The AI/ML patent landscape is expanding

Rapid patent growth

Official studies support a strong growth trend, but they do not support a single universal count for all “AI/ML patents.” Results depend on the search taxonomy, publication date, family definition, and dataset. WIPO’s 2019 AI study identified nearly 340,000 AI-related inventions through 2016 and reported that more than half had been published since 2013.^[1]

• Through 2016: WIPO identified nearly 340,000 AI-related inventions in its study dataset.
• 2013 onward: more than half of those identified inventions had been published since 2013.
• 2023: WIPO later reported approximately 14,000 published GenAI patent families.
• 2025: WIPO’s 2026 update reported more than 37,800 published GenAI patent families, while noting that GenAI remains a subset of AI-related patenting.^[2]

These figures are not interchangeable: the broad 2019 AI study and the later GenAI update use different scopes and time periods. For FTO work, the practical lesson is to document the search definition and cutoff date rather than rely on a headline count.

Patent holders

AI-related patent ownership spans technology companies, semiconductor and automotive businesses, specialist developers, universities, public research organizations, and non-practicing entities. WIPO’s historical 2019 study found that companies represented 26 of the top 30 applicants in its dataset, while universities or public research organizations represented the remaining four.^[1]

That historical snapshot should not be treated as a current ranking. An AI FTO analysis should identify relevant owners from the claims and patent families found in the search, then verify current ownership, jurisdiction, and legal status.

Unique FTO challenges in AI and machine learning

Challenge 1: Broad or functionally worded claims

Some AI-related claims describe a result or processing function at a relatively high level. Illustrative themes—not quotations from any particular patent—include predicting an outcome with a trained model, processing data with a neural network, or optimizing a technical process with machine learning.

FTO consideration: claim breadth cannot be assessed from a title or abstract. Review the complete claim, its required elements, prosecution history where relevant, jurisdiction, and the accused product configuration. For teams organizing that review, Eureka’s FTO Search workflow builds and refines search strategies, screens potentially relevant claims with legal-status context, and organizes evidence into claim-level comparisons.^[3] AI output still requires qualified professional review.

Challenge 2: Validity questions are separate from clearance

In the United States, AI-related claims may raise subject-matter eligibility questions under 35 U.S.C. §101, obviousness questions under §103, or written-description and enablement questions under §112.^[4][5][6] These are distinct legal inquiries, and their application is claim- and fact-specific.

FTO consideration: do not treat a possible validity argument as a finding of non-infringement or freedom to operate. U.S. patents are presumed valid under 35 U.S.C. §282, and infringement, invalidity, unenforceability, ownership, and legal status should be evaluated separately.^[7][8]

Challenge 3: Rapidly evolving technology

Model architectures, training practices, inference techniques, and deployment patterns change quickly. Older patent language may use different terminology from current engineering teams, while later filings may target new model or application layers.

FTO consideration: combine keywords, classifications, citations, assignee searches, semantic searching, and periodic monitoring. Record the search date and product version so that later updates can be compared against a defined baseline.

Challenge 4: Multiple patent layers

AI/ML products can involve several technical layers:

• model architecture and algorithm claims;
• data collection, selection, preprocessing, augmentation, and training-method claims;
• application-specific claims;
• accelerator, memory, networking, and other hardware claims;
• software framework, orchestration, inference, and deployment claims.

FTO consideration: map product components and process steps to each layer before searching. This prevents the review from focusing only on the model while missing hardware, data-pipeline, or deployment claims.

Challenge 5: Open source AI/ML frameworks

Open source availability does not by itself answer patent-clearance questions. Relevant issues include the exact version used, included dependencies, modifications, distribution model, applicable license, contributor patent terms, patent-termination clauses, and third-party patent rights outside the license.

FTO consideration: review the actual license and dependency tree with counsel. Do not assume that all permissive licenses contain the same patent grant.

Challenge 6: Training data and training methods

Patent claims may address how data is generated, selected, labeled, transformed, augmented, or used to train a model. That is more precise than assuming that a dataset itself is necessarily patented.

FTO consideration: document the training pipeline, including preprocessing and augmentation, then compare each potentially relevant claim to the implemented steps.

Conducting effective AI FTO analysis

Step 1: Define the subject technology

The subject-technology definition should identify:

• AI/ML algorithms and model architectures;
• training data sources and training methods;
• the product’s specific AI application;
• hardware and accelerator components;
• frameworks, libraries, dependencies, and versions;
• integration, inference, deployment, and update workflows;
• target countries and relevant commercial acts.

Step 2: Conduct a structured patent search

Search for patents covering similar algorithms, applications, training methods, hardware, frameworks, and deployment techniques. Combine technical concepts with relevant classifications, citations, assignee names, inventor names, patent-family data, and legal-status filters.

Useful official search interfaces include USPTO Patent Public Search, WIPO PATENTSCOPE, Espacenet, and national patent-office databases. General-purpose interfaces can supplement the work, but database coverage, family grouping, translations, and status data should be checked before relying on a result.

Common starting terms include machine learning, neural network, deep learning, transformer, reinforcement learning, inference, fine-tuning, retrieval-augmented generation, and application-specific terms. Search strategies should also account for synonyms and older terminology.

Step 3: Assess claim scope and product mapping

For each potentially relevant patent, identify the claims that matter, break them into required limitations, and map those limitations to the product or process. Confirm the relevant jurisdiction, family members, ownership, legal status, expiration considerations, and product configuration.

Key questions include whether the claim requires the specific architecture, training step, application, hardware component, or deployment behavior used by the product. A feature-level comparison can also reveal possible design-around options, but counsel should assess whether a proposed change actually avoids every required claim limitation.

