Generative AI and the Problem-Solution Approach: Does the Skilled Person Query a (Historic) Large Language Model When Trying to Solve the Objective Technical Problem?

Patentanwälte

This article was written by the authors using a customized retrieval augmented generation (RAG) system employing source referencing with state-of the art reasoning capabilities. LLMs were run on compliant EU cloud servers. In addition, custom on-premise hardware with archived Open Source LLMs was used.

Abstract

The rapid evolution of generative artificial intelligence (AI), particularly large language models (LLMs), raises fundamental questions for the assessment of inventive step at the European Patent Office (EPO). This article explores whether, under the EPO’s problem-solution approach, the notional skilled person should be assumed to utilize such models when tasked with solving the objective technical problem. The analysis considers the legal framework, the nature of the skilled person, and the potential impact of generative AI on the assessment of inventive step, offering scenarios ranging from no impact to transformative change. The article concludes with an outlook on possible implications for the problem-solution approach and alternative inventive step tests.

1. Introduction

The assessment of inventive step at the European Patent Office (EPO) is anchored in the problem-solution approach, a structured methodology that places the “person skilled in the art” at its heart. This notional figure is tasked with solving the “objective technical problem” derived from the closest prior art, using their common general knowledge and the state of the art as of the priority date (EPO Guidelines 2025 ed., G-VII, 3; 5).

In recent years, Large language models (LLMs) have become ubiquitous due to their enhanced capabilities. LLMs are widely used in different fields, including professional patent work. Consequently, the epi has issued guidelines (https://patentepi.org/en/epi/library/main/538e242d-f1be-46bd-bb34-b948e1544d69/file) for LLM use by patent professionals for client work.

Thus, the question arises: should the skilled person, as a legal fiction, be assumed to query such models when seeking solutions to technical problems? This article examines this question with the Llama 2 family of LLMs released in 2023 as a starting point and discusses the broader implications for inventive step analysis at the EPO.

2. The Problem-Solution Approach and the Skilled Person

The EPO’s problem-solution approach comprises four main steps (EPO Guidelines 2025 ed., G-VII, 5, modified numbering):

Determining the closest prior art
Determining the distinguishing features of the claimed invention
Establishing the objective technical problem
Considering whether the claimed invention, starting from the closest prior art and the objective technical problem, would have been obvious to the skilled person.

The skilled person is a legal construct: “a skilled practitioner in the relevant field of technology who has average knowledge and ability” (EPO Guidelines 2025 ed., G-VII, 3). This person is presumed to have access to the entire state of the art and the means and capacity for routine work and experimentation. The skilled person is not an innovator, but is diligent, methodical, and equipped with the common general knowledge of the field.

3. Generative AI as a Tool for the Skilled Person

3.1. The Emergence and Proliferation of LLMs

The field of generative AI is evolving at a remarkable pace, with new LLMs being released almost weekly. These models are trained using vast corpora of text and are capable of generating, summarizing, and reasoning over technical information. An overview of release dates for open weight LLMs can be found at [eugeneyan/open-llms: A list of open LLMs available for commercial use]. At the time of writing, more than 2 million models are publicly available on the popular repository “Hugging Face”, https://huggingface.co/models. Here, LLMs can be downloaded, archived and run at a later date.

A popular example that starts to become potentially relevant for assessment of inventive step is “Llama 2”, released by Meta in August 2023, which serves as a practical example for this discussion. However, the legal and practical questions raised here apply equally to other LLMs, both current and future. Notably, LLMs typically have a publication date – if released. Further, they have a training cutoff date of the data used when training the model. In the following, we will focus on the publication date.

4. Scenarios for the Impact of Generative AI

The EPO Guidelines recognize that the skilled person’s knowledge and tools evolve over time (EPO Guidelines 2025 ed. G-VII, 3 with reference to T 774/89 and T 817/95). If LLMs are becoming more common but are not yet standard, their use may be considered in some fields but not others. The skilled person may be assumed to be aware of LLMs, but not to rely on them as a primary problem-solving tool in all domains. The following sections discuss different scenarios of LLM use for assessment of inventive step and implications.

