Introduction
The standard AI product FTO searches for patents that cover what the model does: its output, its functional capabilities, the system architecture through which it delivers those capabilities at inference time. This is the right starting point. But it is not the complete picture.
A large and commercially significant category of AI patents covers not what the model does at inference but how it was built in training. Data curation method patents, training algorithm patents, fine-tuning technique patents, and synthetic data generation patents all cover the process by which a model acquires its capabilities — not the capabilities themselves. A model that is fully clear at inference may have been trained using a method that infringes a patent covering the training process.
This training layer of AI patent risk is systematically underassessed in most AI product FTO analyses. Our analysis of FTO for software products and how to map claims when a patent covers a method not a product covers the general methodology for method claim FTO. This article covers the specific application of that methodology to AI and machine learning training method patents — the patent category that AI product teams most consistently miss.
The AI Patent Landscape Has Two Distinct Layers
Layer 1 — Inference and deployment patents: The first layer covers what the trained model does when deployed. Computer vision system patents covering how image classification is performed at inference. NLP method patents covering how text is parsed, classified, or generated. Recommendation system patents covering how outputs are ranked and served. These patents attach to the product’s commercial function and are where most AI FTO analyses are scoped.
Layer 2 — Training method patents: The second layer covers how the model was built — the methods used to curate and preprocess training data, the optimisation algorithms used during model training, the fine-tuning and transfer learning techniques used to adapt a base model, and the synthetic data generation methods used to augment training datasets. These patents attach to the model development process. The party that performed the training, not the party that deploys the model, is the potential direct infringer.
Why Layer 2 is consistently missed: Training method patents are filed in different classification codes from inference and deployment patents. A standard AI product FTO scoped to G06N (Computing using neural networks) and G06V (Image/video recognition) will capture many inference patents but systematically miss training method patents filed in G06F (General computing methods) and specific algorithm and optimisation sub-classifications. The scope gap is a classification structure problem — and it requires an explicitly expanded scope to close.
What Training Method Patents Actually Cover
Data Curation and Preprocessing Method Claims Patents covering how training data is selected, filtered, labelled, augmented, and preprocessed. Active curation method claims, data augmentation algorithm claims, and data balancing method claims all attach to the training pipeline rather than the deployed model.
Training Algorithm and Optimisation Technique Claims Patents covering specific gradient descent variants, loss function formulations, regularisation techniques, distributed training coordination methods, and hyperparameter optimisation algorithms. A model trained using a patented optimisation algorithm infringes the patent at the point of training, regardless of the deployed model architecture.
Fine-Tuning and Transfer Learning Method Claims Patents covering how a pre-trained foundation model is adapted to a specific downstream task — fine-tuning scheduling techniques, layer freezing strategies, adapter module architectures, and prompt-tuning methods. Companies building specialised AI products by fine-tuning publicly available foundation models may inadvertently perform patented fine-tuning methods.
Synthetic Data Generation Method Claims Patents covering GAN-based image synthesis, simulator-based data generation for robotics and autonomous systems, and language model-based text augmentation. As synthetic data generation becomes standard for addressing training data scarcity in regulated industries, this category is growing in commercial significance.
Who Holds Training Method Patents
Large technology company portfolios: Google (Alphabet), Meta, Microsoft, and OpenAI hold the most extensive AI training method patent portfolios. Google’s portfolio spans data augmentation methods, distributed training coordination, and neural architecture search algorithms. Meta’s portfolio is concentrated in self-supervised learning methods and contrastive learning algorithms. Microsoft holds significant positions in fine-tuning methods, retrieval-augmented generation (RAG) training approaches, and responsible AI training methodology.
Academic institution spinout and licensing positions: Many foundational training method patents originated in academic research programmes at Stanford, Carnegie Mellon, MIT, and major European research universities. These patents are frequently held by the institution and licensed through technology transfer offices or spinout companies. For AI product teams using training methods developed in academic research contexts, the academic licensing layer is a specific FTO consideration.
