The European Patent Office (EPO) has clarified its expectations for patents relating to Artificial Intelligence (AI) inventions. In a recent decision, the EPO emphasized five criteria for a sufficient disclosure to prevent patents being granted for “black box” AI systems where the inner workings are unknown.
I’ve distilled the key factors into a practical checklist that guides you through the disclosure process step-by-step. Enter your email below and I’ll send it to you:
What is Sufficiency of Disclosure?
In a nutshell, “sufficiency of disclosure” (Art. 83 EPC) means your patent application must describe the invention clearly and completely enough for a skilled person (e.g., a software engineer) to implement it without undue burden. This is similar to the “enablement” requirement in US patent law (35 U.S.C. 112), but the EPO emphasizes enablement across the full claim scope based on the written description.
At Least One Concrete, Workable Example
At the heart of the EPO decision lies the requirement for at least one concrete, workable example in the patent description. To ensure it is truly enabling, the decision has laid out five criteria:
- Specific Computational Model: Detail the architecture and algorithms of the AI model.
- Specific Selection of Input Variables: Clearly define the precise data that feeds the AI.
- Specific Output Variable: Specify the exact result or prediction the AI generates.
- Indication of the Training Method: Explain how the AI model is trained, including data, algorithms, and parameters.
- Plausibility: Demonstrate that the goals of the invention are actually achievable.
These five criteria, summarized in reason no. 1.8 of the decision (which is in German), form a practical checklist for AI patent applicants. They ensure your invention is described with enough detail to clear the “sufficiency of disclosure” hurdle.
These five criteria make a useful checklist for AI patent applicants, ensuring that their inventions are described with enough detail to meet the sufficiency hurdle.
Let’s dive deeper into each criterion…
1) Computational Model: No More Black Boxes
The decision criticizes that the patent did not disclose a concrete, workable example in which a specific computational model (“ein bestimmtes Rechenmodell”; Reasons No. 1.8) is disclosed.
What Went Wrong?
The patent in this case does not give any details about the computational model used. Despite mentioning of a “neural network” once in claim 9, the description offered only vague statements like “according to the invention, a computational model is created from at least part of the measured and determined data or parameters, by means of which these data or parameters are evaluated through calculations and subsequent analyses” (paragraph [0020] of the patent).
What the EPO Decision Says
The decision criticizes in reason no. 1.3.2 that the following characteristics of the computational model were left open:
- Which topology and class (arrangement of nodes and their connections) should the artificial neural network have?
- How are the nodes mathematically modeled (linking of input and output values; propagation and activation functions)?
- Which learning method is used?
How to Do It Better
In reason no. 1.3.2, the decision references an older decision T 0161/18 (Äquivalenter Aortendruck/ARC SEIBERSDORF), which provides a useful benchmark. In that case (EP 1 955 228 A2), the patent application described a neural network with:
- Three layers, with 100 neurons in the first and last layer, and 10 neurons in the middle layer
- A linear transfer function in the middle layer
- A bias added in the last two layers
- A fully connected feed-forward architecture without any shortcuts
- Supervised learning with bias initialization, where the bias of the first layer corresponds to the zero vector
While that application ultimately failed due to insufficient disclosure of the training data, the level of detail provided for the neural network itself seems to have been acceptable. This is confirmed by the present decision, which states that the older case did “at least specify the used neural network”. Therefore, the above list of neural network characteristics seems to be a safe baseline for sufficiently disclosing the computational model.
Key Takeaway
Clearly specifiy the computational model: If the invention involves a machine-learning model, the patent must provide details about the model’s architecture and type. This includes specifying the topology and class of the model, how the nodes are mathematically modeled (including the propagation and activation function), and the learning algorithm employed.
2) Input and Output Variables: Choosing The Right Ingredients
The decision criticizes that the patent did not disclose a concrete, workable example in which a specific selection of input variables (“eine spezifische Auswahl an Eingangsgrößen”; Reasons No. 1.8) is disclosed.
The patent in question dealt with predicting the wear and tear of a metallurgical vessel lining. Claim 1 specifies a long list of input data gathered about the vessel’s lining. The list includes data about the initial lining materials and their properties, operational parameters during use (like melt temperature, composition, and processing times), wall thickness measurements after use, and aspects of how the molten metal is introduced into and removed from the vessel.
What the EPO Decision Says
The decision criticizes in reason no. 1.6.2 that the patent does not give a single example that would answer the following questions:
- Which specific measurement values are particularly relevant for the computational model to learn a correlation between the input and the desired calculation outcome?
A major concern of the decision is the breadth of claim 1, which lists rather broad parameter categories, like “material properties” and “metallurgical parameters”, instead of concrete measurement values, and thereby would try to monopolize many possible combinations of input variables (reason no. 1.6.1).
