In the new AI-age, where Large Language Models (LLMs) like ChatGPT can generate sophisticated textual answers in seconds, the relevance of AI to the patent industry is an active area of debate. The recent EPO Board of Appeal in T 1193/23 specifically questioned the reliability of LLM-generated interpretations for technical terms within patent claims. Central to this case was the question of whether the output of LLMs, trained on everything on the internet, can serve as evidence of how a person skilled in the art would interpret a technical term in a patent claim
Legal background: The person skilled in the art
The concept of the person skilled in the art is fundamental to European patent law, particularly in the context of how claims are interpreted for sufficiency, novelty and inventive step analysis. According to the established case law of the Boards of Appeal, the "person skilled in the art" or "skilled person" is presumed to be a skilled practitioner in the relevant field of technology who has average knowledge and ability (average skilled person). This notional person is the addressee of a patent and serves as the standard against which novelty and inventive step are judged.
 |
| Navigating different technical fields |
The skilled person is also considered to be aware of what was common general knowledge (CGK) in the art at the relevant date. They are also presumed to have had access to everything in the "state of the art," and to have had the means and capacity for routine work and experimentation which are normal for the field of technology in question. Traditionally, determining how the skilled person would understand technical terms in a patent has been a matter of expert evidence and established technical knowledge.
Case background
The appeal case T 1193/23 concerned EP 3118356 owned by Rieter. The Opponent in the case was Saurer Spinning Solutions GmbH. The patent particularly related to improved safety mechanisms for operating rotor spinning machines. These machines, used in textile production, feature numerous rotors operating at extremely high speeds. The invention claimed a method for safely starting and stopping these rotors by checking bearing regulation or data connections before allowing operation. The main legal issue was whether the patent lacked novelty because all the features of the claim were allegedly disclosed in a prior art document. Critical to the decision was how the technical terms, such as "bearing regulation" should be interpreted in the context of the claimed invention.
ChatGPT enters the proceedings
In a first for the EPO Boards of Appeal, the decision addressed the use of AI-generated content as evidence for claim interpretation. During oral proceedings, the Patentee verbally submitted responses from ChatGPT regarding the meaning of "bearing regulation" and other technical terms, arguing for a narrower definition of the terms than was proposed by the opponent. However, the Board of Appeal found that responses from ChatGPT alone could not establish an interpretation of the claims, but should be substantiated by expert evidence. The Board of Appeal went as far as to find the responses from ChatGPT to be "irrelevant, since the interpretation of the claim depends on the skilled person's understanding". As summarised in the headnote, the Board of Appeal concluded:
"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 a received answer - which is based on training data unknown to the user and can also depend sensitively on the context and the exact formulation of the question(s) - necessarily correctly reflects the expert's understanding of the respective technical field (at the relevant time)." (r.1.1.1, emphasis added, machine translation)
With respect to the case in question, the Board of Appeal further noted that no proof other than the responses from ChatGPT were submitted to support the proposed definitions. Therefore, whilst the Board of Appeal found no absolute bar to the submission of evidence from LLMs, they nonetheless required that any such submissions be substantiated with additional proof. As the Board of Appeal went on to explain: "Evidence of how certain terms in the claim of a patent (or patent application) are interpreted by the person skilled in the art can be provided, for example, by appropriate technical literature. No such evidence has been submitted for the alleged different meanings of the above-mentioned terms in the relevant technical field" (r. 1.1.1).
Is there a place for LLMs in claim interpretation?
The Board of Appeal based their decision on what it saw as three key limitations of AI systems like ChatGPT in the patent context. First, LLMs rely on training data unknown to the user. Second, the responses of LLMs depend sensitively on context and question formulation. Third, LLMs cannot be assumed to reflect actual expert understanding, especially at relevant points in time.
For this Kat, the first point, regarding the training data for LLMs, is the weakest reason for rejecting their use for claim interpretation. Foundational LLM models such as ChatGPT, Claude and Gemini are trained on vast amounts of data extracted from the internet, books, and public documents. It is therefore fairly safe to assume that the LLM has access to the common general knowledge of the skilled person in most if not all technical fields.
