20/11/2024
The EPO’s Boards of Appeal have issued a new decision on whether a patent for an artificial intelligence invention was sufficiently disclosed, adding to the growing list of cases on this topic. The EPO’s opposition division revoked EP2789960 for insufficient disclosure and a Board of Appeal confirmed the revocation in T1669/21, published last month.
This decision will not be the last word on the subject and it does not represent a definitive rule on what disclosure is necessary to meet the sufficiency requirements for an AI invention. But it does imply some informal rules that applicants can use to inform their internal invention capturing practices when dealing with AI inventions.
The Board in this case identified a number of aspects where the patent’s disclosure was considered insufficient. It would be tempting to draw from these aspects to create a list of cast-iron requirements for any European patent dealing with AI, but this might be an overreaction at this stage. Firstly, a decision of the EPO’s Boards of Appeal does not automatically create a precedent for other appeal decisions or for EPO examiners. Secondly, this patent is a somewhat unusual case in that the disclosure is very slim and the claims appear vague, at least compared to modern expectations for AI patents. The patent was granted in 2018, i.e. a few years before the Boards of Appeal began seriously questioning sufficiency requirements for AI patents. It is doubtful that the patent would have been granted by examiners assessing sufficiency according to the latest Guidelines for Examination.
Nonetheless, the flaws identified in the patent do offer some guidance to applicants when considering the level of detail to include when drafting a patent application for an AI invention has been reduced to practice, i.e. where the inventor’s contribution to the art is beyond the mere concept of using AI.
The patent concerns the use of AI to determine the condition of the refractory lining of a vessel containing a molten metal. The invention represents the application of AI to a different field, rather than the AI itself. During the appeal process, the proprietor argued that the patent was directed to a ‘comprehensive solution’ to the use of AI in this field. Essentially, the patent’s inventive concept is the principle of using AI in that technical field rather than a field-specific adaptation of known AI technologies. The claims to this ‘comprehensive solution’ are broad and the insufficiency problems arose because patent’s disclosure does not enable this concept across the breadth of the claims.
It is worth noting what the patent did not mention:
Firstly, the patent did not describe the structure of the neural network. The Board found the patent insufficient because it leaves open which topology and class (arrangement of the nodes and their compounds) the neural network should have, how the nodes are mathematically modelled (the linking of their input and output values; propagation and activation function) and which learning method is used.
Secondly, the patent did not precisely define input and output parameters for training or operating the AI. The claims defined the inputs to the AI only in general terms such as “the initial refractory construction of the inner vessel lining” and “production data during use”, rather than specific parameters in the form of numerical values that a neural network could use. The Board also complained that the patent did not include a single ‘workable example’ from which the skilled person could identify important input parameters needed to implement the invention, and which could render the invention plausible.
Thirdly, the patent contained only limited disclosure concerning the training of the neural network. The patent stated that the neural network could be trained using recorded data from normal operation of a metal foundry process. But Board found that normal operation of a metal foundry process would have only minor or accidental variation of control parameters, and thus they were not persuaded that this would result in a trained neural network that could provide a ‘comprehensive solution’ over the whole breadth of the claim.
There is the adage that ‘hard cases make bad law’, that an extreme case is a poor basis for a general law that would cover a wider range of less extreme cases. This patent might be an extreme case of insufficient disclosure. This was not a borderline case where the boundary between sufficient and insufficient disclosure had to be found, and so it does not help us identify where this boundary lies.
But, after the EPO’s first two appeal decisions to find AI patents insufficient, applicants are looking for guidance on how to meet the EPO’s sufficiency requirements and understandably might want to turn the deficiencies of EP2789960 identified by the Board into a drafting checklist.
The relevant statute for sufficiency in Europe is Article 83 EPC, which requires the invention be disclosed “in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art.” A valid patent should allow the skilled person to implement the invention without undue burden, and this is what the EPO will assess when examining compliance with Article 83 EPC. On the other hand, EPO examiners will not actually test that it can be carried out by the person skilled in the art and we would not expect EPO examiners to rely on expert evidence. What matters is whether the patent includes enough detail to satisfy the EPO examiner’s concerns.
Assuming that the invention has been reduced to practice before the application is filed, then the inventor should have enough information to provide a detailed description of how they themselves implemented the invention, which should also be enough to satisfy the EPO examiner’s concerns about whether the skilled person could implement the invention. Those drafting a patent application for an AI invention in particular might help themselves by asking the inventor for specifics of how they have implemented the invention, such as descriptions of:
- the neural network architecture
- the input parameters to the neural network
- how the neural network is trained, such as how to construct a set of training data that will a suitable degree of variation that the neural network’s required functionality is obtained.
US-based drafters might view this as something of a ‘best mode’ requirement, where at least one fully workable example based on the inventor’s implementation is disclosed in a patent specification. Ideally, such an example would be disclosed at varying levels of detail and abstraction to allow more general claim amendments than the narrowest disclosure.
On the other hand, the deficiencies identified in EP2789960 are case-specific rather than a general legal principle for AI inventions. Whether the absence of a description of the neural network architecture in the patent places an undue burden on the skilled person will depend on the facts of the case. The same is true of the input parameters and how the neural network is trained. Therefore the features by which EP2789960 was considered deficient might not be necessary for all AI inventions. But drafters will likely not hurt their chances by including such detail in patent applications. Thus it may help if drafting attorneys have access to such information from inventors for inclusion in the applications. Since we don’t yet know where the boundary between sufficient and insufficient disclosure lies for AI inventions, erring on over-disclosure is a reasonable precaution for the time being.
Ultimately, this patent contained too little detail. It was revoked for insufficient disclosure before inventive step was assessed but it would also likely have been revoked for lack of inventive step. The claims lack the technical features needed to bring about the purported advantages of the invention. They represent the inclusion of an ‘off the shelf’ or ‘black box’ AI without an application-specific adaptation to solve the refractory lining problem. We have previously explained how patenting the basic use of an ‘off the shelf’ AI tool is difficult in Europe. The additional detail needed to meet the inventive step requirements is similar to the additional detail needed to meet the sufficiency requirements, i.e. by relating the AI more closely to the problem solved by the invention.
This decision of the EPO’s Boards of Appeal shouldn’t fundamentally change the drafting process for an AI invention. It has not shone a light on where the boundary lies between sufficient and insufficient disclosure. Applicants that already follow good drafting principles, where they describe the AI used in their invention and relate it to the problem solved, shouldn’t need to drastically alter their drafting process. But applicants can ease this process by ensuring that specifics of the AI implementation are captured from inventors before drafting has begun, such as via internal invention disclosure practices that cater to AI inventions.
Readers that would like to discuss patenting AI inventions in Europe or practices for capturing AI inventions should get in touch with me or any of my colleagues at Reddie & Grose.
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.
