Artificial intelligence and machine learning: sufficiency and plausibility

Wednesday 12 June 2019

Computer-implemented inventions

At the European Patent Office (EPO), examination of computer-implemented inventions (CIIs) is well-established. A CII is defined by the EPO as one which involves use of a computer, computer network or other programmable apparatus (generally, a computer), in which one or more features are realised wholly or partly by means of a computer program.

Artificial Intelligence (AI) encompasses computers that exhibit behaviours perceived as intelligent by humans, including learning, reasoning, inferring and decision-making. Machine Learning (ML), a class of AI, gives the computer an ability to change behaviour according to experience. In other words, the computer learns.

The EPO examines inventions based on AI and ML as CIIs. The Guidelines for Examination (revised November 2018) include illustrative examples for examination of AI and ML by the Examining Divisions and the Opposition Divisions. While the Case Law of the Boards of Appeal is, as yet, limited regarding AI and ML, the extensive corpus of case law on CIIs is expected to apply similarly.

The Guidelines for Examination F-III, 12, note that if the claimed invention lacks reproducibility, this may become relevant under the requirements of sufficiency of disclosure or inventive step. If the claimed invention lacks reproducibility because its desired technical effect, as expressed in the claim, is not achieved, this results in a lack of sufficient disclosure, contrary to Article 83 EPC. Otherwise, if the technical effect is not expressed in the claim but is part of the problem to be solved, there is a problem of inventive step, contrary to Article 56 EPC.

How does this apply to inventions based on AI and ML?


Article 83 EPC requires that an invention is to be described sufficiently clearly and completely to be carried out by a person skilled in the art. This may be satisfied by a detailed description of at least one example of carrying out the claimed invention, though description of well-known features may be omitted. However, where the claimed scope is broad, a single example may not suffice. Lack of sufficiency cannot be remedied after the date of filing, leading to refusal of the patent application or revocation of the granted patent: lack of sufficiency is a ground of opposition (Article 100(b) EPC). Though not commonly encountered, a lack of reproducibility may also be contrary to Article 57 EPC, requiring the invention to be susceptible of industrial application (Case Law of the Boards of Appeal I.E.2).

Objections related to sufficiency are more familiar in the technical fields of chemistry and biology, for example. In contrast, such objections are atypical in mechanical and electrical technical fields: an objection of lack of sufficient disclosure presupposes that there are serious doubts, substantiated by verifiable facts (Guidelines for Examination F-III, 1).

For CIIs generally, the Guidelines for Examination suggest that a clear description of function may be more appropriate than an over-detailed description of structure, to meet the requirements of Article 83 EPC.

AI and ML, however, are not CIIs generally. Particularly, where the computer learns, the behaviour and hence a description of the computer is dynamic until training is terminated (contrast: completed) and is likely to be unpredictable. Indeed, if the behaviour of the computer is static or predictable, the invention may be instead preferably claimed as a conventional CII rather than as ML. How is such dynamic or unpredictable behaviour relevant to sufficiency? How may the trained ML be described? Should the training be described? More simply, how should the claimed invention be described such that it may be carried out by the person skilled in the art?

At the EPO Conference on Artificial Intelligence (05 June 2018), sufficiency was merely noted as a general requirement, together with clarity.

Such questions, under a title of reproducibility, have been debated by leading academics in the field of ML. At The Posner Lecture at the Neural Information Processing Systems Conference (NeurIPS) 2018, Joelle Pineau proposed a reproducibility checklist for academic papers in this field. This reproducibility checklist has since been adopted, as from 02 April 2019, as a mandatory requirement for papers submitted to NeurIPS 2019.

Reproducibility checklist (after Joelle Pineau 2018)

A. For all algorithms presented, check if you include:

  • A clear description of the algorithm;
  • An analysis of the complexity (time, space, sample size) of the algorithm; and
  • A link to downloadable source code, including all dependencies.

B. For any theoretical claim, check if you include:

  • A statement of the result;
  • A clear explanation of any assumptions; and
  • A complete proof of the claim.

C. For all figures and tables that present empirical results, check if you include:

  • A complete description of the data collection process, including sample size;
  • A link to a downloadable version of the dataset or simulation environment;
  • An explanation of how sample were allocated for training / validation / testing;
  • An explanation of any data that were excluded;
  • The range of hyper-parameters considered, method to select the best hyper-generator configuration and specification of all hyper-parameters used to generate results;
  • The exact number of evaluation runs;
  • A description of how experiments were run;
  • A clear definition of the specific measure or statistics used to report results;
  • Clearly defined error bars;
  • A description of results including central tendency (e.g. mean) and variation (standard deviation); and
  • The computing infrastructure used.

While the links to the downloadable versions of the source code and the dataset or simulation environment may be contentious, the remaining requirements of the checklist effectively summarise what should be considered at least as future best practice – if not best practice now.

Indeed, many of the remaining requirements relating to figures and tables are already familiar in the technical fields of chemistry and biology. The Guidelines for Examination are comprehensive and the Case Law of the Boards of Appeal is well-established regarding sufficiency in these technical fields – much of which may be readily transposed to inventions based on AI and ML.

Nevertheless, downloadable versions available from a depository could, in future, become a requirement – as analogously required for micro-organisms under the Budapest Treaty on the International Recognition of the Deposit of Microorganisms for the Purposes of Patent Procedure. Hence, making the ML datasets available now at the time of filing, or using publicly-available datasets, safeguards patent applications in the event that public availability of the datasets becomes a requirement – lack of sufficiency cannot be remedied.