The Board of Appeal's second and third arguments, i.e. that the response of an LLM will depend on the question and context, are more relevant. However, one could imagine a situation where these objections may be overcome. A LLM could, for example, be intelligently prompted to answer a question as if it was a skilled person in the field. In such a prompt, the LLM could be provided as context to the question, the EPO Guidelines and case law on what constitutes a skilled person and the identity of the technical field. For this Kat, the main issue with using LLM definitions for patent challenges, is knowing the cut-off date for the information on which the model is trained and has access to. Sometimes this information is given about a LLM, sometimes it is not. However, it is currently not possible to ask LLM providers to base answers on knowledge up until a certain date, given that this is not something most people have much use for.
Turning back to the decision, a key problem for the Patentee in this case was the limited context they appeared to have provided in support of their submissions from ChatGPT. As the Board of Appeal observed, "the respondent did not submit the extensive, orally read, partly bullet-pointed responses [from ChatGPT] to the file in writing". Details of the prompt and information given to ChatGPT appear to have been lacking. Having rejected the Patentee's interpretation of the relevant terms, the Board of Appeal ultimately found that the main claim lacked novelty over the prior art. The Board of Appeal based their interpretation on technical analysis of the claim language in the context of the claim as a whole. The Board also refused all five auxiliary requests, finding that the additional claimed features in these requests lacked inventive step.
Use of the description to interpret the claims (G1/24)
Claim interpretation is a current hot topic at the EPO, whilst we await the decision of the Enlarged Board of Appeal on whether the description can and should be used to interpret the claims (IPKat). This Kat thought she would see how an LLM might interpret the term of contention in G 1/24.
The case in G 1/24 centres on how the term "gathered" tobacco should be interpreted in a patent claim. The referring Board of Appeal decision in G 1/24 found that the term gathered has a specific meaning in the art. However, the description of the patent defines the term differently, specifically to include tobacco that is spiralled. Unfortunately for the patentee in G 1/24, this meant a product comprising spiralled tobacco was considered novelty destroying prior art for their claimed product.
So how would an LLM have construed gathered tobacco? PatKat asked Anthropic's Claude, using the following prompt:
See attached the guidelines and case law for what constitutes a person skilled in the art in European patent law. Answer the following question: You are a person skilled in the art in the field of the tobacco industry reading a patent related to a new type of electronic cigarette. Claim 1 is as follows: [Claim 1 of EP3076804B1]. How would you interpret the term "gathered" in this claim and does it include spiralled?
Claude supplied the following interpretation:
As a person skilled in the art in the tobacco industry, I would interpret the term "gathered" in this patent claim to refer specifically to a manufacturing technique where a flat sheet of aerosol-forming material is pleated, folded, or crimped to create multiple folds bunched together, resulting in increased surface area within a confined cylindrical space. This is distinct from a "spiralled" configuration, which would involve winding or rolling the sheet in a helical manner around a central axis. These represent different, recognized manufacturing techniques in the tobacco industry with distinct structural outcomes, and the deliberate choice of "gathered" in the claim language would indicate to me that the patent is specifically claiming the pleated/folded configuration rather than a spiralled arrangement.
In a previous post, this Kat explored how the rise of AI tools for patent drafting may increase the risk of the kind of problematic boilerplate language at issue in G 1/24 (IPKat). However, from above, it appears that LLMs can do a reasonably decent job at technical definitions when given the appropriate context. In fact, the use of LLMs to check the specification for overly broad and non-technical definitions might very well have saved the patent in G 1/24, as it may very well have identified the definition provided for "spiralled" as being at odds with the usual definition in the field. However, it is of course possible that the LLM has been influenced by all the commentary on the case...However, it is nonetheless clear that, in patent drafting and claim interpretation, the identity technical field is key. AI tools can therefore only be effective when they are informed of the appropriate technical context for their answers.
Turning back to T 1193/23, it is notable that the Board of Appeal does not say that there is no place for evidence from LLMs in patent disputes, merely that the conclusions of an LLM need to be substantiated by additional supportive evidence. There are now plenty of LLM tools that will provide citations to substantiate their conclusions (which can be checked in order to establish their validity). As with any tool, to be most effective LLMs need to be used by the right person, in the right way for the right job.
