A Comparative Overview of Patenting AI Inventions
Co-authored by: Lin Li (Wolf Greenfield), Ed Walsh (Wolf Greenfield), Nancy Song (Linda Liu & Partners), and Holger Veenhuis (Eisenführ Speiser)
AI flies above the geographical frontiers that often hem in and constrict patent law. Caught between a revolutionary, border-effacing technology and a place-dependent body of law, innovators must navigate how global legal differences shape their patent strategies.
To orient innovators in this evolving landscape, this 2026 update will map out the terrain in three distinct locations: the United States (US), Europe, and China. We will elaborate on Wolf Greenfield’s 2025 mid-year webinar, Patenting AI-Implemented Inventions: A Global Comparison, by considering 1) where patent standards diverge on patent eligibility; 2) how the different regions assess inventiveness; and 3) how to satisfy patent disclosure requirements while safeguarding proprietary AI know-how. We will conclude with practical guidance on how to turn jurisdictional differences into strategy.
Patent Eligibility: Where Standards Diverge
United States
In the US, courts and the United States Patent and Trademark Office (USPTO) approach eligibility differently. Statistically, the Federal Circuit frequently invalidates claims, but outcomes could differ by panel. By contrast, the USPTO has a modified approach, finding no abstract idea where claims reside in technological improvements. This modified approach has contributed to reduced rejection rates and examiner variability.
In December 2025, the USPTO issued additional guidance that reinforces its subject matter eligibility framework for AI-related inventions. In the precedential In re Desjardins, the USPTO confirmed that improvements to the functioning of machine learning models may constitute practical applications.
The USPTO also issued two memoranda addressing the use of Subject Matter Eligibility Declarations (SMEDs) under 37 C.F.R. § 1.132. These memoranda said that applicants may voluntarily submit factual evidence, such as evidence of technological improvement or integration of a judicial exception into a practical application. The memorandum includes examples illustrating how such evidence may inform eligibility determinations involving, e.g., mental processes, improvements to technology, particular treatment or prophylaxis, and claims amounting to significantly more. Such evidence must be evaluated under the preponderance-of-the-evidence (more likely than not) standard.
In practice, successful US patent eligibility strategies emphasize presenting AI inventions not as generalized data processing or decision-making, but rather as concrete technical solutions. Describing meaningful preprocessing or post-processing steps can help demonstrate technological improvements. One can strengthen eligibility positions during prosecution by explicitly aligning claim language and arguments with favorable USPTO examples (e.g., AI-specific eligibility examples introduced in the USPTO July 2024 guidance). Where appropriate, one can also bolster these arguments by submitting factual evidence through SMEDs.
Although one cannot predict how the courts will rule in future cases, courts will look more favorably on patents that clearly emphasize technical improvements and are supported by a robust prosecution record.
China
In China, to establish AI invention eligibility, the applicant needs to show how the AI algorithm technically correlates with the internal structure of computer systems in typically one of two scenarios. On the one hand, if the AI algorithm improves hardware’s computational efficiency or produces other internal performance improvements, it would be considered as possessing specific technical correlation. On the other hand, even if no apparent hardware performance improvement is demonstrated, an invention may still be considered to possess specific technical correlation where the type of data processed by the AI invention is defined and the execution of the algorithm conforms to natural laws (as opposed to that for financial data, etc.).
Note that abstract concepts or applications grounded in social science (e.g., financial fraud detection based on non-technical data) are generally not patentable. However, inventions involving the processing of technical data are more likely to be patent-eligible.
As shown in the examples in the 2026 revised Examination Guidelines section that deals with Article 5 of the Chinese Patent Law, no patent can be granted for any invention-creation that violates the law or social morality or is detrimental to public interests. Also, under the revised guidelines, eligibility turns on privacy. Privacy violations trigger Article 5, preventing the patent application from even entering a novelty or inventiveness assessment. For example, an AI invention that automatically adjusts office lighting might rely on capturing images of individuals’ faces to determine the number of occupants in a room. Such a practice could potentially violate Article 5 of the Chinese Patent Law if it involves improper collection of personal information.
