristics of the training data, feature representation, thresholds, inference sequence, output-to-action logic, deployment environment, and causal connection to the asserted technical result. Validation data and comparisons to an appropriate baseline can support both enablement and the credibility of the improvement. Source code, a complete corpus, or final trained weights are not categorically required, but withholding the information that produces the claimed result creates risk proportional to claim breadth.

China now states those expectations more expressly. For an invention involving model construction or training, the current Guidelines generally require clear disclosure of the necessary modules, layers or connection relationships, and concrete training steps and parameters. For a field application, the specification generally should explain how the model or algorithm is combined with the field, how inputs and outputs are configured, and how their internal relationship permits the skilled person to implement the solution. CNIPA Guidelines pt. II, ch. 9, § 6.3.1, as amended by CNIPA Order No. 84, pt. VII.

Examples 20 and 21 show that the inquiry is practical rather than a checklist demanding every possible detail. In Example 20, the specification did not identify the exact location of a spatial-transformer network within a convolutional architecture, but the skilled person could supply that placement from common general knowledge because the relevant layers, sharing relationship, and input-output operations were otherwise clear. CNIPA Order No. 84, pt. VII, ex. 20. Example 21 involved a model said to predict malignant tumors from blood-test indicators and facial features. The application did not identify which of many blood indicators mattered, did not establish a reliable relationship between general facial features and the claimed tumors, and offered no validation showing that the model could produce the asserted result. The omission was not routine implementation detail; it left the skilled person unable to determine how the solution solved its stated problem. Id. ex. 21.

The cross-border consequence is significant. A priority specification drafted only to a loose, functional U.S. description may lack the architecture, parameter, or input-output disclosure needed in China. Those details ordinarily cannot be added during national-phase prosecution without creating new matter. See 35 U.S.C. § 132(a); PRC Patent Law art. 33. A coordinated filing should therefore disclose the technical mechanism and fallback embodiments at the first filing, even if the initial U.S. claims are broader.

Disclosure also forces a deliberate patent-versus-trade-secret decision. Weights, source datasets, feature engineering, tuning, and deployment techniques may be commercially sensitive. But an applicant cannot claim the benefit of a hidden mechanism while disclosing only the desired result. The better approach may be to patent a reproducible architecture or control layer and retain nonessential tuning know-how as a trade secret, rather than seek a broad claim whose asserted advance depends on undisclosed information.

VIII.       Human Inventorship

The United States and China agree on the basic answer: an AI system cannot be named as an inventor. Their formulations of the required human contribution, however, should be applied with care.

In the United States, “inventor” means the individual who invented or discovered the claimed subject matter. 35 U.S.C. § 100(f). The Federal Circuit held in Thaler v. Vidal that “individual” unambiguously means a natural person and therefore excludes an AI system. 43 F.4th 1207, 1211–12 (Fed. Cir. 2022), cert. denied, 143 S. Ct. 1783 (2023). The court did not decide whether inventions made by humans with AI assistance may receive patent protection. Id. at 1213.

The USPTO’s November 2025 notice is examination guidance rather than a judicial holding or substantive regulation. It rescinded the February 2024 AI inventorship guidance in its entirety and withdrew the use of the Pannu joint-inventorship factors as a general AI-specific framework. The revised notice states that the same inventorship standard applies to every invention regardless of whether AI was used; there is no separate or modified standard for an AI-assisted invention. Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54,636, 54,636 (Nov. 28, 2025) [hereinafter Revised Inventorship Guidance].

That rescission did not withdraw the separate July 2024 subject-matter-eligibility update. For eligibility, the manner in which an invention was developed—including with AI assistance—is not part of the Alice/Mayo inquiry, which focuses on the claimed invention. 2024 AI SME Update, 89 Fed. Reg. at 58,138. But section IV of that update cross-referenced the now-rescinded February 2024 inventorship guidance and repeated its “significant contribution” formulation. That cross-reference is no longer the current inventorship test; the ordinary conception analysis in the November 2025 Revised Inventorship Guidance controls. 2024 AI SME Update, 89 Fed. Reg. at 58,138; Revised Inventorship Guidance, 90 Fed. Reg. at 54,636–37.

When one natural person uses AI, ordinary conception controls. The person must form a definite and permanent idea of the complete and operative claimed invention, possess knowledge of every claim limitation, and have the solution sufficiently settled that only ordinary skill would be needed to reduce it to practice without extensive research or experimentation. Id. at 54,636–37; Burroughs Wellcome Co. v. Barr Labs., Inc., 40 F.3d 1223, 1227–28 (Fed. Cir. 1994). When multiple natural persons participate, the Pannu factors apply only to determine which of those humans qualify as joint inventors; AI use does not alter that analysis. Pannu v. Iolab Corp., 155 F.3d 1344, 1351 (Fed. Cir. 1998); Revised Inventorship Guidance, 90 Fed. Reg. at 54,637.

The revised guidance treats generative AI and other computational models as instruments analogous to laboratory equipment, computer software, or research databases. An AI system may provide services or generate ideas without becoming an inventor or joint inventor. Revised Inventorship Guidance, 90 Fed. Reg. at 54,637.