Step 4: Assess validity without merging it with infringement

For high-priority patents, a separate validity review may consider eligibility, prior art, obviousness, written description, enablement, and other jurisdiction-specific grounds. In the United States, the USPTO’s current eligibility materials include AI-specific examples, but examination guidance is not a substitute for controlling law or litigation analysis.^[4]

FTO consideration: record validity theories as a separate workstream. They may influence mitigation or negotiation strategy, but they do not automatically eliminate infringement exposure.

Step 5: Review open source framework terms

For each framework and dependency, record the precise version, source repository, license text, notices, modifications, and distribution model. Review express patent grants, their scope, and termination provisions where present. Also consider patents held by parties that did not contribute code under the applicable license.

Step 6: Develop mitigation strategies

• modify algorithms, model architecture, training steps, or deployment behavior;
• change hardware, libraries, or framework components;
• seek a license or commercial resolution;
• obtain a focused non-infringement or validity opinion where appropriate;
• consider available challenge procedures with qualified counsel;
• accept a documented residual risk with contingency planning and approval.

Practical examples for AI/ML patent review

Example 1: Large AI/ML portfolios

WIPO’s 2019 study identified corporate applicants across AI techniques and applications, including Alphabet in the historical dataset.^[1] The lesson is not that one company’s portfolio automatically blocks a product. Instead, teams should search by claim scope, family, jurisdiction, status, and the implemented product features.

Example 2: Patent validity analysis

AI-related claims are subject to the same statutory validity framework as other U.S. patent claims. Eligibility under §101, obviousness under §103, and disclosure requirements under §112 remain distinct grounds.^[4][5][6]

Lesson: a credible validity theory can affect strategy, but the analysis must be claim-specific and should not be described as a completed invalidity finding unless an authorized tribunal has made that determination.

Example 3: Open source framework licenses

TensorFlow’s official repository uses Apache License 2.0. Section 3 includes a contributor patent license limited to claims necessarily infringed by the contribution alone or in combination with the work, and it contains a patent-litigation termination condition.^[9]

PyTorch’s current repository license is BSD-style and does not state the same express Apache 2.0 patent grant in that license text.^[10]

Lesson: review each framework’s exact license rather than assuming that open source status supplies a complete or uniform patent defense.

Best practices for AI FTO analysis

1. Start early. Begin during product definition and update the review when architecture or deployment changes.
2. Assess multiple layers. Cover models, data pipelines, applications, hardware, frameworks, and deployment.
3. Use multiple search methods. Combine semantic, keyword, classification, citation, assignee, and family searches.
4. Separate validity from infringement. Track each question in its own analysis.
5. Review open source terms precisely. Record versions, dependencies, license grants, and termination clauses.
6. Evaluate design-arounds carefully. Confirm that a change addresses every relevant claim limitation.
7. Monitor new publications and status changes. Set a review cadence tied to development and launch milestones.
8. Obtain qualified legal advice. Seek jurisdiction-specific review for material risks.
9. Document the record. Preserve search scope, dates, queries, results, assumptions, and decisions.
10. Keep the technology definition current. Update the record when models, data, hardware, or dependencies change.

Emerging AI/ML patent issues

Issue 1: Generative AI patents

Generative AI patent publications are increasing rapidly. WIPO’s 2026 update reported that published GenAI patent families rose from approximately 14,000 in 2023 to more than 37,800 in 2025.^[2]

FTO consideration: search model architecture, training and fine-tuning, retrieval, inference, content generation, safety controls, and application-specific workflows without assuming that a GenAI label captures all relevant claims.

Issue 2: Foundation model patents

Foundation-model products can combine base-model architecture, adaptation methods, orchestration, retrieval, tool use, serving infrastructure, and downstream applications.

FTO consideration: define which layers are developed internally, licensed, accessed through an API, or supplied by customers. The commercial and technical boundaries affect what should be mapped and searched.

Issue 3: AI safety and explainability patents

Searches may need to cover adversarial robustness, interpretability, monitoring, evaluation, access control, bias mitigation, and related technical implementations.

FTO consideration: treat safety and explainability features as separate product components when they have their own architecture, process steps, or deployment behavior.

Conclusion

AI and machine learning can create layered patent-clearance questions across models, data processing, applications, hardware, open source components, and deployment. A disciplined AI FTO analysis defines the product precisely, searches across those layers, maps claim limitations to implemented features, separates validity from infringement, and updates the record as the product changes.

Frequently asked questions about AI FTO analysis

Sources and verification

1. WIPO Technology Trends 2019: Artificial Intelligence—key findings. World Intellectual Property Organization.
2. GenAI Innovation Soaring, With Patent Activity Nearly Tripling in Two Years. WIPO, July 14, 2026.
3. AI Patent Search, FTO & Design Clearance. Patsnap Eureka IP Search.
4. Subject matter eligibility. United States Patent and Trademark Office.
5. 35 U.S.C. §103—Non-obvious subject matter. U.S. House Office of the Law Revision Counsel.
6. 35 U.S.C. §112—Specification. U.S. House Office of the Law Revision Counsel.
7. 35 U.S.C. §282—Presumption of validity; defenses. U.S. House Office of the Law Revision Counsel.
8. 35 U.S.C. §271—Infringement of patent. U.S. House Office of the Law Revision Counsel.
9. TensorFlow official repository and Apache License 2.0. TensorFlow.
10. PyTorch LICENSE. PyTorch.

Sources checked July 31, 2026. Product capabilities are limited to statements supported by the current public Eureka IP Search page. Product features and legal standards may change.