ChatGPT is not a proxy for the skilled person

Background

Rieter CZ s.r.o. (“Rieter”) was granted a patent (EP 3 118 356) on 4 November 2020 relating to a procedure for safely starting and/or stopping a rotor of a rotor spinning machine for the production of yarn. The patent was upheld (in amended form) by the EPO Opposition Division on 18 April 2023 following EPO opposition proceedings brought by Saurer Spinning Solutions GmbH & Co. KG (“Saurer”), but Sauer appealed that decision to the EPO Board of Appeal (the "Board"), seeking revocation of the patent.

Rieter requested that the appeal be dismissed, or alternatively that the patent be maintained in amended form on the basis of one of five auxiliary requests.  Rieter argued that certain features in one of the patent’s claims were not disclosed in the prior art if the claim language was properly interpreted. An interesting issue arose because in support of its claim interpretation Rieter relied on responses it received from ChatGPT after feeding it questions about the meaning of the relevant terms used in the contested patent claim.

The appeal decision

On 14 May 2025, the Board handed down its decision (T 1193/23). The Board ultimately overturned the Opposition Division decision, and revoked the patent. The Board found that claim 1 lacked novelty over a piece of prior art, and that the each of Rieter’s five auxiliary requests also lacked either novelty or an inventive step.

The most interesting part of the Board’s decision comes in its discussion of the relevance to claim interpretation of answers provided by the large-language model (LLM) ‘ChatGPT’. The Board held that, because Rieter had only read out some of the responses received from ChatGPT orally at the hearing, and did not submit in writing either the full responses received, or the details or context of the questions asked, the content of the responses could not be taken into account for the purposes of the Board's decision. 

However, the Board helpfully provided some general guidance on the relevance of LLM responses to claim interpretation.  The Board found that ChatGPT’s answer was “irrelevant” as claim interpretation ultimately depends on the understanding of the skilled person. The Board went on to state:

“The general increase in the spread and use of chatbots based on language models ("large language models") and/or "artificial intelligence" alone does not justify the assumption that an answer received - which is based on training data unknown to the user and may also depend sensitively on the context and the exact formulation of the question(s) - necessarily undermines the expert's understanding of the respective technical field (at the relevant point of time)”.*

The Board acknowledged that “suitable specialist literature” could be deployed to substantiate how certain terms in the claim of a patent might be understood by the skilled person, but no such evidence had been submitted by Rieter in these proceedings.

Analysis

The notional skilled person is presumed to have had access to everything in the state of the art.  There is an obvious attraction in using an LLM, which has access to vast amounts of information, to quickly identify and assess relevant information.  However, the challenge arises in understanding what the LLM has done and limiting its assessment to the appropriate remit and context in each case. 

This decision makes clear that LLMs such as ChatGPT cannot simply be used as a proxy for the understanding of the skilled person before the EPO given the weaknesses it identified in LLMs being used for this purpose.  There were i) the lack of visibility of the training data used to generate the responses, and ii) the sensitivity of the responses to the context and formulation of the questions/prompts input into the LLM. Although more information could have been provided to address the second issue (to some extent), the first issue will always be harder to address because these ‘weaknesses’ are inherently built into the way in which LLMs currently operate.  Therefore, it will likely remain difficult to rely solely on AI-generated content as a basis for claim interpretation. A further difficulty lies in the need to ascertain the state of the art and the knowledge of the skilled person as at the priority date of the patent, where it may not be possible to ask an LLM to effectively cut-off its knowledge after a certain date.

The continued importance of expert evidence as to the understanding of the skilled person, substantiated by “suitable specialist literature” references, therefore seems secure in proceedings before the EPO and in European patent litigation more broadly. However, it is likely that we will see an increase in the use of LLMs and other AI as tools to assist legal practitioners in EPO proceedings and patent litigation, whether in prior art or specialist literature searches, or even in the drafting of submissions.

* All quotations from T 1193/23 in this article are based on machine translation.

Authored by Josh Stickland and Katie McConnell.