NPE accumulation in AI training methodology: Non-practising entities have been accumulating AI training method patents through portfolio acquisition. As major technology companies have sold pre-LLM-era AI patent portfolios, and as foundational machine learning patents from the 2000s and 2010s have been acquired through corporate restructuring, NPEs have built positions in AI training methodology they are beginning to assert against commercial AI product companies.
How to Assess Training Method Patent FTO
Assessing training method patent FTO requires working through four analytical questions that differ materially from inference-stage patent analysis. The general framework for method claim analysis in our guide on when to conduct an FTO search applies — but each question has an AI training-specific dimension that requires separate consideration.
- The infringement geography question — where does training occur?: For method patents, infringement occurs where the method is performed. For AI training method patents, infringement occurs at the computational infrastructure where training runs — the data centres, cloud computing regions, or on-premises hardware where the training computation is executed. A model trained in US cloud infrastructure using a US-patented training method infringes in the US, regardless of where the model is subsequently deployed.
- The single-entity question for distributed training pipelines: Modern large model training pipelines are frequently distributed across multiple organisations. A foundation model may be pre-trained by one organisation, fine-tuned by a second, and deployed by a third. For training method patent infringement, the question is which entity performed the claimed method steps and whether any single entity performed all of them.
- Alice/§101 validity risk: AI training method patents face significant Alice/§101 subject matter eligibility risk. Claims directed to abstract mathematical operations performed on data are vulnerable to §101 challenges. For blocking training method patents identified in the FTO, an Alice validity pre-assessment is valuable — identifying patents that are challenge candidates before committing to a design-around that may be unnecessary if the patent does not survive a validity challenge.
- Design-around options: Training method design-arounds require changes to the model development process rather than the product itself — substituting a different data augmentation algorithm, modifying an optimisation schedule, or using a different fine-tuning architecture. Each requires re-training the model, which is computationally expensive. This is the strongest argument for identifying training method blocking patents before training begins, not after a production model has been built.
“The training method patent is the most expensive AI patent risk to design around — because the design-around requires re-training the model from the point where the infringing method was used. Identifying it before training begins, when the training pipeline is still being designed, is a fraction of the cost of identifying it after a production model has been built on a method that needs to change.”
The Data Layer: When the Blocking Right Is Not a Patent
Copyright in training data: AI training method FTO sits alongside a separate clearance question: whether the training data was obtained and used in compliance with copyright law. This is not a patent FTO question — copyright and patent are different IP regimes — but it is a parallel clearance requirement. Training data copyright clearance covers whether the data was lawfully obtained, whether the licence terms permit commercial model training, and whether the model’s outputs constitute infringement of the training data originals.
How copyright and patent clearance interact: A training pipeline change made to avoid a training method patent may use different training data that raises copyright questions. A training data substitution made to address copyright concerns may require a different preprocessing method covered by a different patent. The two clearance questions are technically independent but practically interdependent — changes made to resolve one may affect the other. Both should be assessed in coordination.
How Our FTO Service Covers AI and Machine Learning Products
Our freedom to operate service covers AI and machine learning FTO across both the inference/deployment layer and the training method layer — with explicit search coverage of training algorithm classification codes, data curation and preprocessing method patents, fine-tuning and transfer learning technique patents, and synthetic data generation method patents alongside the standard AI system and application patent classifications. For AI product teams at any stage of the model development pipeline, we structure the FTO scope to cover the training method risk that standard inference-focused AI FTO analyses miss.
Building or fine-tuning an AI model? The training method is a separate FTO question from the deployed product. Our service covers training algorithm patents, data curation method patents, and fine-tuning technique patents alongside the standard inference-layer AI FTO. → Contact Us
Conclusion: The Takeaway
AI patent risk has two layers. The inference and deployment layer — what the model does — is where most AI FTO analyses are scoped. The training method layer — how the model was built — is where the most expensive and least anticipated blocking patents sit. Training method patents cover data curation, optimisation algorithms, fine-tuning techniques, and synthetic data generation. They attach to the model development process, not the deployed product, and they are infringed at the point of training.
An AI FTO that covers only the inference layer is answering half the patent clearance question. The complete AI FTO covers both layers — and covers the training method layer before training begins, when design-around options are still computational choices rather than expensive re-training decisions.
