Here are some ideas for challenging, in proceedings before the EPO, a document generated by AI 

These issues arise in a context where patent professionals—whether acting during examination, opposition, or appeal—are already confronted with imprecise or speculative disclosures, sometimes originating from the patent literature itself. The massive introduction of AI-generated content exacerbates these difficulties by multiplying disclosures that appear technically plausible but lack any real experimental basis. To date, no clear and formal position appears to have been adopted by professional authorities or patent offices regarding publications generated predominantly or entirely by AI as prior art.

It is in this context that the present discussion seeks to analyse, in light of the Guidelines of the European Patent Office (EPO) and the case law of the Boards of Appeal, how erroneous or speculative technical teachings—such as those found in AI-generated publications—should be assessed as part of the state of the art.

The concept of a prior art document

Pursuant to Article 54(2) and (3) of the European Patent Convention (EPC), the state of the art comprises everything made available to the public before the filing date of the European patent application by means of a written or oral description, by use, or in any other way. In the present article, the analysis focuses on written disclosures, which are the relevant mode of disclosure for AI-generated publications.

With regard to written disclosures, their accessibility to the public is a determining factor in assessing whether they belong to the state of the art. As a reminder, information must be considered publicly available if, on the relevant date, it was possible for members of the public to become aware of it, and if that information was not subject to any measures designed to preserve its confidentiality or to restrict the use or dissemination of the knowledge thus acquired^[1].

In light of these principles, it appears that publications generated entirely by AI, once they are made publicly available without confidentiality restrictions, must therefore be considered part of the state of the art.

Once accessibility is accepted, the question naturally shifts to “when.” However, when it comes to the Internet—and even more so for AI-generated content, which can be volatile—establishing the date of accessibility is often a point of contention.

In general, AI-generated publications disseminated online are subject to the same rules as any Internet disclosure within the meaning of Article 54(2) EPC. They are considered to be accessible to the public on the date on which they were made available online to the public^[2], regardless of whether their content was produced by AI, provided that this date can be established with a sufficient degree of certainty.

Establishing this publication date involves two steps^[3]: verifying the reliability of the indicated date and confirming that the content was actually accessible to the public on that date. In many cases, online publication platforms (such as open archives, institutional repositories, or public databases) expressly indicate a date of upload or electronic publication. This date is, in principle, presumed to be correct and reliable.

It is true that the evolving nature of the Internet may complicate this assessment, in particular where pages do not include an explicit date or when their content is subject to change. However, various elements (online archives, timestamps, version histories, indexing by search engines, dissemination across multiple websites) can help to confirm it^[4]. 

In the absence of serious evidence of manipulation, the date displayed is therefore deemed to be correct, and it is up to the applicant or patent holder to cast sufficient doubt to rebut this presumption, in accordance with the principle of “balance of probabilities” in this field.

Decision T3000/19 illustrates the requirement for rigorous evidence. It also shows that a temporary publication on the internet may be insufficient. In this case, the Board overturned a rejection decision based on a YouTube video that had become inaccessible, preventing it from verifying the merits of the contested decision.  The Board reiterated that when disclosures from the internet are used as prior art, the examining divisions must ensure the reliability of the evidence, particularly with regard to the date of publication and the accessibility of the content in its original version (metadata and original format). Given the evolving nature of online content and the right of third parties to consult the file, electronic evidence must be properly collected, stored, and made accessible.

Thus, any document made available to the public before the effective date is likely to constitute prior art within the meaning of the EPC, without it being necessary for it to have a human author or to have been proofread or validated. However, the analysis cannot stop at this formal observation alone. AI-generated content is likely to contain factual errors (hallucinations), technical approximations, or speculative reasoning. 

However, European patent law does not require a prior art document to be error-free or scientifically accurate. An “imperfect” document may therefore be considered prior art. However, there is an essential limit to this tolerance: the content must remain intelligible to a person skilled in the art. 

Assessment of patentability in the face of erroneous prior art

In order to determine whether the content is intelligible, it is necessary to identify what the skilled person would objectively be able to understand, at the relevant date, in light of their common general knowledge and technical reality.
According to the EPO’s approach^[5], the relevance of a document containing a potential error depends on the ability of a skilled person to identify that error and, if necessary, correct it. The analysis is thus guided by a pragmatic assessment of the teaching actually disclosed.

Three situations are distinguished:

1. If the error is immediately recognisable and only one obvious correction is possible, the document is interpreted as implicitly containing that correction, and it is the corrected version that forms part of the state of the art.
2. If the error is recognisable but several plausible corrections exist, the passage cannot be relied upon as a relevant disclosure, for lack of a clear and unambiguous technical teaching.
3. If the error is not identifiable as such, the document is taken into account as such.

