AI And Inventive Step: Does Machine-Generated Mean Obvious?

Examining the impact of artificial intelligence on patentability at the EPO

One question we are now hearing regularly is whether the use of AI to generate inventions, particularly in biotech and pharma, risks undermining inventive step. Put simply: if AI can produce a solution, does that make it obvious?

In our recent webinar, we explored this question through the lens of the EPO’s established problem-solution approach. Under that framework, the EPO assesses inventive step by asking what the notional skilled person would have done in view of the prior art. This “would, not could” distinction remains central to the assessment of inventiveness, and is particularly important in the context of AI-assisted inventions: would the skilled person have used AI in the first place? And would they have selected the claimed solution even if they had used AI?

The risk of using, or disclosing, AI is therefore often overstated. First, it is unclear whether AI has reached the threshold of becoming a routine tool that would be used by a skilled person working in the biotech or pharma industry. Secondly, most AI-related disclosures do not contain enough detail to allow a skilled person to recreate the model exactly, particularly given the importance of training data, labelling strategies, and model parameters. Even small differences in these factors can lead to different outputs. As a result, while a skilled person could attempt to reproduce the research approach that led to the invention, it does not follow that they would arrive at the same solution. Likewise, the fact that an AI tool could, in principle, generate a particular result does not mean that the skilled person would have used the tool in the particular way that led to the invention, or would have selected the inventive solution.

There is therefore still considerable scope to argue that the specific path taken by the inventors – whether in setting out to solve a particular problem, choosing a particular model, applying it in a particular way, or interpreting its outputs – was not an obvious one.

What is the skilled person’s expectation of success based on an AI output?

A key issue related to the points above is how AI-generated outputs should be treated when assessing expectation of success. At first glance, one might think that an AI-generated prediction – such as a protein structure, an epitope, or a novel drug target – would point the skilled person directly to a solution. In practice, however, the position is more nuanced. AI outputs are not inherently reliable, and their value depends heavily on the context, e.g. the type of model used, its known limitations, and the nature of the underlying scientific problem.

For example, if an invention relies on an AI-predicted structure or a newly identified biological target, it is far from clear that this alone would give rise to a strong expectation that the approach will work in practice. This is especially so in life sciences, where the EPO typically requires supporting experimental data to establish technical credibility. In that sense, AI outputs are likely to be assessed in a similar way to other forms of in silico data: potentially useful, but probably not decisive on their own.

There may be cases where the lower “try-and-see” threshold comes into play; this can apply where a suggestion can be tested easily and without difficulty. However, whether this applies to AI-generated ideas will depend on the facts. In many cases, AI outputs may not provide a sufficiently concrete or reliable basis for routine testing, particularly where the model cannot explain how it arrived at its conclusions.

Practice points

It is worth stressing that the inventive step analysis at the EPO is objective: it does not depend on how the invention was actually made. Using sophisticated AI tools does not automatically make an invention inventive, but nor does it make it obvious. Equally, avoiding AI will not insulate applicants from any shifts in how the EPO assesses inventive step as AI becomes more widely used and becomes a more “routine” tool in the skilled person’s arsenal.

From a practical perspective, the way AI is used is important. There may be advantages in relying on in-house tools, particularly where these are not routinely available to the skilled person. By contrast, where publicly available models are used, it may be helpful to emphasise any non-standard or creative aspects of their use. More broadly, it remains important to support AI-driven inventions with appropriate real-world experimental data where possible, especially any data showing surprising or unexpected technical effects.

Looking ahead, the core legal test is unlikely to change in the short term. For now, the key message is that the increasing use of AI does not fundamentally alter the inventive step assessment at the EPO, but it does add a layer of complexity that may in time require a shift in how we argue inventiveness in Europe.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.