The USPTO generally presumes that the inventors named in the application data sheet or oath or declaration are the actual inventors. If an application identifies an AI system or another non-natural person as an inventor or joint inventor, however, the guidance instructs examiners to reject all claims under 35 U.S.C. §§ 101 and 115 or take other appropriate action. Revised Inventorship Guidance, 90 Fed. Reg. at 54,637.

Ownership or operation of an AI system does not itself establish conception, and the revised guidance does not preserve the rescinded 2024 prompting-and-selection heuristics as categorical rules or safe harbors. Prompting, selecting outputs, modifying results, testing, and validation matter only insofar as the evidence helps prove that a natural person possessed a complete mental picture of every limitation of the claimed invention and could describe it with particularity. See id.; In re Jolley, 308 F.3d 1317, 1323 (Fed. Cir. 2002). The analysis remains fact-intensive and tied to the claims.

The revised guidance applies to utility, design, and plant patents and applications. Design-patent inventorship uses the same standard as utility-patent inventorship. For a plant patent, the natural person must contribute to creating the plant; merely appreciating its uniqueness and asexually reproducing it is insufficient. Revised Inventorship Guidance, 90 Fed. Reg. at 54,637.

China’s Implementing Regulations define an inventor as a person who makes a creative contribution to the substantive features of the invention. A person who only organizes work, facilitates use of material conditions, or performs other auxiliary tasks is not an inventor. PRC Implementing Regs. art. 14 (2023). CNIPA’s DABUS reexamination concluded that an AI system lacked the civil status necessary to hold the personal and property rights associated with inventorship and that an inventor must be a natural person who completes the invention through human intelligence and talent. Reexamination Decision No. 1373038, Application No. 201980006158.0 (CNIPA 2024) (published Aug. 13, 2024).

The 2025 amendments now state expressly that the inventor must be a natural person, all inventors’ identity information must be truthful, and neither an organization nor an AI name may be entered. CNIPA Order No. 84, pt. I. CNIPA’s AI Guidelines distinguish AI-assisted inventions, for which a natural person makes the substantive creative contribution, from AI-generated inventions produced without human substantive contribution. CNIPA AI Guidelines pt. II. For the latter, current law supplies no qualified AI inventor, and a human who did not make the required contribution cannot be named simply to preserve the filing.

Inventorship records should therefore be treated as part of the filing process. Development teams should preserve contemporaneous evidence showing which natural person formed the definite and permanent idea of each claimed invention, knew its limitations, and could describe it with particularity. Prompts, outputs, evaluation records, modifications, and validation results may be retained as part of that evidence, but none independently establishes conception. Counsel should audit inventorship before the first priority filing and again when claims change.

Priority and benefit claims require particular care. Under the notice’s announced policy, the prior-filed application and the later United States application must name the same natural-person inventor or have at least one natural-person joint inventor in common. A foreign application naming an AI system as the sole inventor therefore cannot support a United States priority claim. If the foreign application names both natural and non-natural persons, the United States application data sheet should identify only the natural persons, including at least one natural person in common with the foreign filing. The same approach applies upon national-stage entry under 35 U.S.C. § 371. Revised Inventorship Guidance, 90 Fed. Reg. at 54,637.

IX.             China’s AI Ethics Screen

China’s express ethics review is the most visible substantive divergence. Article 5 denies a patent to an invention that violates law or social morality or harms the public interest. PRC Patent Law art. 5(1). The 2025 Guidelines amendments apply that provision specifically to AI and big-data applications whose data acquisition, label management, rule setting, recommendation, or decision logic contains unlawful, immoral, or public-interest-harming content. CNIPA Order No. 84, pt. VII, § 6.1.1.

The examples are intentionally concrete. One claims a retail mattress-sales system that captures facial images and performs identity recognition for targeted marketing without establishing lawful acquisition or individual consent. Because the data practice is embedded in the solution rather than merely a collateral use, it fails Article 5. Another constructs an autonomous-vehicle emergency-decision model that assigns collision priorities based on a pedestrian’s age and sex. CNIPA treats the discriminatory rule itself as contrary to social morality. Id. exs. 1–2.

The boundary matters. Implementing Rule 10 states that an invention does not violate Article 5 merely because implementation of the invention is prohibited by law. PRC Implementing Regs. art. 10 (2023). The stronger Article 5 case arises when illegality or immoral decision logic is built into the claimed solution or necessary disclosure, not whenever a lawful technology could be deployed unlawfully. Compliance boilerplate cannot cure a scheme whose operation inherently requires covert data collection or discriminatory rules, but an application need not become a general regulatory treatise. It should accurately describe lawful data provenance, consent, de-identification, authorization, or bias controls when those features are technically true and relevant.

United States patent examination has no comparable general AI morality screen. Patentability does not authorize commercial practice and does not displace privacy, discrimination, medical-device, export-control, or other substantive law. Professional duties concerning candor, signature, confidentiality, and verification likewise should not be recast as a morality-based § 101 doctrine. See 37 C.F.R. §§ 1.56, 11.18.