To be successful, applicants should show how AI algorithms affect the hardware system, such as by showing improvements in computational efficiency or other internal performance metrics. Where this cannot be easily demonstrated, the applicant can still establish the required “specific technical correlation” by identifying the specific application domain or how the AI algorithm processes the types of the technical data.
In addition, for AI systems that process personal data, the specification should address privacy considerations and regulatory compliance (e.g., compliance with China’s Personal Information Protection Law).
Europe
In Europe, AI inventions must pass a “two-hurdle” test. The first determines eligibility and the second analyzes the inventive step.
To pass the first hurdle, one must show that under Article 52 of the European Patent Convention (EPC), the claimed subject matter possesses technical character. This is relatively easy. The European Patent Office (EPO) Guidelines expressly state that while neural networks and other machine learning algorithms are per se of an abstract mathematical nature, they are not excluded when claimed in a technical context (Guidelines, G-II, 3.3.1). According to established case law, notably T 258/03 (Hitachi), the use of technical means, such as a computer, is sufficient to confer technical character on a claim.
However, passing the second “inventive step” hurdle is often more challenging, as will be explained below.
Same Prior Art, Different Results: Assessing Inventiveness
United States
Under the Graham factors and the Supreme Court’s decision in KSR, inventiveness (non-obviousness) of AI inventions is assessed in an obviousness analysis that is the same as with other technologies. This obviousness analysis encompasses all claim limitations, including features that other jurisdictions may characterize as non-technical. As a result, US patent law can include elements relating to data processing, information handling, or decision logic when supported by the specification and shown to interact with the prior art in a non-obvious way—although other jurisdictions might discount these factors.
China
Chinese law assesses AI-related claims features like algorithmic features or business-method-related features.
The key consideration is whether the AI algorithmic features are functionally integrated with, and mutually supportive of, the technical features. Under the Examination Guidelines, this can be shown in one of three scenarios: 1) the algorithmic features are applied in a specific technical field to address a specific technical problem; 2) the algorithmic features require corresponding adaptations of technical means; or 3) the algorithmic features, in combination with technical features, produce an objective effect that improves user experience.
Regarding scenario two, it is important to demonstrate specific adaptations to new technical domains. One cannot just apply a known AI model to a new field without also adjusting parameters or modifying the algorithm. The Examination Guidelines provide a negative example where the algorithm merely switches from counting fruit to counting ships. Therefore, it is critical to clearly document the technical challenges addressed.
Regarding scenario three, an improvement in user experience when the improvement results from the interaction between algorithmic features and technical features would be considered for inventiveness assessment. The Examination Guidelines provide an affirmative example in which optimization of data architecture and communication methods (technical features) improves both parcel delivery efficiency and the user experience of parcel pickup.
Europe
As explained above, although AI inventions typically clear the first eligibility hurdle, they may encounter challenges in the second hurdle of the inventive step analysis. To overcome this second hurdle, one should consider that the EPO applies the problem-solution approach, as defined by the COMVIK doctrine (T 641/00) for inventions involving mixed technical and non-technical subject matter.
First, the EPO construes the claim. It identifies which features contribute to the claim’s technical character, in that these are features that produce or serve a technical effect. The EPO also identifies which features are non-technical, such as the abstract mathematical content of an AI model, data-analysis concepts, or business-driven aims.
Second, starting from the closest prior art, the EPO chooses the “distinguishing features” that credibly contribute to a technical effect. These are retained as potential contributors to the inventive step. On the other hand, it can treat the non-technical distinguishing features as part of the constraints given to the skilled person.
Third, based on the technical distinguishing features, the EPO formulates the objective technical problem statement. The problem statement may include non-technical distinguishing features, but they cannot provide the inventive contribution. For example, a non-technical distinguishing feature can be a target to be achieved.
Finally, the EPO determines obviousness. Looking to prior art and common general knowledge, it asks if the skilled person would have implemented the technically contributing features to solve the objective technical problem. Technically contributing features could be any technical adaptations of the AI, such as those improving a technical process, signal or image processing, or the internal functioning of a computer.
