The “New” Relevance of the Enablement Requirement for Drafting AI Patent Applications
Artificial intelligence (AI) – and in particular machine-learning (ML) and neural-network–based inventions – has rapidly become one of the most active fields of patent filing before the European Patent Office (EPO). While discussions often focus on the assessment of inventive step under Art. 56 EPC, recent case law shows that the requirement of sufficiency of disclosure (enablement) under Art. 83 EPC has gained renewed and significant importance for AI-related inventions.
The Boards of Appeal have repeatedly stressed that AI inventions cannot rely on generic references to “machine learning” or “neural networks”. Instead, they must be enabling: the skilled person must be able to reproduce the trained model, to select appropriate data, and to achieve the claimed technical effect across the entire scope of the claims (see decisions T 161/18, T 1669/21, and T 1191/19).
In more detail, in AI inventions, the training phase is often the core of the technical contribution. If said training cannot be reproduced, the invention is not enabled (see T 1669/21). According to T 1669/21, a detailed disclosure of the AI model, input and output data, and training is required. Particularly, the Board held that the AI/ML model must be disclosed in detail, and that the patent must disclose the model's architecture (e.g., topology, layers), the mathematical modelling of nodes, the learning methods and algorithms, and how the model is implemented. The use of broad categories (e.g., “physiological data”, “environmental parameters”) is insufficient. Concrete measurements and examples must be disclosed so that the model can be reproduced. The patent must guide the selection of data, and it must provide criteria, rules, or examples for selecting appropriate parameters, not just list general data categories. Training data must be disclosed in quantity and quality, particularly addressing the source of the training data, the properties of the dataset, how the dataset covers the relevant parameter space, and why the data is suitable for achieving the claimed technical effect. The Board made clear that it cannot be assumed that the skilled person knows “what the correct data would look like”.
Further, in T 1191/19, the Board held that missing data and missing model structure destroy sufficiency with respect to Art. 83 EPC. The underlying application in T 1191/19 related to a meta-learning scheme for predicting personalized patient interventions, but is failed to disclose any example set of training or validation data, the structure or topology of the neural networks used as classifiers, the activation functions, the learning mechanism, or the stopping criteria. At the abstraction level of the underlying application, the Board found that the skilled person would face an undue burden in attempting to execute the meta-learning scheme. The invention was not sufficiently disclosed, violating Art. 83 EPC.
Even beyond AI-specific cases, the EPO continues to tighten its sufficiency standards. For example, the decisions T 0149/21, T 941/16, and T 438/19 may provide crucial analogies for AI applications: According to T 0149/21, whole-range sufficiency may apply to mechanical and control systems. One may derive that, based on that, whole-range sufficiency may also apply to AI patent applications. In detail, according to T 0149/21, a single enabling embodiment is insufficient when the claims cover broad functional variants. For AI inventions, this may mean that, if the claim covers several types of input data, the description must enable them. Moreover, if the claim covers any model capable of achieving a result, the patent must teach how that result can be achieved across the entire scope.
According to T 941/16, functional claims require structural or experimental guidance. This may mean that, if claims define an AI model by its function (e.g., “a model configured to classify X” or “a network trained to detect Y”), the description must enable the skilled person to produce substantially all models achieving that function. Again, a single example of a trained model may not suffice if the claim covers much more.
According to T 438/19, parameter-based claims require measurable and achievable parameters. AI claims often rely on parameters such as recognition accuracy, confidence scores, threshold values, or generalization errors. Following T 438/19, this may mean that the patent must teach how to measure, reproduce, and achieve such parameters. Otherwise, the claim may fail under Art. 83 EPC.
Thus, drafting an AI application may now require a level of detail comparable to disclosing a chemical or a biotechnological invention.
Historically, sufficiency objections were most common in biotechnology and chemistry. With AI, however:
Training data determine the technical performance
AI models are only as good as the data they are trained on. Without disclosing the data (or at least how to obtain it), the invention may not be reproducible.
Model architecture directly affects the technical effect
Vague references to “a neural network” cannot show how the claimed technical improvement is achieved.
Functional claims are pervasive in AI drafting
Many AI claims are formulated in terms of tasks or functions (“classifying”, “predicting”, “detecting”), which may trigger the strict “whole-scope enablement” doctrine.
AI models are inherently non-deterministic
Without clear training instructions, different skilled persons would obtain different models with different behaviors, violating Art. 83.
In general, the EPO does not consider data selection or training details to be “non-technical”. Decisions such as T 1669/21 confirm that disclosure of training data and training methods is part of the technical teaching required for enablement.
All these developments make it clear: Art. 83 EPC has become a central, and often decisive, requirement for AI patents.
Based on the case law above, practical recommendations for drafting AI patent applications include:
Explicitly state the use of AI/ML
Avoid general statements such as “using machine learning techniques”.
Describe the model in detail
Include the architecture, training algorithm, hyperparameters, model topology, and implementation aspects.
Precisely define input and output data
Provide examples, formats, ranges, and measurement methods.
Provide guidance on selecting data or parameters
Explain the criteria, thresholds, or decision rules for data selection.
Disclose the training data
If the data cannot be given verbatim (e.g., privacy reasons), one may disclose: how it is obtained or derived, its statistical characteristics, its domain, its representativeness.
Ensure “whole-range” sufficiency
If claims cover many model types or data types, provide multiple enabling embodiments or clear instructions that apply across the full scope.
Sufficiency of disclosure (enablement) is now a key battleground for AI patent application and patents. The enablement requirement of Art. 83 EPC has gained new, practical importance for AI patent drafting – and ignoring it puts even seemingly strong AI inventions at risk.