Medical AI presents a separate statutory overlay in China because methods for diagnosing or treating disease are excluded under Article 25(1)(3). PRC Patent Law art. 25(1)(3). Claims directed to devices, systems, image-processing operations, or intermediate technical results may require a different analysis than a claim whose endpoint is the diagnostic judgment itself. In the United States, medical AI instead encounters the ordinary § 101 framework, including the separate case law governing laws of nature and diagnostic correlations.

X.                Claim Formats and a Coordinated Drafting Strategy

AI development is a lifecycle, not a single “processing” step. Where supported, a portfolio should consider separate claims to data acquisition and preprocessing, model construction and training, deployment and inference, output-driven technical control, monitoring and retraining, rollback, distributed edge or cloud execution, and specialized hardware. The objective is not multiplication for its own sake. It is to place enforceable claim sets around the actors who practice distinct portions of the system: the model developer, training-platform provider, cloud inference provider, device manufacturer, and enterprise deployer.

In the United States, common formats include method, system or machine, and nontransitory computer-readable-medium claims. Claims directed only to intangible information or a transitory signal remain problematic. China permits method, apparatus, computer-readable storage-medium, and computer-program-product claims. Its 2023 Guidelines amendments clarified that a computer program product may be claimed as a software product without being confined to a physical storage medium. CNIPA Guidelines pt. II, ch. 9, § 5.2.

The common specification should begin with a concrete technical problem. It should identify the specific architecture, training, data transformation, resource-allocation, or control mechanism that addresses the problem and explain why it produces the technical result. It should provide fallbacks at several levels: system arrangement, model modules and connections, training steps and parameters, input and output definitions, preprocessing, deployment, and post-inference action. Comparative performance data should use a meaningful baseline and identify the claimed feature responsible for the improvement.

The independent claims need not recite every implementation detail, but they should not omit the only feature that distinguishes the invention from generic AI. In the United States, that omission invites Recentive and Dental Monitoring: the specification may promise an improvement while the claim covers any suitable model producing a desired result. In China, the same omission can prevent the algorithmic feature from contributing to inventive step or leave the claim without the necessary technical means. The safest common principle is simple: put the contribution-producing mechanism in the claim and disclose more than one way of implementing it.

Jurisdiction-specific prosecution should then emphasize different parts of the same disclosure. A U.S. response should identify the claimed technological implementation, distinguish generic field-of-use cases, and explain how the claim improves model, computer, or other technological operation. When facts would assist, an applicant may voluntarily file a Subject Matter Eligibility Declaration (“SMED”) under 37 C.F.R. § 1.132 with evidence of the filing-date state of the art, skilled-artisan understanding of the original disclosure, or objective results tied to the claimed improvement. The examiner memorandum’s AI example uses comparative regression testing for a disclosed neural-network architecture. U.S. Pat. & Trademark Off., Subject Matter Eligibility Declarations: Memorandum to the Patent Examining Corps 2–5 (Dec. 4, 2025) [hereinafter SMED Corps Memorandum]. The evidence must have a claim nexus and cannot supply information required in the original disclosure; an opinion on the ultimate legal conclusion receives no weight. Id. at 3. Properly submitted evidence must be considered with the whole record under the preponderance standard, but it is not a safe harbor and does not control the result. Id. at 3–4; U.S. Pat. & Trademark Off., Best Practices for Submission of Rule 132 Subject Matter Eligibility Declarations 1–4 (Apr. 30, 2026).

A Chinese response should articulate the technical problem, technical means, and technical effect; identify the natural-law or internal-computer relationship; and connect the claimed algorithmic features to the technical features with which they interact. Architecture, training parameters, and input-output relationships needed for that account must be present in the original filing. A China-directed application also should undergo an Article 5 review before filing, not after the specification has committed to a questionable data or decision architecture.

Literal translation is not claim coordination. Terms such as training, inference, parameter, feature, label, embedding, model, and output should retain stable technical meanings across languages, while the independent claims should be tailored to each office’s legal test and local infringement realities. Filing before disclosure, preserving inventorship and model-development records, and aligning the priority specification to both countries’ disclosure requirements are more important than forcing identical claim text.

XI.             Conclusion

The comparison does not support a simple conclusion that one jurisdiction is categorically friendlier to AI patents. United States law foregrounds eligibility and increasingly asks whether a claim improves AI or computer operation rather than merely deploying familiar analytics. Chinese law routes much of the same inquiry through the technical-solution requirement, inventive step, and disclosure, while adding an express ethics and data-governance screen.

The doctrinal vocabulary differs, but generic functional claiming is vulnerable in both systems. “Use AI to generate an optimized result” identifies neither an invention nor an inventor. A durable claim instead identifies technical inputs, a specific model or training adaptation, an output-driven action, and the mechanism connecting those elements to a measurable technical result. The specification must then disclose enough detail to support that scope, and the record must identify the natural person or persons who conceived the claimed invention under United States law and those who made the requisite creative contribution to its substantive features under Chinese law.

AI assistance does not itself defeat patentability, and the presence of AI does not establish it. In the United States, ordinary natural-person conception applies without an AI-specific modification; in China, a natural person must make the prescribed creative contribution. In both jurisdictions, the invention must also be technically specific and adequately disclosed.