Accordingly, improvements attributable only to the choice of model architecture, training strategy, or predictive quality in a non-technical field typically do not support the inventive step. By contrast, AI features that are functionally linked to a technical purpose or that produce a further technical effect may support the inventive step.
Navigating Patent Disclosure Requirements While Safeguarding Proprietary AI Know-How
United States
The USPTO infrequently rejects claims based on enablement, definiteness, or written description grounds. However, these issues remain crucial. In demonstrating how an AI model achieves a claimed technical effect, one can strengthen both eligibility and Section 112 positions.
We see this in the contrast between the ineligible machine learning training claim in Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), and the eligible claim in USPTO’s example 47. The difference lies in the disclosures. While both claims seek to cover real-time results produced by trained machine learning models, only the latter provides technical disclosure explaining how the real-time results are achieved.
At the same time, applicants should carefully tailor disclosures to avoid inadvertently revealing unclaimed implementation details that may be better protected as trade secrets.
China
Although China National Intellectual Property Administration (CNIPA) does not require detailed descriptions of well-known AI architectures or models, for inventions that improve the training or generation of the architectures or models, one should be transparent regarding the model structures, parameter modifications, and other relevant aspects of the architectures or model. For inventions on specific applications for AI, one should specify input and output data types and explain the relationship therebetween and how they relate to the technical outcomes.
Along these lines, it is advisable to include verification data in the specification demonstrating the advantageous effects achieved. Such data is particularly important when the claims involve non-intuitive correlations. Examples include diagnosing dementia based solely on button-press duration, or identifying certain cancers based only on facial features.
Europe
Under Article 83 EPC, an AI invention must be disclosed in a manner sufficiently clear and complete for it to be carried out by the skilled person over the entire scope claimed. According to the EPO Guidelines (F-III, 1; F-III, 3) and the established case law of the Boards of Appeal, this requires that the application must contain a concrete technical teaching that enables the skilled person, using common general knowledge, to reproduce the claimed technical effect without undue burden and without exercising inventive skill.
To show that the inventions can be implemented across the whole claimed range, one should disclose the AI system’s essential components—such as the nature of the input and output data, relevant aspects of the model architecture, preprocessing steps, and the training framework. While exhaustive parameter listings or proprietary datasets are not required, the disclosure must provide sufficient technical detail (e.g., structure of the neural network and interaction of its components) to ensure that the skilled person with average knowledge can realistically carry out the invention and achieve the asserted technical effect.
Practical Guidance: Turning Jurisdictional Differences into Strategy
Regardless of where we are on the map, our compass can be guided by some constant practices, such as:
- Technical Storytelling: Have a clear problem-solution narrative showing how the invention solves a technical problem. While improving hardware performance is generally regarded as technical, contributions in other fields can also be considered technical, depending on the problem addressed. Avoid emphasizing purely non-technical considerations, such as business-related objectives.
- Layered Claiming: Include both broad, high-level claims and narrower, technically detailed dependent claims. Incorporate into the claims non-obvious technical features. For example, specify the technical purpose of the output data or what technical effect the AI system achieves.
- Necessary Disclosure: Describe the AI invention sufficiently to enable a skilled person to carry it out, including relevant aspects such as the underlying system or model architecture, input data, preprocessing steps, training data, output data, and postprocessing steps. Avoid unnecessary disclosure of unclaimed implementation details that might be better protected as trade secrets.
- Clarity and Consistency: Use consistent terminology and clearly defined technical terms to be clear and to avoid examiner objections.
- Privacy-Conscious Design: Where applicable, explain how the invention handles personal data responsibly. For example, address compliance with relevant data protection regulations.
Similar to how artists differ in their style and expression, patent practitioners may apply these principles in different ways, depending on their skill and judgment. Innovators seeking to protect AI inventions globally will benefit from partnering with experienced counsel who can thoughtfully translate technical innovations into effective patent strategies across jurisdiction.
Related content: