When AI Floods the Prior Art, Enablement Becomes the EPO Battleground
Debate is intensifying across European patent practice over a practical question that used to sound theoretical: what happens when generative AI starts mass-producing technical disclosures that are later cited against fresh filings? What can be verified publicly is not a newly published EPO rule aimed specifically at “synthetic prior art”, but the continued force of a familiar one in the 2026 EPO Guidelines: a prior-art document must give the skilled person enough information to carry out the relevant technical teaching. A polished text is not enough on its own.
That point matters much more once AI systems can generate vast numbers of seemingly plausible combinations, formulations and molecular proposals at industrial scale. In chemistry, materials and life-science cases, the pressure created by a large machine-generated disclosure can be real. But if the cited document lacks a reproducible route, key conditions, credible data or any serious experimental footing, applicants still have room to shift the argument away from surface overlap and back to a harder question: does this document actually enable what it seems to disclose?
The real issue is not authorship by AI, but whether the technical teaching can be worked
European patent law does not need a special doctrinal category before applicants can push back against weak AI-generated disclosures. The existing framework already asks the right question. If a document is relied on for novelty or inventive-step purposes, the relevant teaching must be made available to the public in a form that the skilled person can actually practise with common general knowledge. In chemical and formulation cases, a name, a formula or a broad conceptual statement does not automatically do that work.
This is why AI-generated material does not fail simply because it was machine-produced, but it also does not become powerful prior art simply because it looks complete. Once examiners or opponents lean on a document that appears to have been mass-generated, the pressure point is practical implementation: where are the critical parameters, what conditions are disclosed, what result is credibly supported, and which claimed effect is backed by something more than prediction? If those answers are missing, the disclosure may look broad while remaining fragile.
Markush-heavy and combinatorial cases are where the pressure will be strongest
The problem is especially sharp in chemistry, pharma, advanced materials and some biotech cases. Generative tools are very good at producing long lists of compounds, substitutions, parameter ranges and use hypotheses that look technically coherent at first glance. That can create immediate prosecution pressure because the applicant appears to be boxed in by a forest of earlier text. Yet once the analysis gets more granular, many of those combinations lack a workable synthesis route, purification conditions, stability data, activity data or even a credible basis for expecting the proposed result.
That gives applicants a concrete line of response. For Markush-style disclosures, breadth is not the same thing as a specific enabling disclosure. The better question is whether the skilled person, at the relevant date, could really make the cited compound, composition or embodiment and reasonably expect the promised effect using the document plus common general knowledge. If not, the document may still exist as background noise, but its value as a serious novelty-destroying or inventive-step starting point becomes much harder to defend.
This changes prosecution strategy more than many teams expect
When confronted with a massive reference that appears to cover everything, many teams instinctively narrow first and argue distinctions later. That reflex may become expensive in the AI era. If the cited material carries clear signs of automated generation, applicants should first test its enabling core before conceding too much claim scope. In European practice, once the case is reframed around enablement, a reference that initially looked overwhelming can lose much of its force.
The order of attack matters. Instead of rushing to explain why the claimed invention differs from the reference, it can be more effective to dismantle the reference itself: which process steps are missing, which effects are unsupported, which embodiments exist only as auto-generated permutations, and which conclusions would require undue experimentation to verify. A shift in sequence often changes the centre of gravity of the whole case. Some AI-built references look formidable only from a distance.
What applicants should start preparing now
First, development records need to be cleaner and more granular. The most persuasive rebuttal is rarely a general complaint that an AI-generated document is unreliable; it is a specific explanation of why the skilled person could not make, test or validate the cited teaching from the document. Second, experiments and technical parameters should play a larger role in office-action responses and oppositions. The party that can explain necessary conditions, failure points and technical bottlenecks most concretely is more likely to make the enablement challenge stick.
Third, monitoring practice needs an upgrade. If competitors, platforms or third parties begin publishing large volumes of automatically generated technical text online, companies should track publication dates, document stability, version history and the actual technical gaps inside those disclosures. The central shift here is not that the EPO has given AI prior art a new label. It is that examination and post-grant disputes are becoming more willing to ask a blunt question: did this reference really place the technology in the hands of the public, or did it merely pile possibility upon possibility into a polished document?