4.1. Scenario 1: No Impact

Legal Argument based on the Fictional Nature of the Skilled Person

One could argue that the skilled person, as a legal fiction, does not use real, existing tools. The skilled person is a construct for a thought experiment, not a proxy for actual experimental practice. The assessment of inventive step is not an empirical exercise, but a legal and conceptual one. Thus, output generated from an LLM — such as Llama 2 — should only be considered if it was actually available as part of the state of the art before the relevant date (i.e., before the priority date of the application or patent).

Argument: Hypothetical LLM output is irrelevant

Under this view, it is not sufficient that an LLM could have produced a particular output before the priority date; instead, there must be evidence that such output was actually available to the public before that date. This is consistent with the established legal principle that only information made available to the public before the priority date forms part of the state of the art (EPC Art. 54(2); EPO Guidelines 2025 ed. G-VII, 3.1). The mere theoretical possibility that a model could have generated a solution is not enough. In other words, what an LLM could have potentially generated can be deemed as irrelevant; just as prior art that potentially could have been published by a human writer – but was not.

Argument: No Mixing of Fiction and Experiment

To provide the objective technical problem (which is itself derived from a legal analysis of the prior art and the invention) to an LLM and then use its output as evidence of obviousness would mix the fictional domain of the skilled person with the experimental domain of a real-world tool. Since the objective technical problem can only be determined in view of the closest prior art, introducing a (non-fictional) LLM into this process risks undermining the legal fiction of the skilled person. Doing so would add to the conceptual analysis an empirical, post hoc element into what is meant to be a conceptual analysis.

Case Law Support

Recent EPO Board of Appeal decisions T 1193/23 and T 206/22 have held that the output of LLMs such as ChatGPT does not reflect the skilled person’s understanding, as their responses are context-dependent, based on unknown training data, and not anchored to the knowledge at the priority date. However, at least the last objection would be overcome by using archived versions of LLMs which were available before the priority date, e.g. Llama 2 when assessing inventive step for a patent application filed in September 2023 or later.

Implication

Inventive step is assessed as before; the skilled person does not use LLMs, and the presence of generative AI does not affect the obviousness analysis unless the output itself was made available to the public before the priority date. Under such assumptions, any non-published output by an LLM is irrelevant, as it was not available to the public and the possibility that it could have been created in the past is as irrelevant as a publication that could have been written but was not.

4.2. Scenario 2: Strong Impact

Legal and Practical Implications

In this scenario, the skilled person is assumed to use LLMs as a standard tool. For applications filed after the release of a given model, the skilled person would routinely input the objective technical problem into the LLM and consider its suggestions as part of their problem-solving process.

Including generative AI in inventive step assessment would have profound implications for both the EPO and law firms practicing patent law. It would become necessary to conduct experiments on different checkpoints of LLMs (for example, those available at Huggingface.co https://huggingface.co/models or other repositories) corresponding to the relevant priority date. These checkpoints would need to be preserved and made accessible, and the models would play a key role in assessing inventive step.

Stochasticity and Experimental Parameters

However, this approach is complicated by the fact that LLMs behave stochastically. Querying LLMs (“inference”) is done by providing a seed number (usually, a random number leading to variations in outcome for each run of the model even with identical question) and model parameters such as temperature, top-p, and others which can be arbitrarily configured and govern how the sampling of the actual LLM output is obtained (see How to Tune LLM Parameters for Top Performance: Understanding Temperature, Top K, and Top P | phData for an overview with basic examples how such parameters affect output).