The “Black Box” Argument
The patentee made an interesting argument that goes to the heart of any machine-learning innovation (see reason no. 1.6.4). They argued that precise knowledge of the relationships and the influence of the individual parameters on the wear is not required for an enabling disclosure, because it is precisely the essence of machine learning that the ability to predict the output variable without knowledge of the causal relationships is acquired through training in a self-learning manner. In the process, the influence of irrelevant input variables is filtered out by itself. If the selected input variables include parameters that are less relevant to the wear, this does not affect the predictive power of the calculation model. A “research program” would only be necessary if one wanted to clarify the causal relationships and the significance of each input variable, but this effort is not required for a machine-learning calculation model.
But the board of appeal did not follow. They argued in reason no. 1.6.5 that due to the lack of a specific example in the patent, there is no principle proof that a successful prediction of wear is possible with a selected set of parameters. The board found that a corresponding correlation was not self-evident. On the one hand, it depends on the quantity and quality of the training data set used. On the other hand, even if the inclusion of less relevant parameters is filtered out during training and the predictive power of the calculation model is not affected, the training would not lead to reliable results if an important influencing variable were missing. In this case, the calculation model might not be able to learn a reliable correlation between the wear and the stressed input variables.
How to Do It Better
A strong patent description should explain the following (see reason no. 1.6.3):
- At a minimum, at least one concrete, workable example with a specific selection of input parameters for the computational model.
- Specific instructions which input parameters are relevant for the desired output
- A principle proof that the computational model can learn a robust correlation.
Key Takeaway
Give clear guidance for parameter selection: When the claims include broad categories of data or parameters (e.g., “material properties”, “metallurgical parameters”), the patent needs to offer guidance on selecting specific, suitable parameters within those categories. This guidance should help the skilled person to identify relevant parameters and exclude irrelevant combinations.
3) Training: Teaching the Right Lessons
The decision criticizes that the patent did not disclose a concrete, workable example in which an indication of the used training method (“eine Angabe des verwendeten Trainingsverfahrens”; Reasons No. 1.8) is disclosed.
What the EPO Decision Says
The decision criticizes in reason no. 1.7.1 that the patent did not disclose information to answer the following questions with regard to the training:
- For which specific application situations should the computational model be trained?
- How to determine the range of necessary variations of the input variables in the training data?
- Where to obtain the training data from?
How to Do It Better
AI patent applicants are advised to explain in the patent description:
- The quantity and quality of training data required for the invention to function successfully.
- The scope of input variable variations needed in the training data, considering the intended application.
- How to obtain the necessary training data.
Key Takeaway
Transparency in training is essential: Provide a clear and comprehensive description of the training process and the training data used.
4) Plausibility
The decision criticizes that the patent did not disclose a concrete, workable example which makes it plausible that the claimed results can be achieved in principle (“mit dem zumindest plausibel gemacht wird, dass bzw. inwieweit sich damit die beanspruchten Ziele … grundsätzlich erreichen lassen”; Reasons No. 1.8) is disclosed.
What the EPO Decision Says
A core concern in reason no. 1.7.5 of the decision was that during normal operation of a steel plant, the aim is to keep the process parameters as constant as possible in order to achieve reproducible results. It was also not common to vary the initial refractory lining of the inner vessel or the maintenance procedure.
Therefore, the data obtained during normal operation contains no or only slight random variations in the parameters mentioned in the claim. In this context, the decision criticizes that the patent does not not plausibly demonstrate that the calculation model can be successfully trained with such a limited training data set, in particular that or how the learning of artifacts can be prevented by random correlations of the only slightly fluctuating input variables with the wear.
How to Do It Better
AI patent applicants are well advised to provide at least one concrete, workable example that demonstrates in principle how the invention can achieve the desired result. In particular:
- Include supporting evidence: Present data from experiments and tests that validate the performance of your AI model.
- Address potential limitations: If there are constraints or conditions that might affect the AI model’s performance, acknowledge and explain them.
- Consider the claim scope: The broader your claims, the more evidence you’ll need to provide to demonstrate plausibility across the entire claimed scope (Reasons No. 1.6.6).
Key Takeaway
Evidence is key: Plausibility is about showing, not just telling. Back up your claims with concrete examples and evidence to demonstrate that your AI invention is not just a theoretical concept but a practical solution.
Conclusion
You’re well advised to diligently address as many of the five criteria as possible when describing the inner workings of the AI system in a patent document. To make it easy for you, I’ve compiled the key questions into a free cheat sheet to guide you through the disclosure process. Sign up to my mailing list and I’ll send it to you.
Hope it helps!
Bastian