EPO Tightens Disclosure Expectations for AI-Assisted Inventions
The EPO’s public guidance is making one point harder to ignore: for AI-assisted inventions, the real issue is no longer whether AI was used, but whether the application clearly ties a claimed technical effect to a reproducible technical solution and identifiable human technical contribution. For filings built on model training, data selection or AI-assisted discovery, that goes directly to patentability, sufficiency and later validity risk.
The point many applicants still misread is disclosure. The EPO does not generally require applicants to hand over the specific training dataset. But where the technical effect depends on characteristics of the training data, those characteristics may need to be disclosed in enough detail to reproduce the effect. Combined with the EPO’s repeated human-centric framing of AI use in examination, the practical message is clear: internal records, contribution mapping and a better-written specification now matter more than generic references to “AI optimisation”.
This is not just about naming AI in the file
A lot of commentary around AI patents still assumes the key compliance question is whether an applicant should openly state that AI tools were used. That is too shallow. What current EPO materials actually suggest is a change in examination emphasis: examiners are more likely to ask, earlier in the process, what exactly delivers the technical effect, which steps are technically meaningful rather than abstract, and where the human-led problem definition and validation sit in the invention story. Once those questions move forward, drafting discipline matters much more than it did a few years ago.
This is especially relevant for teams using large language models in drafting or machine learning in screening materials, sequences or process parameters. AI can assist discovery. But a patent application cannot stop at a black-box narrative that says the model produced a better result. The less concrete the story, the more likely the file will attract pressure on reproducibility, technical effect and inventive-step reasoning.
Human contribution has to be visible in the technical story
One practical weakness in many AI-assisted filings is that real human contribution gets flattened out on paper. Valuable technical contribution often lies in setting the problem, choosing constraints, excluding weak candidates, designing experiments, interpreting outputs and deciding what was actually validated. If none of that is visible in the application, the document can read as if the core inventive contribution came from system output alone and the human role was merely administrative.
That is where applicants lose control of the narrative. A better approach is to describe, with restraint but real specificity, how humans defined the technical target, selected operational parameters, connected model output to technical effect and verified that link experimentally or through implementation. Inside the company, that also means inventorship analysis, lab notebooks, version control and development records can no longer be treated as separate housekeeping exercises.
Dataset disclosure is becoming more selective and more demanding
This is the point that deserves the most careful reading. The EPO’s guidance draws a nuanced line: applicants do not generally need to disclose the specific training dataset itself, but if the technical effect depends on characteristics of the training data, the characteristics needed to reproduce that effect may have to be disclosed. In practice, that means the Office is not asking applicants to dump the whole data warehouse. It is asking them to explain the part of the data story that actually matters to the claimed technical result.
That changes how specifications should be written. Phrases such as “the model was trained on historical data” or “parameters were optimised using a sample set” are often too vague to carry real weight. Applicants increasingly need to explain the relevant scope of the data, how it was selected or labelled, which structural or statistical characteristics matter, and how model output was connected back to a validated technical outcome. In areas such as new materials, sequence design and process optimisation, that bridge between data and effect is often where the application stands or falls.
The weakness often appears later, not on filing day
Many teams still assume that if an AI-related application gets through initial examination, the difficult part is over. In reality, weaker drafting tends to resurface later. Opposition and revocation disputes can bring sufficiency, technical character, inventive step and evidential credibility back into the same conversation. If the application rests on an under-explained model process and the applicant cannot show a coherent human-led validation path, patent durability becomes much more fragile.
That is the practical takeaway from the EPO’s current direction. The real question is not simply whether AI was used. It is whether the filing can still explain, years later and under pressure, who contributed what technically, why the claimed effect is credible across the scope claimed, and whether the record behind the invention still holds together. Applicants that fix those three issues early are far more likely to end up with enforceable assets rather than fashionable but brittle filings.
