Technical character first, inventive step next
“Can AI be patented?” is too broad a question. Article 52 EPC excludes, among other things, mathematical methods, computer programs and business methods as such, while allowing patents for inventions meeting the legal requirements. The analysis concerns a particular claim and technical solution, rather than an artificial intelligence label. [1]
Our editorial proposal separates two questions: does the claimed subject matter as a whole have the required technical character, and do its differences from the prior art support inventive step? Merely naming a processor or neural network does not complete the second inquiry. The team should explain the problem, technical means and causal relationship with the effect. These are project organisation questions, not a completed patentability opinion.
What T 0641/00 actually establishes
In T 0641/00 of 26 September 2002, known as COMVIK, the Board held that inventive step takes account of features contributing to technical character; features making no such contribution cannot support it. A non-technical aim may appear as a constraint in the problem. The decision does not exclude the entire software industry. [2]
The 2026 EPO Guidelines, G-VII, 5.4 develop this approach for mixed claims. A feature’s contribution in the invention’s overall context matters, rather than mechanically labelling a word technical or non-technical. We propose a table recording each feature, its system effect, evidence and difference from a reference document. A new business rule alone does not explain a technical solution for implementing it. [3]
Two fictional projects, different questions
One project automatically sets discounts using customer segments. Its team describes faster sales and higher revenue, but has not identified a particular technical solution. Another processes an industrial sensor signal and changes device control under interference. That project requires examining the processing path, interaction between components and actual effect. Both are fictional editorial examples; neither establishes patentability or non-patentability of every implementation falling under those descriptions.
EPO guidance distinguishes technical AI applications from purely linguistic classification; implementation-specific analysis remains necessary. [4] A changed purpose, implementation or claim may require a fresh assessment. Examples in the Guidelines should not be read as a catalogue of products automatically entitled to a patent without examining the other requirements.
Preparing evidence of the effect
Our proposed experiment record identifies the system version, hardware environment, inputs, baseline, performance measure and conditions in which an improvement appears. If the team claims lower latency, document the measurement method and workload. If it claims more robust control, describe the disturbance. A result obtained with one prototype should not become an unconditional statement about every imaginable implementation. State the test’s actual boundary and any assumptions needed to interpret it.
Preserve negative results and configurations in which the advantage disappears. They can help distinguish essential features from incidental ones. Separate a model’s statistical test performance from the operation of the complete device. This editorial evidence practice does not imply that every application requires the same experiment. The supporting explanation depends on the claimed solution and assertions, and should remain tied to what the team can actually demonstrate.
Training data and sufficient disclosure
Article 83 EPC requires disclosure enabling a skilled person to carry out the invention. A promise that an unspecified model will produce the desired outcome is therefore insufficient as a working description. Before drafting, the team should organise what is needed to reproduce its solution and distinguish those essentials from merely organisational details. [5]
Where reproducing a technical effect requires particular training-data characteristics, EPO guidance requires sufficient disclosure of those characteristics, not necessarily the actual dataset. [4] We propose separately recording provenance, sampling, labelling and necessary properties. Rights to use the data require their own review: the ability to explain a model does not settle every question about obtaining and using its material.
From demonstration to a reviewable technical package
A working package can contain a flow diagram, implementation variants, comparison table and disclosure history. Engineers should describe what they built as well as what the presentation promises. A reviewer should be able to identify the passage supporting each assertion. Version control should prevent a test result from accidentally being assigned to a different architecture or dataset. Clearly mark planned functions that have not yet been implemented or evaluated.
The output should identify questions for further examination: relevant prior art, important differences, effects, reproducibility and a defensible scope. Inventive step concerns what would have been obvious to the skilled person in light of the prior art. [6] This guide describes the EPO approach; it does not automatically transfer outcomes to Poland, the United States or other systems, and it does not replace FTO.
An additional useful output is a terminology list explaining the technical meaning of words used in the claims. If “adaptive” refers both to updating a model and adjusting a sensor threshold, clarify the difference before drafting. Review the drawings too: an arrow between modules may suggest a relationship that the actual code and architecture do not implement. A coherent package makes it easier to distinguish an established mechanism from an attractive but untested future function.
Sources and further reading
- 1European Patent Convention — Article 52 ↗
EPC, art. 52
Exclusions “as such” and basic patentability requirements.
- 2EPO Board of Appeal, T 0641/00 — Two identities/COMVIK ↗
ECLI:EP:BA:2002:T064100.20020926; 26.09.2002
Headnotes concerning technical and non-technical features.
- 3EPO Guidelines 2026 — G-VII, 5.4 ↗
EPO Guidelines 2026, G-VII, 5.4
Mixed claims; checked 18 October 2026.
- 4EPO Guidelines 2026 — G-II, 3.3.1 ↗
EPO Guidelines 2026, G-II, 3.3.1
AI and machine learning: technical applications, effects and training data characteristics; checked 18 October 2026.
- 5
- 6
Rekel Patent — an intellectual-property information website. Publisher: Vadym Rekel.