Patentability of AI inventions – updates to the EPO Guidelines for Examination 2024

The EPO has published a preview of the amended EPO Guidelines for Examination that will enter into force on 1 March 2024. This is available to view on the EPO website here. 

Major updates reflect decisions from the Enlarged Board of Appeal (EBA) in G2/21 (regarding post-published evidence to support technical effect / plausibility) and G1/22 (entitlement to priority). However, there are also several updates particularly relevant to patents in the field of artificial intelligence (AI). These include updates on the assessment of inventive step and sufficiency for AI inventions, as well as clarification on the topic of AI inventorship. The relevant changes and considerations for those working in the field of AI are summarised below. 

Inventive Step and Sufficiency

The amended guidelines clarify the assessment of technical effects stemming from algorithms in AI inventions, with a new passage added to section G-II-3.3.1 under the heading “Artificial Intelligence and Machine Learning”.  The new passage states the following: 

“The technical effect that a machine learning algorithm achieves may be readily apparent or established by explanations, mathematical proof, experimental data or the like. While mere allegations are not enough, comprehensive proof is not required, either. 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 (see also F-III, 3 and G-VII, 5.2)”

This new passage also refers to “particular characteristics of the training dataset used”.  This is expanded upon in another new passage in the amended guidelines under F-III, 3 “Insufficient disclosure”: 

“Another example can be found 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 (see also G-II, 3.3.1)”

In short, these amendments make it clear that any features of the training data set that are necessary for reproducing the purported technical effect must be disclosed in the application as filed. However, this is only required if these features cannot be derived by the skilled person, without undue burden, using their common general knowledge. The Guidelines do not require the disclosure of the specific training data set itself. However, such disclosure would likely satisfy these requirements where necessary. There is a balance to be struck by patentees here, in disclosing enough information regarding the training data set such that a skilled person can reproduce the invention (and technical effect) across the breadth of the claim, and protecting commercially sensitive information, as the data sets themselves can be very valuable. In some cases, an alternative to disclosing the training data itself might be disclosing learned coefficients/weights of a model​. However, best practices will only become more clear as we see further AI inventions proceed through examination and opposition procedures. 

These changes are aligned with the findings of the Technical Board of Appeal (TBA) in T 0161/18, which found that a lack of information on details of AI training data could lead to an invention being considered insufficiently disclosed.

In T 0161/18, a method for determining cardiac output (11) from an arterial blood pressure curve (7)​ was claimed.  The method used a peripheral blood pressure curve (7) to estimate an equivalent aortic pressure (9), where weighting values used in the estimation are calculated with the help of an artificial neural network (8)​.

The application in question mentioned that training data must cover patients of different ages, sexes, health conditions, etc. to avoid specialisation of the network but gave no further details. The TBA considered that this was not enough, and with no disclosure of more specific parameters or examples of input data, the skilled person could not carry out the training of the network to achieve the purported technical effect. As such, the application did not meet the requirements of Art. 83 EPC (sufficiency). 

In practice, the technical effects of AI algorithms should be mentioned in patent applications, and care should be taken to ensure these effects are both plausible and applicable across the entire scope of the patent claims. If a technical effect is plausible only with respect to specific training data, necessary features of that training data should be included in the patent application. 

These additions to the Guidelines go some way to align the assessment of AI inventions with the disclosure requirements for inventions in the biotech field (as clarified in the G2/21 “Plausibility” decision). As such, we appear to moving in a direction where the Biotech Guidelines and associated case law could be considered analogous for the assessment of AI algorithms. 

Inventorship

A-III, 5.1 has been amended to specify that a designated inventor must be a natural person and A-III, 5.3 has been amended to specify that the EPO will check whether the designated inventor is a natural person. These amendments follow from the decision in J 8/20 (DABUS), where the Legal Board of Appeal found that an AI cannot be designated as an inventor, a decision in line with many other jurisdictions (including most recently the UK Supreme Court). 

Corresponding guidance in the UK 

In what may mark a major shift in UKIPO practice, on 21 November 2023, the High Court handed down its decision in Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks.  The Court found that the UKIPO had erred in finding a neural network implementing a recommendation system as being excluded from patentability as a “program for a computer… as such” under s.1(2)(c) of the Patents Act 1977. The UKIPO has since appealed the decision. 

Following this decision, the UKIPO has suspended its guidance on the examination of AI inventions. In what may signal a shift in approach to the patentability of AI more generally, it has also indicated that the Manual of Patent Practice and the Office’s guidelines for examining patent applications relating to AI inventions will be updated in due course. 

However, the guidance previously stated: 

“…the extent to which a training dataset should itself be disclosed is a matter to be decided by considering each case on its own merits. However, we note the recent decision of the EPO Board of Appeal in T 0161/18. We believe this decision both reflects, and is consistent with, the principles set out in Eli Lilly v Human Genome Sciences…”

We are closely monitoring for further UKIPO updates and guidelines, and it remains to be seen whether the UK will remain aligned with the EPO’s current approach, as set out in the updated Guidelines. 

Conclusion

The amended EPO Guidelines for Examination set clearer standards for the assessment of AI inventions and clarify the need for disclosure of technical effects and the circumstances in which details of the associated training data are necessary to reproduce those effects. Those working in the field of AI should consider the following best practices. 

For patentees:​

• Disclose the structure of AI models, e.g. via a diagram or shorthand​;
• Describe novel components in precise detail;​
• Disclose at least one example of input data and output data; 
• Consider what information about the training data set would be required for the skilled person to reproduce the invention (and technical effect) across the breadth of the claim; and 
• Where necessary, include a description of the way the model is trained and consider including references to training data or learned coefficients/weights of the model​

For opponents: ​

• Consider sufficiency as a possible ground of invalidity against AI patents; and ​
• Begin to consider Biotech Guidelines and case law as analogous