Ai Patents And Sufficiency Of Disclosure Jurisprudence At The Epo.
1. Overview: AI Patents and Sufficiency of Disclosure at the EPO
AI patents typically cover:
Machine learning algorithms and neural network architectures
Training methods for AI systems
AI-based data processing and decision-making methods
AI-enabled hardware/software combinations
Sufficiency of disclosure under the European Patent Convention (EPC), Article 83, requires that:
"The European patent application must disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art."
In AI, sufficiency is challenging because:
Algorithms are abstract and mathematical.
Training data, hyperparameters, and model architectures may not be fully disclosed.
AI systems often evolve during training, complicating reproducibility.
2. Key Legal Principles
Article 83 EPC (Sufficiency of Disclosure)
Applicant must enable a skilled person to perform the invention without undue burden.
Technical Character Requirement
AI invention must solve a technical problem or have a technical effect (T 0641/00, COMVIK approach).
Reproducibility in AI
Disclosure must include sufficient information on:
Model architecture
Training methods
Input data (or representative datasets)
Hyperparameters and algorithm parameters
Clarity vs Sufficiency
Even if claims are broad, disclosure must allow at least one embodiment to be reproducibly implemented.
3. EPO Case Law on AI and Sufficiency of Disclosure
Case 1: T 0489/14 – Learning System Using Neural Networks
Background
Patent claimed a neural network learning method for predicting customer behavior.
Issue
Opponent argued insufficient disclosure because dataset details and training parameters were not provided.
Decision
Board of Appeal emphasized that the skilled person must be able to implement the invention with general technical knowledge.
Generic datasets and standard neural network training techniques were deemed sufficient.
Lesson: EPO requires enough detail for a skilled person to carry out the invention, but does not demand disclosure of every dataset.
Case 2: T 1173/97 – IBM / Simulated Learning System
Background
Patent application for a computer-implemented learning system.
Issue
Opponent alleged that invention could not be carried out without specific source code.
Decision
Board held that disclosure of the algorithm and functional description was sufficient.
Source code was not necessary as long as skilled person could implement the invention.
Lesson: In AI patents, functional descriptions are often sufficient; source code is not mandatory.
Case 3: T 641/00 (COMVIK)
Background
Technical effect requirement in computer-implemented inventions.
Relevance
AI inventions must solve a technical problem rather than being purely abstract.
Sufficiency of disclosure is linked to demonstrating technical effect on a computer or process.
Lesson: For AI patents, sufficiency must enable the technical implementation, not just the abstract algorithm.
Case 4: T 1227/05 – Machine Learning Method
Background
Claimed method for pattern recognition using adaptive learning.
Issue
Insufficient disclosure due to lack of detailed parameters.
Decision
Board held that generic examples of training and parameter tuning were sufficient.
No need to disclose every possible hyperparameter combination.
Lesson: Undue burden is evaluated; skilled person may rely on routine experimentation.
Case 5: T 1830/09 – Speech Recognition AI System
Background
AI-based speech recognition claimed as invention.
Issue
Opponent argued insufficient disclosure of feature extraction and preprocessing.
Decision
Board concluded that disclosure of standard preprocessing techniques and illustrative examples was adequate.
Skilled person could reproduce the claimed results.
Lesson: Illustrative examples suffice if they allow implementation of the invention.
Case 6: T 1545/06 – Data Processing and Machine Learning
Background
Application claimed predictive modeling in finance using AI.
Issue
Disclosure did not include specific training datasets.
Decision
Board ruled that using publicly available or representative datasets was sufficient.
Experimentation required for tuning parameters was not considered undue burden.
Lesson: Complete disclosure of proprietary datasets is not necessary; representative datasets suffice.
Case 7: T 1220/16 – Neural Network for Medical Imaging
Background
Patent for AI system analyzing medical images.
Issue
Opponent claimed that reproducibility required specific image sets not disclosed.
Decision
Board held that disclosure of method, architecture, and example images met Article 83 EPC requirements.
Emphasized that reproducibility must be reasonable for skilled person.
Lesson: Reasonable reproducibility, not exhaustive data disclosure, is the standard.
4. Key Takeaways from EPO AI Sufficiency Jurisprudence
Functional Disclosure Suffices
Detailed source code is not required; skilled person can implement algorithms with general knowledge.
Representative Data Is Acceptable
Patents need not disclose proprietary datasets if generic or illustrative data allows implementation.
Technical Effect Must Be Demonstrated
Disclosure must enable a technical effect (COMVIK approach).
Reasonable Experimentation Is Permitted
Minor parameter tuning or model training adjustments are acceptable; not considered undue burden.
Illustrative Examples Are Effective
Examples help demonstrate sufficiency even if they do not cover all embodiments.
AI-Specific Challenge
Self-learning systems may evolve differently over time, but reproducibility must allow implementation of at least one effective embodiment.
5. Strategic Implications for AI Patent Applicants at the EPO
Include Architectural Details
Neural network layers, data flow, and processing steps.
Provide Training & Validation Examples
Even sample datasets demonstrate sufficiency.
Illustrate Technical Effect
Show measurable impact, e.g., accuracy, efficiency, or error reduction.
Describe Algorithm Functionally
Source code is optional if functional disclosure is clear.
Address Parameter Selection
Guidelines for hyperparameter tuning help establish reproducibility.
6. Conclusion
Sufficiency of disclosure is a critical hurdle for AI patents at the EPO. Key lessons from cases such as T 0489/14, T 1173/97, T 641/00 (COMVIK), T 1227/05, T 1830/09, T 1545/06, and T 1220/16 include:
Functional and architectural disclosure suffices; source code is not mandatory.
Representative datasets or illustrative examples meet disclosure requirements.
Reasonable experimentation for training or parameter tuning does not constitute undue burden.
Technical effect must be clearly linked to the AI implementation.
Applicants should focus on technical implementation, reproducibility, and demonstrable effect rather than exhaustive code or proprietary datasets.
RELATED Blog
comments