In other words, a vast parameter space is available to generate LLM outputs and re-runs may yield different, potentially contradictory, outputs. As with all empirical sciences, the question arises: what constitutes a “significant” output? How many runs are needed? Must the LLM solve the objective problem with a given probability as in the claimed invention to render it obvious or is a single LLM output sufficient to demonstrate obviousness? What is the threshold for reproducibility or consensus? These are not trivial questions and would require negotiation and standardization, leading to disputes over methodology and interpretation.

Implication

If the notional skilled person is assumed to routinely use LLMs, the bar for inventive step could rise significantly. When an LLM presented with the objective technical problem is capable of outputting the claimed solution, the conclusion may be that the invention is obvious.

This would be different from adopting the broader “MOSITA” (Machine of Ordinary Skill in the Art) paradigm — under which the skilled person is wholesale replaced or continuously augmented by AI (see Fabris, IIC 2020, 685). Instead, injecting an LLM only at a single, well-defined stage of the problem-solution approach would in effect create an “LLM-augmented skilled person”.

On a conceptual level, the attempt to solve the objective technical problem by the skilled person would be transformed from a purely fictional legal consideration to a hybrid of empirical data (experimental LLM outputs) and legal reasoning (the fictional skilled person), in a way comparable to a witness testimony or an expert report.

If accepted, numerous practical questions arise, namely which questions are asked to the LLM, may they be reformulated, how many outputs are generated, what happens if the outputs are contradictory etc. Notably, a key question would be: who runs the LLM experiments? The European Patent Office, each party (for example in opposition proceedings), an independent expert or all of the above?

4.3. Scenario 3: Medium Impact

LLM output as hybrids between legal argument and evidence

In this intermediate scenario, a fictional interaction of the skilled person with an LLM could be considered analogous to expert opinions or witness testimonies. In inter-partes proceedings, such evidence could be provided by the parties legally assessed by the EPO.

If the skilled person could have used an LLM, an argument might be required whether or not the skilled person would have used an LLM in the case at hand (similar to the could/would approach). Consequently, the EPO might include LLM assessments for inventive step as well.

5. Discussion: Legal and Practical Considerations

5.1. The Nature of LLM Outputs

The EPO Boards of Appeal have cautioned against equating LLM outputs with the skilled person’s understanding (T 1193/23, T 206/22). LLMs generate responses based on vast, but not always transparent, training data, and their outputs may not reflect the knowledge or reasoning of a skilled practitioner at the priority date. Moreover, LLMs may incorporate information published after the priority date, which is inadmissible for inventive step analysis.

5.2. The Evolution of Common General Knowledge

For LLMs to be considered part of the skilled person’s toolkit, their use must become established as common general knowledge in the relevant field. This is a factual question, to be determined by evidence (e.g., textbooks, industry standards, expert testimony). Until such time, a discussion may be needed whether or not the skilled person would have used an available LLM or not.

5.3. Enabling Disclosure by the LLM?

A further question is whether everything an LLM “spits out” actually amounts to an enabling disclosure. Some LLM output may meet these requirements, other output may not meet this requirement. LLMs are notorious for “hallucinations”, i.e. plausible sounding text that is not factual. Therefore, the LLM output will likely have to be scrutinized for enablement. In some cases, the LLM output may provide an incentive to add features.

5.4. The Risk of Hindsight

A central purpose of the problem-solution approach is to avoid hindsight bias. To this end, the objective technical problem must be formulated in such a way that it does not contain pointers to the technical solution (T 229/85). Thus, it may be argued that the skilled person using a LLM as a tool trying to solve the objective technical problem is admissible and not an ex post facto view of inventive activity.

However, it may be argued that generating new prior art via an LLM based on the objective technical problem (which is itself derived from a legal analysis of the prior art and the invention) may create a risk of hindsight. The LLM that existed at the relevant date is confronted with a question — the objective technical problem — that could not have existed at that date, as it is formulated only after the invention is known. This introduces a temporal paradox: the model is being asked to solve a problem that was not posed at the time, potentially leading to an artificial inflation of the model’s capabilities and an unfair assessment of obviousness.

