This line of questioning ignores an important aspect of the EPO’s approach to determining non-obviousness. The EPO does not consider how the invention was made, nor is there any obligation under the European Patent Convention to disclose how the invention was made, including whether AI was used. This is because the obviousness test is an objective assessment, not a subjective one.

The question also neglects one of the considerations underlying the principle that an AI cannot be an inventor: the research leading to an invention, even using AI, always involves human input.

For now, the inventiveness of an invention derived via AI is assessed in the same way as any other invention, and so, such inventions are not inevitably obvious.

The EPO’s inventive step assessment is from the perspective of the notional skilled person, a legal fiction equipped with the common general knowledge of the relevant technical field, and capable only of routine experimentation using routine tools.

While researchers in pharma and biotech companies may create, train, or refine bespoke AI models as part of their inventive activity, the skilled person would generally be limited to using publicly available, routine AI tools. This is important because the AI systems used by inventors are often not the routine AI systems available to the skilled person.

Although certain public tools (e.g., AlphaFold, Claude) could now be regarded as routine, many of the most powerful models used in drug discovery are proprietary systems developed and deployed in-house. Consequently, the existence and use of such models is largely irrelevant to the obviousness assessment because they are not readily available and their use is therefore highly unlikely to be considered routine to the notional skilled person.

Even where a public AI model could have generated the invention, that does not necessarily make the invention obvious. The EPO, when assessing obviousness, asks whether the skilled person would have followed a particular route with a reasonable expectation of success – not whether they could have done so. This is especially relevant in the life sciences, where researchers routinely face vast numbers of potential targets, molecules, and experimental approaches.

For instance, AI may help identify promising candidates more efficiently, but inventors will still generally need to make critical decisions about which outputs to pursue, how to interpret them, and which experiments to conduct. An AI-generated suggestion, whether a protein structure, drug target or candidate molecule, does not automatically provide a reasonable expectation that, for example, a particular therapeutic strategy will be successful – at least not for currently available AI models. Where substantial judgement, interpretation, and follow-up experimentation are required, such non-routine elements may underpin inventiveness.

In life sciences, this means that patent applications in this field tend to include experimental data – typically from in vitro assays, but often also from in vivo animal models – showing that a drug works in the way the inventors claim. For instance, if the invention is a new antibody, the patent application might include data showing that the antibody binds to a specific receptor and has a specific function, e.g., it inhibits signalling, or prevents binding to a ligand. If the invention is a new medical use, then data showing that the drug has an effect on a biological pathway that is relevant to the disease in question might be provided, and there may also be preclinical data from studies in mice.

As AI becomes more heavily integrated into early-stage drug discovery, there is increasing interest in whether AI-generated predictions can replace some or all of the experimental data that applicants have traditionally provided. For now, however, our view is that AI-generated data is unlikely to be regarded by the EPO as a viable substitute for experimental evidence of a therapeutic effect.

Fundamentally, AI model outputs are statistical predictions that depend heavily on the quality of training data, model assumptions, and underlying biases. Many models provide limited insight into how an output has been reached. While a mechanistic explanation is not always required, the absence of a clear rationale for why a predicted therapeutic effect is expected to occur may make it challenging to persuade the EPO that the prediction is credible.

As a result, the EPO is likely to continue placing significant weight on experimental verification, particularly in therapeutic applications where real-world biological effects are central to the invention. Good-quality in vitro and in vivo data therefore remain important cornerstones of any new life sciences patent application.

This does not mean that AI evidence is irrelevant. On the contrary, the EPO’s principle of free evaluation of evidence means AI-generated data may increasingly form part of the package of evidence supporting a patent application. Over time, as AI models become better understood and ever more widely used, it is possible that the evidential value of such material will increase. Nevertheless, for the foreseeable future, applicants are advised to support AI-generated inventions with conventional experimental data wherever possible.

These considerations should not be taken to suggest that the use of AI in R&D is inherently detrimental from a patentability perspective. Avoiding AI is unlikely to be a realistic option for many companies, nor would doing so necessarily provide any patent advantage. There is generally no reason to avoid using AI out of concern that doing so will inevitably make an invention obvious. Equally, there is no general requirement under European patent law to disclose that AI was used during the inventive process.

In any event, where AI has played a significant role in R&D, applicants should identify and document which aspects of the work were genuinely non-routine. Relevant factors could include the selection and training of the model, the design of inputs, the interpretation of outputs, and the experimental work required to validate the results. These human-led decisions may help explain why the skilled person would not simply have arrived at the invention as a matter of course, even if a powerful AI model was within the skilled person’s arsenal (which is unlikely to be the case for proprietary models).

Applicants should continue to prioritise including robust experimental support in new patent applications, particularly where the invention is based on a new therapeutic use or strategy. Such data often underpins both inventiveness and sufficiency and is therefore central to many life sciences patent cases. Poor-quality or insufficient data can be fatal, whereas strong data showing a surprising effect can prove decisive in high-value cases.