The sufficiency of disclosure of AI inventions

Abstract The complex and data-driven nature of artificial intelligence (AI) raises questions for the sufficient disclosure of patent applications in this field. What are the European patent disclosure requirements for AI inventions? One challenge is that, prior to training, AI systems can be considered generic models. But after training, they transform into specialized AI systems to solve a particular problem. This transformation requires training data, making it an integral part of the AI system’s definition. But to what extent is the disclosure of the training data or training process necessary for patent disclosure? The Boards of Appeal of the European Patent Office (EPO) first dealt with this challenge in case T 0161/18, which involved a medical AI invention to calculate cardiac output. It held that the specialized artificial neural network (ANN) in the patent could not be carried out by a person skilled in the art due to insufficient disclosure of input data suitable for the training of the ANN or at least one data set suitable for solving the technical problem. Furthermore, without specialization, the invention lacked an inventive step. But, is it always necessary to disclose the input data or at least one data set suitable for solving the technical problem? Are there alternative ways for applicants to satisfy the disclosure requirements for AI inventions? And what evidence is there that patent applicants are disclosing specific details of the AI/machine learning (ML) training or specific AI/ML model architecture? In this article, we analyse case T 0161/18 and subsequent sufficiency of disclosure decisions (T 1539/20; T 0606/21; T 1526/20; T 1191/19) and consider these foundational questions for applicants drafting patent applications with claims directed to AI inventions. We also analyse the EPO’s examination guidelines on sufficiency of disclosure for AI inventions, which were updated in early March 2024.