Final thoughts
T 1193/23 marks an important first step for the EPO in addressing the role of AI in patent proceedings, establishing that LLM outputs alone cannot completely replace expert evidence. However, in this Kat's view LLMs will undoubtedly have a place in the patent industry. The business of patents involves debating the smallest nuances of meaning hidden in technical terminology, and LLMs represent an extraordinary resource for this task. Crucially, the Board of Appeal decision in T 1193/23 did not rule out the use of LLMs for claim interpretation, but merely indicates that any evidence based on the output of an LLM be supported and substantiated by additional evidence.
To get the most out of LLMs in the context of patents, it is necessary to explicitly frame any queries within the specific technical field in question. Without this crucial contextual guidance, an LLM's interpretation of highly specialised terms risks being too broad or misaligned. The nuanced and highly technical language of inventions demands an equally technical-field focused approach to the use of AI tools.
Further reading
- Use of large language models in the patent industry: A risk to patent quality? (Oct 2023)
- Review of AI patent drafting software for life sciences: Qatent (Feb 2025)
- [Guest Post] In too deep? Considering the legal risks when using DeepSeek (Feb 2025)
- Beware of boilerplate: Practical lessons for patent drafting from G1/24 (Claim interpretation) (Apr 2025)
Reviewed by Dr Rose Hughes
on
Monday, May 19, 2025
Rating:
Follow the IPKat on LinkedIn 
![[UPCKat] Has the UPC set too high burden to stop generic entry in Boehringer Ingelheim v Zentiva?](https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEiq8nBoNt-e3wprqLLKpCrLrwgq-jlUJjQkoVYKzUR7ppQ2ssWrPpA6uQunU49Y3s216DkeyQHN2bw0C2cDUlnwYWlQZIn7fi9vynFF5rS0JxrDx4qHKB82RM5P45wh5JHkI-f0I9sGFuQ4OVkPIIfvjec1UtVj-UyEvcBM-ApvTjK39-QxbXEz6w/s72-c/A%20whimsical%20cat-themed%20jigsaw%20puzzle%20partially%20assembled%20on%20a%20wooden%20table,%20with%20the%20cityscape%20of%20Li.jpeg)
Posts
Well said. Surely another issue with using an output from a large language model like ChatGPT is that, when given the same prompt multiple times, the model will output slightly different responses. My understanding is that this is baked into how LLMs work.
ReplyDeleteAnother issue, and one that touches on the remark that the response of an LLM will depend on the question and context, is that the output of current LLMs are heavily influenced by the phrasing used. For instance, if asked whether a particular statement is correct, then the LLM will often provide a quite reasonable analysis. If instead prompted with the phrase 'why is this statement correct?' then the LLM will instead hallucinate a reasonable-sounding argument for why. This is apparently a product of the retraining done on the initial models, which make them 'agreeable' to the users (meaning more or less sycophantic) and so might be fixable for specific implementations of the same underlying model.
The statement in your blog reminds me of a glass. For some it is half empty and for some it is half full. You appear to belong to the second category. The board did not exclude the use of LLM or AI in patent matters, but it put rather stringent conditions.
ReplyDeleteFrom the day of its publication, provided it is properly established and not merely a transient event, the output of a LLM or an AI belongs to the prior art. This fact is irrespective of what was the question leading to the output.
Problems start when it comes to determine what was known, be it as common general knowledge or specific one, or to interpret the meaning of claimed features, ‘before the effective date of a claim. In patents the effective date of a claim is the key element.
Trying to determine common general knowledge or a specific one, or to interpret the meaning of claimed features with the output of a LLM or AI which became public after the effective date of a claim, is actually doomed to fail, if it is not insured that the data input to the LLM or AI has become public before the effective date of the claim.
The situation is similar to that of Internet citations as prior art. There is presently a large body of case law making clear what the conditions are in order to take them into account. The most current is to use a way back machine. Unless there is a similar tool available for LLM or AI, it will not be possible to use a LLM or an AI in order to determine common general knowledge or to interpret a claim. The question given to the LLM or AI will also have a big influence of the outcome.
Unless all those conditions are satisfactorily resolved, the output of a LLM trying to prove any technical knowledge or general meaning before the effective date of a claim will not be possible.
