30 November 2023

The proof is in the patent – on the question of experimental data in AI-related patent applications before the EPO

The recent Enlarged Board of Appeal decision G2/21 has brought the spotlight back onto the topic of the use of experimental data for providing proof of a technical effect in European patent applications. Whilst the use of experimental data is a mainstay of Chemistry and Life Sciences (CLS) applications, it has typically not been a feature of the European patent process outside of CLS. However, recent decisions issued by the Boards of Appeal show a potentially growing need for certain types of AI-related applications to include some form of experimental data.

The general need for experimental data

A European patent can only be granted if an invention is new and involves an “inventive step”. The European Patent Office (EPO) uses what is known as the “problem and solution approach” for the assessment of whether an invention involves an inventive step. The decision of G2/21 sets out the steps of the “problem and solution approach” as:

“(a) identifying the “closest prior art”;

(b) comparing the subject-matter of the claim at issue with the disclosure of the closest prior art and identifying the difference(s) between both;

(c) determining the technical effect(s) or result(s) achieved by and linked to these difference(s);

(d) defining the technical problem to be solved as the object of the invention to achieve these effect(s) or result(s); and

(e) examining whether or not a skilled person, having regard to the state of the art within the meaning of Article 54(2) EPC, would have suggested the claimed technical features in order to obtain the results achieved by the claimed invention.” (paragraph 24, G2/21)

The Enlarged Board of Appeal in G2/21 further explains:

“Step (c)…requires that, in order to determine the objective technical problem, the technical results and effects achieved by the claimed invention as compared with the closest prior art must be assessed. According to the established case law of the boards of appeal (see CLB, 10th edition, I.D.4.2, and the decisions therein) it rests with the patent applicant or proprietor to properly demonstrate that the purported advantages of the claimed invention have successfully been achieved.” (paragraph 26, G2/21)

In other words, it is up to the applicant or proprietor to prove that the differences over the closest prior art achieve some technical advantage. Depending on the particular field of the invention, this may involve providing experimental data as evidence of the existence of a technical effect. For example, in general it cannot be determined whether a particular drug would provide a therapeutic effect for a particular disease simply by looking at its structure on paper. Hence, experimental evidence is normally required to prove that a drug would have a beneficial effect on a particular disease. For a mechanical type invention however, the technical advantage is normally derivable from how the invention works on paper and our understanding of physics whether or not a technical effect is achieved and hence, experimental evidence is not normally necessary.

Turning back to the question of AI-related inventions, given the black-box aspect of AI, it can be difficult to determine on paper whether one AI algorithm is necessarily better than another AI algorithm for a given problem. In this case, there can be a potential need for experimental data to provide evidence of an improvement. This is exactly the case in a recent Board of Appeal decision, T2803/18, discussed next.

AI-related decisions

The case of T2803/18 was an appeal against a decision by the Opposition Division to maintain European Patent No. EP2582341 in amended form.

The patent in question related to monitoring of moisture in absorbent articles arising from incontinence. The method involved applying clustering to estimate the volume of exudates in a wetness event. The patentee had attempted to argue that the use of clustering provided a technical effect of more accurate estimation over the closest prior art method. The Board of Appeal was sceptical as to whether there was indeed an improvement over the prior art method, remarking in the Decision:

“The Board is not convinced by the respondent’s argument that the accuracy of the estimation would be increased. The accuracy would depend on many factors (size of training sets, number and type of elements/variables constituting the representative vectors, etc.), none of which are defined in claim 1, so that the results obtained by the claimed method are not necessarily more accurate than the results obtained by the regression analysis, the resulting mathematical model and the threshold criteria applied in D2. The patent in suit does also not support such an alleged benefit by comparative data.” (reasons 4.2, T2803/18)

As discussed above, the onus is on the patentee to prove the existence of a technical effect. In this case, the Board doubted that an improvement existed over the closest prior art and there was no further evidence to back-up the patentee’s arguments. As such, the Board remained unconvinced and considered the invention to be an obvious alternative to the closest prior art and the patent was subsequently revoked.

In a similar vein, we have a second case, T2147/16, which related to spam filtering and again, the use of clustering. This case was an appeal against a decision to refuse the grant of a patent. In this case, the applicant attempted to argue (amongst others) the technical effect of reduced processor load on an end user computer. Once again, the Board was sceptical over the technical effect being argued, remarking:

“The present invention, page 19, lines 1 to 8, discloses that the cluster size has to be optimised in order to reduce the load on the computer. However, further details are not provided, for example, the range of the optimal cluster size, relevant parameters, the amount of memory saved or the ratio of increased speed.

The Board is of the opinion that specific details as to how an algorithm is implemented in practice and how the load is reduced must be provided in order to give evidence that the algorithm has any further technical effect with respect to known algorithms and that it provides an improvement over the prior art.” (reasons 5.3.11, 5.3.12, T2147/16)

Once again, we have a Board specifically wanting to see further evidence to prove the existence of a technical advantage without which, the Board did not see fit to grant a patent.

The take-home 

As we have seen with the decisions of T2803/18 and T2147/16, should the EPO be sceptical over a purported technical effect, without experimental data or other evidence to persuade, applications may be rejected and patents revoked.

Those familiar with the problem and solution approach and the keen-eyed among you may have noticed that the technical effect under the problem and solution approach is determined with respect to the claimed invention as compared with what the EPO has identified as the “closest prior art”. Of course, it is not possible to know when drafting a patent application what the EPO would identify as the closest prior art. That however does not mean that including experimental data in a patent application is worthless if it is not a comparison with the closest prior art. Rather, any experimental data tells its own story, of the potential benefits of the invention and under what conditions and circumstances these may arise. This alone could be enough to prevent doubts regarding the invention from forming in the first place and an argument that the invention performs better than the closest prior art could well be accepted without the need for further evidence explicitly comparing against the closest prior art.

With the widespread adoption of AI in many fields, it is increasingly difficult to obtain a patent on the basis of being the first to use algorithm X on problem Y. A clear technical advantage of using algorithm X may be necessary in order to obtain a patent. In this regard, as we have seen, the inclusion of suitable experimental data demonstrating a technical advantage would stand the application in good stead, ready for any potential challenge by the EPO.