The rapid rise of artificial intelligence naturally leads businesses to question whether solutions incorporating AI can be protected by patents. These questions are all the more legitimate given that AI is profoundly transforming numerous technical fields and that the architectures employed are evolving at a particularly rapid pace.
Nevertheless, the presence of an artificial intelligence engine in an invention does not call for the application of a new patentability framework. An invention incorporating AI remains, for the most part, a computer-implemented invention. Its patentability must therefore be assessed according to the established principles applicable to such inventions: identifying a technical problem, the solution provided to that problem and the features that actually contribute to achieving a technical effect.
Artificial intelligence does, however, have certain specific characteristics. First, an AI engine considered in isolation generally falls within the category of mathematical methods. Second, its operational function generally results from a training process that must be described in sufficient detail. Finally, the sometimes difficult-to-explain behaviour of a trained model may raise particular questions regarding the credibility of the technical effect relied upon.
These characteristics do not establish a separate legal framework. Rather, they require adjustments to the way in which the invention is identified, analysed and described.
1. An invention incorporating AI remains a computer-implemented invention
The case law of the Boards of Appeal holds, in particular, that a neural network, when considered as such, defines a class of mathematical functions. Decision T 702/20 states this particularly clearly: a neural network defines a class of mathematical functions which, considered as such, falls within excluded subject-matter. However, the Board immediately clarifies that such a network may contribute to inventive step when it helps solve a technical problem, for example when it is trained on specific data for a particular technical task.
Decision T 1952/21 adopts the same approach in relation to reinforcement learning. The Board declines to regard reinforcement learning as inherently technical: the technical character must be found in the invention’s actual contribution, rather than merely in the classification of the AI method employed.
Fundamentally, this approach is not very different from that applicable to more traditional software. An abstract algorithmic innovation does not become patentable merely because it is executed by a computer. Similarly, an innovation relating to an AI model does not become patentable merely because it involves recent or sophisticated technology. In both cases, the decisive question remains whether it contributes to solving a technical problem.
Recent UK case law also illustrates a convergence of approaches. In its judgment in Emotional Perception AI Ltd v Comptroller General of Patents, delivered on 11 February 2026, the UK Supreme Court revised the approach previously followed in the United Kingdom regarding the exclusion of computer programs and moved significantly closer to the reasoning developed by the EPO. In particular, the judgment confirms that an artificial neural network may be regarded as a computer program, without its inclusion in a claim comprising technical means automatically resulting in the exclusion of the claimed subject-matter.
The legal frameworks remain different across jurisdictions. Nevertheless, their recent evolution reveals increasing practical convergence for inventions with a clearly identifiable technical purpose.
2. The invention does not necessarily lie in the AI engine
A common difficulty is to equate “AI” with “the invention” too readily. However, the artificial intelligence engine may be just one software component among others within a much more complex architecture.
It is first necessary to distinguish between the architecture of a model and the model that is actually operational. A neural network architecture may be entirely known. When its parameters are determined through training, it is this training that gives it the function sought in the application concerned. The fact that an architecture is known therefore prejudges neither the use to which it will be put nor the patentability of the solution incorporating it.
Conversely, simply replacing conventional processing with a neural network is not necessarily sufficient to establish an inventive step. In case T 1526/22, the Board of Appeal considered it obvious, in the circumstances at issue, to replace a known physical model with an appropriately trained neural network, given the technological developments already underway at the relevant date.
The inventive contribution may, in fact, lie at different levels.
It may lie in the choice of data used. Data whose relevance to solving the technical problem was not obvious may constitute an essential element of the solution. However, a distinction must be drawn between the mere discovery of a correlation, which is not in itself a patentable invention, and the use of that correlation in a specific technical solution. The discovery that a previously overlooked variable makes it possible to detect or predict a technical condition may therefore be decisive when that information is actually exploited by the claimed system.
The invention may also lie in the preparation or combination of data, in the way an engine is trained, in the use of its output or in its interaction with other software or hardware components.
Finally, an invention may incorporate several artificial intelligence engines. These may process different data, perform distinct functions, produce intermediate representations or cooperate within an architecture adapted to the technical problem concerned. In such cases, the inventive contribution may lie in the allocation of functions between the engines, their interactions, the flow of data or the combination of their results. It would then be artificial to isolate each engine in order to determine whether it is individually novel or inventive. It is the claimed technical architecture as a whole that must be analysed.
3. Training cannot be left in a “black box”
The main practical specificity of inventions incorporating AI often emerges at the drafting stage.
An engine whose parameters must be determined through training does not spontaneously perform the function attributed to it in the invention. The description cannot therefore be limited to stating that an “artificial intelligence engine is configured to determine” a given result, without providing the information necessary to understand how it acquires that capability.
This issue was examined particularly clearly in decision T 161/18. The Board considered the invention insufficiently disclosed because the training of the neural network, which was necessary to achieve the claimed function, had not been adequately taught.
Decision T 1669/21 provides another significant illustration. The claim covered a particularly broad method for determining the condition of a refractory lining using a computational model. The Board noted the discrepancy between the scope of the claimed invention and the level of detail provided in the disclosure. In particular, the patent did not provide any concrete, reproducible example or sufficient information to identify the relevant parameters and implement an appropriate computational model.
However, these decisions should not be interpreted as establishing a fixed list of information that must systematically be disclosed for every machine learning invention.
