AI innovation crosses borders faster than legal systems can adapt. A model trained in one jurisdiction can be deployed globally within days. A breakthrough by a U.S. company may encounter patentability standards that differ sharply in Europe or Asia.

For innovators building scalable AI technologies, securing protection now requires navigating a rapidly evolving patchwork of international frameworks. Significant new guidance from the USPTO, CNIPA, the UK Supreme Court, and the EPO has emerged between mid-2025 and early 2026, reshaping examination practice, disclosure expectations, and strategic filing decisions.

According to WIPO's World Intellectual Property Indicators 2025, more than 19.7 million patents were in force globally across 142 jurisdictions in 2024, with patent filings rising 4.9% that year. WIPO's Patent Landscape Report on Generative AI documented more than 54,000 GenAI patent families filed between 2014 and 2023, with China alone accounting for over 38,000—roughly six times the U.S. total over the same period. China, the United States, and Europe continue to dominate global AI patent activity. Understanding the differences among these systems is essential for companies seeking robust, enforceable portfolios globally.

Key Takeaways

  1. Eligibility standards are diverging at the examination level but converging on core principles such as human inventorship.
  2. China now imposes the strictest disclosure and ethics requirements; drafting to satisfy CNIPA often meets or exceeds other jurisdictions' needs.
  3. The U.S. has adopted a more applicant-friendly examination posture under new USPTO leadership, while Federal Circuit precedent continues to constrain "apply generic ML to new domains" claims in litigation.
  4. The UK has materially lowered its eligibility threshold post-Emotional Perception, aligning more closely with EPO practice.
  5. Strategic claim drafting focused on model architecture or training improvements, combined with early inventorship documentation, consistently improves outcomes across major offices.

The European Patent Office: Technical Effect Remains the Standard

The EPO continues to evaluate AI inventions under its established framework for computer-implemented inventions. The April 2025 Guidelines for Examination (effective April 1, 2025) clarified rather than fundamentally altered the approach. AI and machine learning methods are not patentable on their own unless they demonstrate a further technical effect beyond abstract computation.

If a claim is directed to a method involving technical means (such as a computer) or to a device, the subject matter has technical character as a whole and is not excluded under Article 52(2) or (3) EPC. Computational models and algorithms contribute to technical character only if they help solve a technical problem—either by application in a specific technological field or by adaptation to a specific technical implementation.

AI used to improve image compression, optimize industrial control systems, or enhance cybersecurity infrastructure may qualify. Claims directed solely to business logic or abstract decision-making face greater scrutiny under the COMVIK approach to mixed inventions, which assesses inventive step using only the features that contribute to technical character. Where technical effect depends on training data characteristics, those characteristics should now be disclosed with sufficient specificity.

Practical implications: EPO examination rewards precise linkage between algorithmic features and concrete technical outcomes. Vague functional language or black-box model references rarely succeed.

China: Rapid Growth and Tightening Quality Standards

China leads the world in AI patent filings. By the end of 2024, the number of valid invention patents in China reached 4.756 million, making it the first country to exceed 4 million (CNIPA).

On December 31, 2024, CNIPA issued the Guidelines for Patent Applications for AI-Related Inventions (Trial Implementation), categorizing applications into four types: AI algorithms or models themselves; functions or applications based on AI; AI-assisted inventions; and AI-generated inventions. Inventors must be natural persons; AI-generated inventions without substantial human contribution are presumed ineligible.

Less than a year later, on November 13, 2025, CNIPA enacted Order No. 84—comprehensive amendments to the Patent Examination Guidelines effective January 1, 2026. These established a dedicated AI and big data section addressing legality and ethics under Article 5(1) of the Patent Law (explicit examples reject facial-recognition data without consent and discriminatory autonomous-vehicle decision models), sufficient disclosure for algorithm-related inventions (model architecture, training steps, parameters, and input/output relationships), and inventiveness assessment that rewards genuine improvements to algorithmic processes over mere domain reapplication.

Practical implications: Companies in sensitive sectors (e.g., autonomous systems, surveillance, hiring algorithms, or generative content) should integrate ethics and data-provenance reviews into their patenting workflow. Tailoring disclosures to CNIPA's heightened disclosure standards typically satisfies other major jurisdictions.

The United States: A Substantial Policy Reset at the Examination Level

The U.S. landscape shifted significantly in the second half of 2025 through agency action under new USPTO leadership.

The Federal Circuit's April 2025 decision in Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025), remains good law. The Federal Circuit held that "patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101."

At the examination stage, however, the USPTO has moved toward a more permissive posture:

  • The July 2024 AI Subject Matter Eligibility Guidance Update provided AI-specific examples (47–49) integrating AI/ML into the existing framework.
  • The August 4, 2025 Kim Memo (to TCs 2100, 2600, and 3600) narrowed the mental-process abstract-idea grouping, distinguished claims that "recite" a judicial exception from those that merely "involve" one, required holistic Step 2A Prong Two analysis, and instructed examiners to issue a § 101 rejection only when it is more likely than not (>50%) that the claim is ineligible. Close calls now favor the applicant.
  • In Ex parte Desjardins, Appeal No. 2024-000567 (ARP Sept. 26, 2025, designated precedential Nov. 4, 2025), the PTAB vacated a § 101 rejection of claims directed to training a machine learning model on multiple tasks while preserving prior-task performance. Claims directed to improvements to the machine learning model itself qualify under Enfish-style analysis.
  • The November 28, 2025 Revised Inventorship Guidance for AI-Assisted Inventions rescinded the February 2024 guidance in its entirety, withdrew application of the Pannu factors to AI-assisted inventions, and confirmed that AI systems are treated as instruments (analogous to laboratory equipment or software). Only natural persons may be named as inventors, consistent with Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).