This decision is to be approved. I generally agree with "interested third party".
ReplyDeleteFirst, there is no cut-off date for ChatGPT, therefore no evidence that the output was available on the filing date. This is a requirement for an assertion regarding the CGK or the meaning of a claim term, it has to be documented on the basis of documents available on the filing date.
Second, the output of ChatGPT is not « sourced » i.e. it does not cite the data which have led to the output, so that it is impossible to assess its reliability, and it is not explainable since ChatGPT is a black box. In this case, the Board points out that the applicant did not provide any literature to support their position.
As to the interaction with G 1/24, I do not quite agree with Rose’s depiction of the description of EP3076804 at stake in the referring decision T 439/22 as « problematic boilerplate ». The definition of ‘gathered’ in para [0035] of the description is worded as an authoritative statement which leaves no doubt as to the patent owner’s intent regarding the meaning of « gathered ». It starts with « as used herein, the term ‘gathered ‘ denotes…». The term « as used herein » is typical language in the preamble of agreements, and it implies that the specific meaning provided by the definition prevails over any other meaning for the claimed term, regardless of whether it is « ordinary » or not. This can by no means been equated to « boilerplate » conditions in an agreement.
This is consistent with the case law of the BOAs ( II.A.6.3.3) citing a long list of decisions :
« According to the established case law, the patent document may be its own dictionary (see e.g. T 311/93, T 523/00, T 1192/02, T 61/03, T 1321/04, T 620/08, T 2480/11, T 1817/14; see also in this chapter II.A.6.3.5)."
In its preliminary opinion, the EBA stated that question 3 of the referring decision T 439/22 relating spefically to whether the patent could be its own dictionary was inadmissible, without providing explanation. A logical reason may be found in the EBA’s positive answer to question 2 in the preliminary opinion. If it is permissible to use the description to interpret the claims, then question 3 becomes pointless.
This would also be in tune with the case law of the UPC. The statement that « the patent may be its own lexicon » can be found in a number of UPC decisions. Consistency of the UPC’s case law with the EPO’s is obviously important to owners of UPs for legal security.
The chapters of the CLBA quoted by Mr Hagel, have for title:
ReplyDelete- II.A.6.3.3 “Interpretation of ambiguous terms or confirmation of the text of a claim
- II.A.6.3.5. “Use in the examination relating to the clarity requirement pursuant to Article 84 EPC”.
T 1271/05, Headnote, says clearly the application cannot be its own dictionary. In T 2221/10, the board referred to established case law according to which the description can be used as the patent's "dictionary" to assess the correct meaning of ambiguous terms, but added that, if a term used in a claim has a clear technical meaning, the description cannot be used to interpret such a term in a different way.
In T 197/10 the board explained that if the claims are worded so clearly and
unambiguously as to be understood without difficulty by the person skilled in the art, there
is no need to use the description to interpret the claims.
T 2221/10 and T 197/10 are mentioned in CLBA II.A.6.3.1 "General principles".
The chapters of the CLBA quoted by Mr Hage, and the further decisions quoted above here as examples, can thus not be considered giving a general line of case law according to which the patent might be it own dictionary. It is only in case of ambiguities or when it comes to Art 84. When the claim is clear as such, no recourse to the description should be accepted. This is the core of G 1/24.
The statement that « the patent may be its own lexicon » can indeed be found in a number of UPC decisions. However, those decisions correspond to German practice which is clearly inspiring UPC decisions.
I would also refrain from overinterpreting the position of the EBA when it considers Question 3 in T 439/22 as not admissible. We will have to wait for G 1/24 to be issued in order to understand why the EBA considers question 3 not admissible.
Even if the answer to question 2 is positive in its broadest interpretation, i.e. “general” or systematic recourse to the description, it does not mean whatsoever that the EBA will accept that the patent is its own dictionary.
If the EBA takes care of not fostering angora cats, I cannot imagine that the EBA will accept that the patent is its own dictionary. A limiting feature in a claim should not be allowed to be blown up in the description, for the sake of a dictionary. Similarly, a limiting feature in the description should not be read into a claim.
Let's hope that G 1/24 comes soon.