AI-Related Inventions Before the Patent Office: Best Practices for Ensuring Sufficient Disclosure

We have reviewed seven cases where the EPO found lack of sufficiency in AI cases. In one of them, the first-instance finding was overturned on appeal.

Although the cases relate to different applications and types of AI, some common conclusions can be made.

General drafting good practice

The general drafting good practice holds also when it comes to AI. In particular, it is recommended to use unambiguous terminology. All terms should be used in the same way as known in the art or unambiguously defined. Otherwise, they can cause difficulties when it comes to sufficiency, see T 1539/20.

Statements of the kind “Defining the services from the link classes depends on how the IT system 800 has been implemented by developers and system integrators, and how it is operated“ make the EPO examiner suspicious since they seem to indicate that the invention cannot be put into practice without the developers and system integrators becoming inventive. This should be avoided, see T 1539/20.

Specifics of computer-implemented inventions – Levels of abstraction

As has been observed generally for computer-implemented inventions (see for example Schwarz/Kruspig: Computerimplementierte Erfindungen, Carl Heymanns Verlag, 3^rd ed. 2024, p. 16-17), inventions involving AI may be described on different levels, and both very high-level and low-level descriptions can cause difficulties:

• The highest level is a purely functional description, which is most likely not sufficient as it only states the problem and not the solution. An example of this is T 1191/19 Neuronal plasticity – Institut Guttmann.
• An upper mid-level description usually describes items / parts of the AI as black boxes, e.g. “an RNN”. They include structural features and can usually be put into practice by the skilled person. AI researchers know what RNNs are and can easily name one or more examples. However, only mentioning these structures may not be enough since the technical effect is not in all cases proved for any network of this type. Therefore, issues may arise as to sufficiency of disclosure over the full range claimed, support, or credibility as an inventive step question.
• Lower mid-level descriptions usually cite well-described AI types, e.g. an LSTM. These are more specific and can usually be put into practice by the skilled person, even without citing literature (such as the classic by Hochreiter and Schmidhuber). In particular, it is not necessary to name a specific trained model or an implementation in most cases, as far as the term for the AI is not too general. If amendments to the neural network have been made, they can be explicitly described. This may include a description of a structure down to the node level, and/or stating a loss function (e.g. as a formula) used for training.
• Low-level descriptions – anything at the implementation level – do usually not provide further insight. In particular, lines of code should only be included where they actually illustrate the invention, see the EPO Guidelines F-II 4.12.

Aspects of AI

Machine Learning Algorithms are mathematical models that automatically learn from raw data. This means that their functionality stems in part from explicit programming, and in part from acquiring, i.e. extracting, their knowledge from the raw data. This implies that Machine Learning algorithms have three aspects to them that may be patentable alone or in combination:

AI Structure: The general structure of the model is human made. In other words, programming the model is a mental activity as much as programming any other algorithm. The resulting software product may implement a patentable computer-implemented method. For EPO and German proceedings, its features contribute to inventive step under the Comvik approach if they effect or conduct sensor data analysis, control of a machine, or are specifically adapted to the working of the computer. To disclose a claimed structure of an AI, the features must be described down to a level on which the invention can be put into practice by using the standard building blocks available to the skilled person.

To sufficiently disclose a claimed AI, the information on the AI in the description must be more detailed than a mention of one or more functionally defined features, such as “classification process” (T 509/18).

When drafting applications, it is sensible to include examples of data types or data items generated by intermediate steps of a method. Even if each processing step of a method is in principle well-known, the Boards often look specifically for examples of what data is generated by each of them (e.g. in T 509/18). If such data is exemplified in the description on a low level, this makes the argument that the skilled person knows at least one way of carrying out the invention easier.

Training: A method directed to training an AI may be patentable. In general, the training data may be synthetic or measured data. Measured data are technical, synthetic data only under the conditions set out in G1/19. In order to sufficiently disclose training of an AI, some indication of the training data is necessary.

After decision T 161/18 Äquivalenter Aortendruck – ARC Seibersdorf, some observers warned that the decision could be the beginning of a much stricter approach in AI sufficiency (e.g. F. Hagel, EPI Information 4/2020, p. 22), this did not seem to have materialized. In particular, we have not seen any requirement emerge that training data must be filed, let alone that an open-weights implementation of the trained AI is required.

Inference: The application of the ML model to a particular problem is technical under the conditions set out in G 1/19. In order to sufficiently disclose an inference phase, a method to obtain the applied model must also be disclosed. This includes some information on the structure of the AI and the training data needed to provide the trained AI – bearing in mind that such information may be implicit and that information known to the skilled person need not be provided. However, it is – at least on a conceptual level – more difficult to sufficiently disclose an application of an AI that only its structure and/or the training, so that statements that some computation can be done by an AI are in general not enough (see e.g. T 1079/17).

In particular, different parts of the application relating to different phases of use of the AI (training and inference) should be made compatible at the drafting stage (T 606/21).

Comparison to the US proceedings

In four cases (T 509/18, T 161/18, T 1539/20, T 606/21) the EPO refused the application although parallel US patents were granted. In one case (T 1191/19), the USPTO refused a parallel application but for other reasons: obviousness, lack of eligibility. Only in one case did the USPTO object to the application referring to the written description requirement (T 1669/21). Therefore, the EPO can be seen to be stricter than the USPTO.