This analytical framework is firmly rooted in the case law of the Boards of Appeal, which consistently emphasises that the skilled person is primarily concerned with technical reality, rather than with a purely formal or abstract reading of prior art documents.

Decision T 77/87 is a fundamental reference in this respect. The Board held that, in determining the state of the art within the meaning of Article 54 EPC, what must be considered is what has actually been made available to the skilled person in terms of technical reality. In principle, the literal disclosure of a prior document is taken into account as it stands. However, where reliable elements show that this disclosure is manifestly erroneous and does not correspond to the intended technical reality, that erroneous disclosure must not be regarded as forming part of the state of the art.

In that case, an abstract published in Chemical Abstracts contained a manifest error compared with the original document. The Board held that the abstract could not be analysed in isolation and that, in the face of a substantial inconsistency, the original document had to prevail, the erroneous abstract being disregarded.

This approach was confirmed and refined in numerous later decisions, including T 1088/96, where a manifest error in a prior document was corrected by the skilled person in light of their common general knowledge and corresponding documents published simultaneously. The Board expressly recalled that the assessment of the content of the state of the art must proceed on the assumption that the skilled person seeks technical coherence and plausibility (explicitly referring to T 77/87).

Similarly, in T 89/87, T 412/91 and T 1088/96, typographical errors or internal mistakes were considered immediately detectable and correctable, leading to the disclosure being taken into account in its corrected form.

Conversely, decision T 161/98 confirms that only the factually incorrect information (for example, an incorrect boiling point value) had to be excluded from the state of the art, without disqualifying the entire technical teaching of the document.

For AI-generated publications, the consequence is straightforward: the fact that an AI text contains errors is not enough to disqualify it, but often provides an angle of attack to reduce (or even neutralize) what it is supposed to disclose, by applying the above analysis grid and drawing on the general knowledge of the skilled person.

However, while in fields such as mechanics, simply reading a document often allows the skilled person to detect an inconsistency or technical impossibility, the situation is slightly different in fields such as chemistry, biology, or biotechnology. In these fileds, the existence of an error is not necessarily obvious, as the validity of a teaching can only be assessed by means of tests aimed at reproducing the teaching. The question is then no longer limited to identifying an error, but also concerns the sufficiency of the disclosure.

The requirement of reproducibility in patent law

In many cases, the most effective challenge is not to demonstrate that the document is “false,” but that it cannot be reproduced because it is insufficiently described. AI-generated literature is particularly vulnerable to this criticism: it can describe plausible results without providing the parameters, conditions, or operational paths needed to obtain them.

In this respect, it is similar to documents whose disclosure is described as “speculative” or “prophetic” examples.

A speculative disclosure presents hypothetical ideas or effects without sufficient demonstration, experimental evidence, or technical indications to enable reproduction. In principle, it does not constitute sufficient teaching if a skilled person cannot implement the disclosed subject matter. In this regard, courts, particularly in France^[6], have already classified patents whose technical effect was not demonstrated as “speculative.” However, speculative nature does not automatically exclude the reproducibility of the disclosure: if the information provided, combined with the general knowledge of the skilled person, allows implementation without excessive effort, it may be taken into account in the prior art.

Prophetic examples, also known as fictitious, hypothetical, or “paper examples,” consist of hypothetical examples inserted into patent applications. Prophetic examples are presumed to be valid, and it is therefore incumbent upon those who wish to challenge the technical teaching of the prior art to demonstrate their inadmissibility. While in the United States the admissibility of prophetic examples is based on reasonable scientific grounds, their non-misleading nature, and their distinction from real examples^[7], the EPO accepts them without requiring them to be explicitly differentiated from experimental examples, which can make them difficult to identify. Provided they are reproducible.

In fact, whether speculative, prophetic, or merely conceptual, a disclosure is relevant before the EPO only if its teaching is reproducible. This requirement—that a disclosure must enable the implementation of an invention—is consistent with the principle set out in Article 83 EPC. It would therefore be possible to challenge the relevance of a prior art document whose disclosure is insufficient. The reproducibility of the disclosure has been assessed on several occasions by the EPO Boards of Appeal.

Challenging a document cited for lack of novelty

It is well known that only prior art whose technical teaching can be reproduced is destructive of novelty^[8]. In decisions T 206/83 and T 719/12, it is considered that a chemical compound is not disclosed if the skilled person is unable to prepare it, even if its structure is disclosed. 

In decision T 113/17, the Board reiterated that a prior art document can only destroy novelty if it describes the invention sufficiently, i.e., in a credible or plausible manner, and without merely referring to theoretical concepts. However, in the case of the document cited, the Board considered that the information provided, namely vaccine compositions, was neither sufficiently detailed nor sufficiently credible to establish that the compositions described could actually be used as vaccines, as no real efficacy had been demonstrated. The examples concerned were described as hypothetical, and the document was therefore not considered capable of destroying the novelty of the claim. 