Practical implications: Prosecution-stage opportunities have expanded for claims that genuinely improve model architecture or training processes. However, enforcement and validity challenges remain subject to Recentive scrutiny. Companies should document early any human inventive contribution and model-level improvements.

Japan and South Korea: Case-Driven Guidance and Patent-Density Strategy

Japan has adopted a case-example-driven approach. On March 13, 2024, the JPO added ten new AI case examples covering inventive step, support and enablement, and patent eligibility—with particular focus on generative AI inventions, trained models, and how prompts to large language models are evaluated for inventive contribution. A trained model described as a combination of a neural network and a parameter set operating on a computer is patent-eligible; a mere parameter set without hardware integration is not. Japan's generative-AI inventive-step analysis is among the most developed globally.

South Korea pursues a deliberate patent-density strategy. According to WIPO data, Korea remains the most patent-intensive filer relative to GDP. KIPO's guidelines for AI-related inventions (originally 2021, updated 2022) address subject matter eligibility, claim formats, enablement, and inventive step, complemented by accelerated examination programs for priority technologies.

The United Kingdom: A Landmark Overhaul Post-Emotional Perception

On February 11, 2026, the UK Supreme Court delivered its judgment in Emotional Perception AI Ltd v Comptroller-General of Patents [2026] UKSC 3—its first decision addressing the patentability of a computer-implemented invention and a unanimous overhaul of UK practice.

The Court abandoned the Aerotel four-step test, finding it had misinterpreted Article 52 EPC by collapsing the threshold "invention" inquiry into inventive-step assessment. In its place, the Court endorsed two EPO-derived approaches working together: the "any hardware" approach (from T 154/04 Duns Licensing and ratified in G 1/19) under which a claim involving a computer, database, network, or other technical means clears the initial exclusion hurdle; and the COMVIK-style "intermediate step" approach for assessing whether mixed technical and non-technical features contribute to a technical effect.

The practical effect is a materially lower eligibility threshold at the "invention" stage, with the substantive work shifting to the COMVIK-style technical-character analysis. The case was remitted to the UKIPO Hearing Officer. For international portfolios, the UK is now considerably more accessible for AI patent protection than it was a year ago and more closely aligned with EPO practice than at any point in the past two decades.

Practical implications: Forum considerations are important when deciding where to pursue global protection. The most robust global portfolios are built from claim sets drafted to satisfy the standards of the strictest jurisdiction(s) of interest (e.g., China disclosure + EPO technical effect + U.S. model-improvement focus) while leveraging more permissive examination where available.

Where Jurisdictions Are Converging — and Where They Are Not

Despite differences in technical-effect analysis, eligibility thresholds, and disclosure standards, the major jurisdictions have converged on one core principle: AI cannot be an inventor. The U.S. (Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), codified in the November 2025 USPTO guidance), China (December 2024 AI Guidelines plus January 2026 amendments), the UK/EPO, and Japan (Tokyo District Court DABUS ruling and subsequent practice) all require named inventors to be natural persons, with AI-generated inventions facing presumptive ineligibility absent substantial human contribution.

Genuine divergence remains at the eligibility and disclosure layer. A claim that succeeds before a permissive U.S. examiner under the Kim Memo and Desjardins may still face rigorous COMVIK analysis at the EPO or UKIPO. An invention meeting China's heightened algorithmic disclosure requirements may contain detail unnecessary elsewhere. A UK claim clearing the post-Emotional Perception "any hardware" threshold may still fail at the intermediate technical-contribution step.

Practical Takeaways for Practitioners and Innovators

For companies across technology, life sciences, and consumer sectors, these moves can consistently improve outcomes:

  1. Include model-architecture or training-method improvements to support data or domain applications. Recentive in the U.S., the EPO's technical-effect requirement, CNIPA's January 2026 inventiveness standards, and Japan's generative-AI case examples all reward genuine algorithmic contribution and penalize mere reapplication.
  2. Treat inventorship as a unified natural-person-only discipline from the disclosure stage forward. Verify that inventor declarations capture the humans who made substantive contributions to conception. Document AI's role as a tool, not a co-inventor. This approach is consistent with guidelines in all major jurisdictions.
  3. Sequence China filings to capture the strictest disclosure standards early. CNIPA's requirements for model architecture, training data characteristics, parameter relationships, and input/output relationships are now the most demanding. Tailoring applications for China typically satisfies other jurisdictions and reduces later-stage amendments.
  4. Reassess the UK as a primary or co-equal filing jurisdiction. The post-Emotional Perception threshold is meaningfully lower than the Aerotel era and broadly aligned with the EPO. Portfolios that previously deprioritized direct UK filings may benefit from revisiting that decision, particularly for computer-implemented inventions with arguable technical effect.

The future of AI patenting is international by default. Companies and their counsel who build claim sets aware of where these systems converge and diverge will build global portfolios worth defending.