Artificial intelligence and patents

The expression covers the body of questions raised by the patentability of inventions that implement artificial intelligence techniques, machine learning, neural networks or statistical models, together with inventions made with their assistance. Positive law contains no special regime: such inventions are governed by the ordinary conditions of Article 52 of the European Patent Convention and Article L. 611-10 of the French Intellectual Property Code, which exclude mathematical methods and computer programs, but only to the extent that the application relates to those elements as such.

Patentability of inventions based on machine learning

A learning algorithm, taken in isolation, is a mathematical method excluded from patentability. The position differs where the claim places it in a technical context: control of a device, processing of a physical signal, improvement of the internal functioning of a computer. The European Patent Office then applies the framework developed for computer-implemented inventions. Technical character is assessed on the claim as a whole, and only the features that contribute to that character are taken into account in the assessment of inventive step.

Disclosure and reproducibility requirements

The nature of learning models raises a specific difficulty with regard to the sufficiency of disclosure required by Article 83 EPC and Article L. 612-5 IPC. The applicant must disclose the invention in a manner that enables the person skilled in the art to reproduce it. Where the technical effect depends on the quality of the training data or on a particular architecture, the description must account for it, failing which the claim may be held irreproducible or found to rest on an effect whose plausibility has not been established. The disclosure of the data, parameters and training protocol then becomes part of the application, which forces a choice between patent protection and the preservation of trade secrets.

Litigation and proof of infringement

A model is often run on remote servers, sometimes outside the territory covered by the patent, and its parameters cannot be observed from the outside. Proving infringement therefore requires recourse to the saisie-contrefaçon, to the right of information and, where appropriate, to court-ordered expert examination in order to access the code, the data and the execution logs.

Key points

An invention that uses artificial intelligence is patentable under the ordinary conditions provided that it produces a technical effect, the algorithm alone remaining excluded. The description must enable that effect to be reproduced, where necessary by disclosing the data and the training protocol. Before any filing, check that the claims relate to an identifiable technical application and choose, for each element of the system, between patent and secrecy. In the event of infringement, prepare the evidence at an early stage by identifying the observable elements of the opposing system.

---

This glossary entry is general in scope and does not constitute legal advice.

Dhenne Avocats litigates patent disputes before the Paris Judicial Court, the Paris Court of Appeal and the Unified Patent Court. When the question arises in a real case, the right moment is almost always before proceedings are issued.

Patent litigation · Unified Patent Court · UPC case law · Tell us about your situation