However, it should be recalled that an insufficient disclosure does not necessarily mean that it is technically incorrect.

Reproduction experiments play a key role in assessing sufficiency and make it possible to verify whether a disclosure truly meets this requirement.

In case T 465/92, a prior document had been relied upon to challenge the novelty of an invention relating to an alloy. However, reproducing the process described in that document did not make it possible to obtain the claimed product. The Board therefore held that the document did not constitute a sufficient disclosure and did not destroy novelty, since it did not enable the skilled person to reproduce the claimed subject-matter.

This case perfectly illustrates how a lack of practical details in a disclosure may render its reproduction impossible for the skilled person. Although the document mentioned general steps for processing alloys, it omitted the specific conditions (precise temperatures, exact durations, appropriate cooling rate) necessary to obtain the desired product. As a result, even the skilled person could not reproduce the invention without undue burden or additional experimentation.

In conclusion, only a technically reproducible disclosure can truly destroy novelty. Simply mentioning a concept or structure is not enough if the skilled person cannot actually obtain the claimed object based on the information provided. Reproducibility thus plays an essential role in destroying novelty.

Challenging a document cited for lack of inventive step

In decision T 1764/09, the Board of Appeal had to determine whether document D1 could constitute the closest prior art for the assessment of inventive step. It concluded that D1 was speculative in nature and was therefore unsuitable as a realistic starting point for the analysis of inventive step.

This assessment was based on the hypothetical nature of the document, as evidenced by:

• its title, which itself indicated its speculative nature (“On the possibility of…”), suggesting that the article explored hypothetical concepts rather than established technical achievements;
• the absence of any concrete technical achievement, D1 presenting a theoretical examination of future possibilities in intraocular adaptive optics, without providing any detailed description or applicable technical solutions; 
• the gap between the theoretical considerations set out and any actual practical application. 

As a result, D1 was not retained as a relevant basis for the assessment of inventive step.

Conversely, in decision T 184/10, the Board decided that a document could not be disregarded as the closest prior art solely on the grounds that it contained forward-looking elements. Although document D14 contained speculative statements about the future use of certain substances in the treatment of diabetes, it also mentioned that these compounds were already undergoing clinical trials. The Board considered that these elements, combined with the general knowledge of the time, provided a credible basis for the predictions made. Consequently, the Board concluded that D14 could not be dismissed as pure speculation, but had to be considered a serious attempt to interpret the existing state of the art. This led to D14 being classified as the closest prior art, justifying the subsequent revocation of the patent.

However, decision T 161/98, already cited above, also clarified that the speculative nature of certain information, in the sense of extrapolations or generalizations based on limited experimental data, is not in itself a sufficient reason to disregard a disclosure of prior art. Much of the patent literature is based on such generalizations, which those skilled in the art accept as long as they remain technically plausible. 

However, the question may shift to another, distinct but related area: that of reasonable expectation of success. When a prior art document merely suggests a course of development or offers a forward-looking incentive, the skilled person must still have had objective reasons, at the relevant date, to anticipate the success of such an approach. Otherwise, the incentive may appear to be mere speculation, insufficient to form the basis for obvious reasoning.

According to the EPO Guidelines, inventive step must be assessed on the basis of a reasonable expectation of success, and not on a mere “try and see” approach^[9]. Thus, when a document—particularly one generated by AI—refers to a theoretical possibility without a credible technical or experimental basis, it can be argued that such an expectation was lacking, thereby weakening the allegation of obviousness within the meaning of Article 56 EPC.

In terms of inventive step, the reproducibility of a teaching may influence the assessment of the realistic nature of the starting point chosen. Furthermore, even when a document is taken into account for the assessment of inventive step, obviousness can only be accepted if the skilled person would have followed its teaching with a reasonable expectation of success. The mere presence of a suggestion or a prospective opening is therefore not sufficient: it must also be based on a credible technical basis that allows the success of the envisaged approach to be objectively anticipated.

Notes

1. Guidelines G-IV, 1 and G-IV, 7.2.1
2. See the EPO press release on Internet citations (OJ EPO 2009, 456) and Guidelines B-VI, 7 and G-IV, 7.5.
3. Guidelines G-IV, 7.5.1
4. Guidelines G-IV, 7.5.4
5. Guidelines G-IV, 9
6. TGI Paris, 6 octobre 2009, Teva / Sepracor, RG n°07/16446 ; TGI Paris, 28 février 2013, Sanofi / Mylan, RG n°12/16120, TGI Paris, 13 février 2014, Virbac / Merial, RG n°13/00455 ; Cass. Com., 6 décembre 2017, N° de pourvoi: 15-19726
7. USPTO, MPEP, “Properly presenting prophetic and working examples in a patent application”, 7 January 2021.
8. Guidelines G-IV-2
9. Guidelines G, VII, 13

February 2026