Where disclosure of the dataset is not desired, such as for confidential or proprietary data, complete details of how the data were acquired should be included in the description. In this way, the skilled person may recreate the dataset. Again, such description is already familiar in the technical fields of chemistry and biology. Many patent applications for inventions based on AI and ML are likely to be multi-disciplinary, involving scientists or engineers, data scientists and computer scientists. Hence, the skilled person may be a team of such scientists and engineers, so the descriptions should meet the sufficiency requirements with respect to each of these.

But is the reproducibility checklist not merely a wish list? No.

On 30 January 2019, the Japan Patent Office (JPO) moved decisively towards such sufficiency requirements for inventions based on AI and ML, as set out in Annex A of the Japanese Patent Examination Handbook. ‘Determination on the Description Requirements for the Description and Claims’ presents six case examples (Cases 46 – 51) that explore the JPO essential description requirements of:

  • the condition where it can be recognized that there is a certain relation such as a correlation among the multiple types of data based on the disclosure in the description, or
  • the condition where it can be presumed that there is a certain relation such as a correlation among the multiple types of data in view of a common general technical knowledge.

That is, a correlation between a training dataset and an output should be explicitly identified in the application as filed, or test results (or other proof of validation), of a resulting model presented therein ‘unless an estimation result by AI can be a substitution for an evaluation on a product that has actually been made.’

So, if Japanese patent applications are planned, these JPO requirements must be met now. Certainly, the EPO will be considering how to evolve its requirements. Applying future best practice now, such as according to the reproducibility checklist, is not too soon.


In T 939/92, the Board of Appeal considered plausibility under inventive step for the first time. In this agrochemical case, the Board of Appeal noted that the application provided data for only a few compounds, to support a technical effect, such that it was not credible (plausible) that the technical effect was achieved by all the compounds within the scope of the claim. The application was refused as being contrary to Article 56 EPC, lacking an inventive step across the whole claimed scope. T 939/92 contains fundamental rulings on broad claims in the field of chemistry (Case Law of the Boards of Appeal I.D.9.8.3) – which may also be readily transposed to inventions based on AI and ML.

The case law regarding plausibility is now well-established.

In contrast to the requirements for sufficiency, post-published evidence may support arguments for plausibility. In T 1642/07, the Board of Appeal accepted post-published evidence as confirmation of a technical effect already made plausible based on a convincing theoretical explanation in the application as filed. The Board of Appeal in T 433/05 took a similar view. Particularly, for post-published evidence to be taken into account, it must be established whether or not the claimed invention was made sufficiently plausible at the effective date, based on the patent application as filed and the common general knowledge of the skilled person.

However, in T 0488/16, the Board of Appeal decided that post-published evidence could not be taken into account since the technical effect relied on had not been made plausible in the application by way of any data in the application, or similar information available to the skilled person as common general knowledge. The Board of Appeal reasoned that if the nature of the invention is such that it relies on a technical effect, which is neither self-evident, nor predictable or based on a conclusive theoretical concept, at least some technical evidence is required in the application to show that a technical problem has indeed been solved. At issue was not the absence of any in vivo or clinical data, but rather the absence of any verifiable data with regard to the stated technical effect.

Again, the reproducibility checklist addresses potential issues regarding plausibility, even if plausibility has not, as yet, become at issue at the EPO for inventions based on AI and ML – but it may become an issue. Considering these decisions regarding plausibility is also a reminder regarding overly broad claims, particularly where fall backs are limited or evidence rare.

Here, again, the JPO has moved decisively towards such plausibility requirements for inventions based on AI and ML, as also set out in Annex A of the Japanese Patent Examination Handbook, optionally in combination with the evidential requirements outlined under sufficiency. ‘Determination on the Inventive Step’ presents four case examples (Cases 33 – 36) that offer helpful guidance. Example 34 relates to a hydroelectric power generation and explains:

Generally, an input of data of which correlation is unknown may cause a noise in machine learning. However, the invention of Claim 2 uses an input data containing a temperature of the upper stream of the river during a predetermined period between a reference time and a predetermined time before the reference time. This enables a highly accurate estimation of a hydroelectric power generating capacity, taking an increase of inflow rate due to melt water in the spring into consideration. It is a significant effect that a person skilled in the art cannot expect.

Accordingly, it not considered to be a mere workshop modification that can be carried out in application of the well-known art … by a person skilled in the art to contain, in an input data in an estimation of a hydroelectric power generating capacity, a temperature of the upper stream of the river during a predetermined period between a reference time and a predetermined time before the reference time.

Again, if Japanese patent applications are planned, these JPO requirements must be met now. Certainly, the EPO will be considering how to evolve its requirements. Applying future best practice now, such as according to the reproducibility checklist, is not too soon.


For applicants, patentees and opponents alike, these developments help assess whether a claimed invention, that uses AI or ML, may succeed or fail. Where AI or ML is used, the application should comprehensively describe the detailed implementation, functionally and structurally, together with a theoretical explanation, explicit correlation and/or experimental evidence. The reproducibility checklist is a practical starting point, whether with a view towards the JPO requirements for patentability or anticipating how practice at the EPO may evolve. Applying future best practice now is not too soon.

At Appleyard Lees we have a large and experienced team of attorneys who combine a broad range of legal and technological expertise with real world commercial insight.  If you need help with navigating the complex world of intellectual property, contact us here.

By continuing to use this site, you agree to the use of cookies. more information

The cookie settings on this website are set to "allow cookies" to give you the best browsing experience possible. If you continue to use this website without changing your cookie settings or you click "Accept" below then you are consenting to this.