Decision T 48/24, issued on 13 November 2025, provides an important clarification in this respect. The Board expressly states that there is no reason to treat machine learning inventions differently from other inventions when assessing sufficiency of disclosure. The implementation of the model, its training and its ability to produce the claimed output may naturally be important elements of the assessment, but the level of disclosure required depends on the circumstances of each case, particularly the common general knowledge of the skilled person.
This case law does not mean that the source code, network weights or entire training dataset must systematically be disclosed. The requirement concerns the information necessary to enable the skilled person to carry out the invention across the claimed scope without undue burden.
In practice, similar reasoning may be applied when a proprietary or commercially available pre-trained model is used as a component of the invention. The situation is not fundamentally different from that of an off-the-shelf software module. It is not necessary to know all its internal details, but the application must describe the properties that make this component suitable for the intended function. A commercial product may usefully be cited as an example of an embodiment, without the protection having to be limited to that particular product.
4. The technical effect must be sufficiently credible
A second question arises when the link between the means described and the claimed technical effect is not immediately apparent.
In many mechanical inventions, the relationship between the structure described and the result obtained is directly understandable. The same is often true of software implementing deterministic processing, where the consequences can be deduced from the algorithm described.
The operation of a trained model may, in some cases, be less directly predictable. It may therefore be useful to provide evidence supporting the reality of the technical effect relied upon.
This requirement certainly does not mean that an invention incorporating AI must systematically be accompanied by an experimental programme comparable to those encountered in certain fields of chemistry or pharmaceuticals.
Decision T 48/24 is again instructive in this respect. The Board noted that the patent in question contained no concrete and reproducible example demonstrating that a trained model could actually produce the claimed result, nor any indication of the accuracy achieved. Nevertheless, it clarified that such an example is not an absolute requirement: it is necessary only insofar as the general disclosure and the skilled person’s knowledge are insufficient to enable the invention to be carried out.
The credibility of the effect may therefore be established through various elements: a convincing technical explanation, examples of operation, performance indicators, success rates or, where useful, comparative results. A comparison between an embodiment implementing the claimed feature and one without that feature may be particularly relevant when it establishes that the alleged technical effect is indeed attributable to that feature.
Decision G 2/21 of the Enlarged Board of Appeal also emphasises the importance of the technical teaching contained in the application at the filing date. Evidence submitted subsequently cannot be disregarded solely because it was produced after filing. However, the applicant or patent proprietor may rely on the technical effect supported by such evidence only if the skilled person, having regard to their common general knowledge and on the basis of the application as filed, understands that effect as being encompassed by the technical teaching and embodied by the same invention originally disclosed.
It is therefore prudent, even when all experimental results are not yet available, to identify and explain, at the filing stage, the technical effect on which inventive step may be based.
5. Training and credibility of the technical effect: two converging issues
The sufficiency of the description of training and the credibility of the technical effect are often presented as two distinct issues. Yet they are closely linked.
Training is precisely the mechanism through which a model architecture acquires the technical function attributed to it in the invention. Describing this training therefore serves not only to enable the skilled person to reproduce the trained model. It also helps explain why that model is capable of achieving the intended technical result.
The sequence can be summarised as follows: the training data and the objective assigned to the model determine the capability acquired by the trained model; this capability then enables the model to perform a technical function within the system; and this function ultimately contributes to achieving the technical effect relied upon.
A sufficiently precise description of the relevant data, the model’s inputs and outputs, the training objective and, where necessary, the training methods thus makes it possible to understand the link between the means implemented and the stated effect.
From this perspective, the description of training plays, in certain AI-based inventions, a role similar to that of explaining a causal chain in a more conventional invention. It explains why the means described are actually capable of producing the intended effect.
In practice, the more clearly this link emerges from the application, the more likely the description itself will be sufficient to establish the credibility of the technical effect, without the need to support it with detailed experimental results. Conversely, where the function acquired by the model remains difficult to deduce from its training, a limited number of results may usefully supplement the description and support the existence of the technical effect.
The issues of sufficiency and credibility are therefore not legally identical. However, in drafting practice, they converge around the same requirement: explaining how the model becomes capable of performing the technical function on which the invention relies.
6. An ultimately very conventional approach
Artificial intelligence therefore does not require a reconsideration of the fundamental principles governing the patentability of computer-implemented inventions.
The AI engine may be one element of the solution, sometimes central, sometimes secondary. The invention may lie in the data supplied to it, the identification of a non-obvious correlation, its training, its integration into a complex architecture, the interaction between several engines or the technical use of the results produced.
The analysis must therefore focus on the technical architecture as a whole and on the features that actually contribute to solving the technical problem concerned.
The main area requiring particular attention ultimately concerns drafting. Where the function attributed to the engine in the invention results from training, the application cannot simply refer generically to an “artificial intelligence engine”. It must provide, taking into account the skilled person’s common general knowledge and the role played by that engine, the information necessary to understand how it acquires or possesses the technical capability attributed to it. The application must provide the information necessary to understand how this function is acquired and, consequently, why it is capable of producing the technical effect relied upon.
AI therefore does not create a new category of patentable inventions. Above all, it requires the rigorous application of already familiar principles: identifying the technical contribution, sufficiently describing the means that make it possible and credibly establishing the link between those means and the intended effect.
October 2026
Contributing author