6. Outlook

6.1. Implications for the Problem-Solution Approach

If the problem-solution approach doubles down on the requirement that the objective technical problem is truly derived without hindsight, there may be a strong argument not to use an LLM available at the time by the skilled person trying to solve it. This would preserve the conceptual integrity of the approach and avoid the risk of hindsight. However, as LLMs become better and consequently more widely adopted, the bar for inventive step may rise significantly if their use is accepted as routine. Alternatively, LLMs could be excluded from the assessment of inventive step, as outlined above, but this would make the problem-solution approach more artificial and the storytelling aspect of the skilled person less convincing. It is worth recalling that the skilled person has always been a somewhat implausible character, required to know all the literature at the relevant date while being devoid of any inventive activity.

Moreover, as LLMs become more prevalent, it is foreseeable that, in opposition or revocation proceedings, parties may seek to introduce output from historic LLMs (corresponding to the relevant priority date) to argue for or against inventive step. This could lead to a new evidentiary battleground, with both patent proprietors and opponents seeking to demonstrate what a model could or could not have produced at the relevant time.

6.2. Implications for Alternative Approaches

The potential integration of generative AI into the skilled person’s toolkit raises questions for alternative inventive step tests.

Notably, the Unified Patent Court (UPC) has so far not decided if the problem solution approach must be applied or if alternative approaches should be used. In UPC_CFI_501/2023 (issued on 4 April 2025) the Local Division Munich held that both “the ‘German’ test and the PSA […] should lead to the same results in the majority of the cases”. In the decision, the court favored the PSA. However, regardless of legal standard used to assess inventive step: If AI models become the primary means of technical problem-solving, hypothetical LLM output may be required to be taken into account in one way or the other. An alternative approach to assessing inventive step based on embeddings is, for example, discussed in R. Free, epi Information 02/2024, p.13.

In case the abilities of AI continue to grow, inventors will increasingly be able to use LLMs to augment their inventiveness, for example by adding further embodiments not found in the prior art. Using LLMs may counterbalance these new possibilities when assessing inventive step.

Thus, if LLMs are used to assess inventive step, the threshold for non-obviousness may rise, and the focus may shift from human creativity to the capabilities of AI systems (see previous discussion e.g., by R. Abbott, 66 UCLA L. Rev. 2 (2019) or D. Fabris, IIC 2020, 685). As AI systems may also augment the inventive process, a new patentability equilibrium may emerge as a sublation of the former purely (fictional) human centered inventive step assessment.

7. Conclusion

As of fall 2025, the prevailing view is that the skilled person at the EPO is not yet assumed to routinely use generative AI models when solving the objective technical problem. However, this may change as such tools become more deeply embedded in technical practice. Notably, the PSA could “naturally” integrate the use of “historic” AI systems at the stage of the objective technical problem. The presentation and assessment of such arguments provide new challenges both for practitioners, offices, and courts. Very likely, the capability for inference of historic large language models will be required - for both submissions and decisions.

References

EPO Guidelines 2025 ed., April 2025, Part G, Chapter VII.
Wikipedia, “Llama (language model)”, accessed September 2025
(https://en.wikipedia.org/w/index.php?title=Llama_(language_model)&oldid=1311943860)
eugeneyan/open-llms: A list of open LLMs available for commercial use.
T 1193/23, T 206/22 (EPO Boards of Appeal).
Fabris, D., “From the PHOSITA to the MOSITA: Will ‘Secondary Considerations’ Save Pharmaceutical Patents from Artificial Intelligence?”, IIC 2020, 685.
R. Free, “A new test for inventive step”, epi Information 02/2024, p.13.
Abbott, R., “Everything Is Obvious”, 66 UCLA L. Rev. 2 (2019).
Richard, P., “ChatGPT Person Skilled in the Art at EPO”, LinkedIn, 14 May 2025
(https://www.linkedin.com/posts/prestonrichard_epo-patents-epo-activity-7328373865401724928-m8B7).