31/05/2024
We are now encountering insufficiency objections raised by EPO examiners against claims to inventions implemented using artificial intelligence. Where a claim specifies an artificial neural network or machine learning model configured to process a particular input and provide a particular output, EPO examiners are questioning whether the application as a whole teaches the skilled person how to implement such an element that achieves the claimed result.
This can be seen in the following excerpt from the written opinion prepared for an international application by the EPO acting in its capacity as International Searching Authority.
“7.3 Claims 1 and 13 recite various determinations using a machine learning module, which are not clear in that they amount to the expression of a result to be achieved, without specifying the technical features associated with such use, e.g. the properties of the machine learning module.
8 In addition, contrary to the requirements of Article 5 PCT, the description fails to disclose in sufficient detail how a machine learning module is to be used in order to carry out the claimed determinations.”
In other words, these sufficiency objections are not hypothetical. We are seeing them in practice.
The background: from Boards of Appeal to Guidelines for Examination
European patent law requires inventions to be capable of implementation by the skilled person without undue burden. If the implementation is unspecified and left to the reader as a research project, then the EPO will consider the invention insufficiently disclosed.
We previously reported on two decisions from the EPO’s Boards of Appeal where patent applications were found invalid for insufficient disclosure on how to achieve the claimed functionality of the AI features: Artificial intelligence, insufficiency and inventive step: detailed disclosure needed at the EPO.
The EPO has recently updated its Guidelines for Examination to reflect these decisions. The Guidelines are general instructions to EPO examiners rather than rules of law in themselves, so some departure from these Guidelines may be allowable, but applicants should normally expect the EPO examiners to follow the Guidelines.
At section F-III-3, the updated Guidelines refer in particular to AI inventions as an example where insufficient disclosure is found due to undue burden on the skilled person:
“in the field of artificial intelligence if the mathematical methods and the training datasets are disclosed in insufficient detail to reproduce the technical effect over the whole range claimed. Such a lack of detail may result in a disclosure that is more like an invitation to a research programme.“
Similarly, section G-II-3.3.1 of the Guidelines states:
“If the technical effect is dependent on particular characteristics of the training dataset used, those characteristics that are required to reproduce the technical effect must be disclosed unless the skilled person can determine them without undue burden using common general knowledge. However, in general, there is no need to disclose the specific training dataset itself.“
This issue was raised in the earlier appeal decisions. It is now in the EPO’s Guidelines for Examination. It is now being raised by EPO examiners against pending applications. Applicants should not ignore this issue.
Fixing issues before filing
Applicants are not allowed to fix sufficiency issues by adding new matter to the application after filing. The best time to address these issues is before the application is filed.
The purpose of the sufficiency requirements is to avoid ‘armchair inventions’, where a speculative application is filed directed to a desired result without the inventor actually solving the problem of providing the AI with the desired functionality. Assuming the inventor has solved the problem before the application is filed, then applicants should include as much detail as possible on how this is achieved.
Helpfully, the EPO Guidelines tell us “there is no need to disclose the specific training dataset itself.“
It should be sufficient to disclose how to create a training dataset that, when used to train the AI, at least plausibly achieves the claimed functionality. In essence, this could be viewed as a form of procedural generation for the training dataset.
There is no single boilerplate paragraph that can be included in all patent applications for AI inventions that would satisfy this requirement. Any description of a method to generate a training dataset should be written on a per-case basis, considering the problem solved by the AI. Ultimately, this requires engagement with the inventor during the drafting process, asking questions as to how they implemented the invention.
In addition to a detailed description of how the AI is trained, the application should preferably also include a higher-level abstracted version of this description, preferably in claim-like language. It could become necessary during prosecution to amend the claims to specify the training needed to achieve the AI’s functionality, in response to inventive step objections or result-to-be-achieved objections such as the objection set out at section 7.3 of the written opinion excerpt above. It would be better to include basis for such an amendment in claim-like language to avoid overly narrow limitations to the claims or added-matter objections if the amendment uses broad language than the detailed description that is not disclosed in the application as originally filed.
Can we overcome these issues by argument alone?
In some technical fields, the training process to obtain a trained AI model might be implicit to the skilled person based on their common general knowledge.
This is true where the use of AI is well-established in that field, such as in computer vision applications where an AI model is trained to recognize objects in images. A sufficiently large and diverse set of labelled images may be all that is required, with the aim being that the AI model should process an input image and output the correct label or classifier for that image.
This may also be for other AI models that have well-defined inputs and outputs. A suitable training process based on a large and diverse set of training inputs and outputs might be readily apparent to the skilled person.
On the other hand, given current uncertainty as to what will satisfy the EPO examiners on the matter of sufficiency, we would recommend erring on the side of caution and providing as much detail as possible on the AI and how to obtain it in the application as originally filed. This includes both the training process and any special features of the AI model needed to achieve the desired functionality.
But, if the application has already been filed and we do not have the opportunity to fix these issues by amendment, then applicants might reasonably argue along these lines that the skilled person can implement the invention based on their general knowledge. Time will tell whether these arguments will be successful.
Conclusion
Sufficiency objections of this sort for AI inventions are a new development in examining practice at the EPO. We have seen two high-profile refusals by the EPO’s Boards of Appeal, from which we can draw conclusions as to what an insufficiently disclosed patent application looks like. But we do not yet have counterexamples of sufficiently disclosed AI patents to tell us where the line between sufficiency and insufficiency lies.
Hopefully, this line will become more apparent in the coming years as applicants react to these developments and include greater detail in their originally filed applications or as we gain more experience in overcoming these objections by argument alone.
If you have any questions about the sufficiency requirements for AI inventions or would like to discuss AI inventions more generally with European patent attorneys with expertise in this field, please contact your Reddie & Grose attorney for more details.
This article is for general information only. Its content is not a statement of the law on any subject and does not constitute advice. Please contact Reddie & Grose LLP for advice before taking any action in reliance on it.
