Much has changed regarding AI patents. Today, the bar is far higher than in 2014: sufficiency of disclosure for medicine and biotechnology—now supported by AI—is a conditio sine qua non.
The Trained Model as Part of the Invention
In T 0161/18, the statement that the input data should cover "a wide range" was deemed insufficient. The application disclosed neither which input data were suitable for training nor a specific training dataset. Consequently, the training of the neural network could not be reproduced, and the invention was refused for lack of sufficiency of disclosure.
It follows that if a specifically trained model forms part of the invention, the application must enable the person skilled in the art to reproduce the training. The assay, measurands, data processing, target variable, and specific training data may all be relevant. However, no rigid minimum requirements regarding data volume and data variability should be derived from T 0161/18.
From Measurement to Technical Effect
Even an AI invention that is sufficiently disclosed must make a technical contribution. In T 1741/22, it was decided that the mere generation of further data from measurement data already collected from the human body does not constitute a technical effect. Therefore, the additional statistical processing of glucose measurement data could not prevent the refusal of the patent application.
The chain "biological measurement – AI – calculation of a further parameter" is thus not automatically technical.
T 0660/22 illustrates this for the microscopic analysis of cells in which image data were evaluated. The Board of Appeal distinguished between information about physical reality and indirect measurement. Correlations between physical quantities can demonstrate dependencies of biological states. Without proof of such a substantive relationship, these values are merely "data (...) to gain scientific knowledge about a (...) natural system" (G 1/19, Reasons 98), and determining them does not constitute an indirect measurement within the meaning of G 1/19.
Technical Feedback as a Guideline for Patentability
Things become particularly interesting where AI feeds back into the physical world. While extensive patent literature exists for such AI inventions, there is not yet a citable decision on this matter. The more integral the model is to the technical teaching, the more precisely the application should disclose the technical causal chain: measurand, training data and model training, technically relevant parameter, prediction, physical selection, and, where applicable, experimental feedback.
The lesson for biotech AI:
if a specifically trained model forms part of the invention, the patent application must enable the person skilled in the art to actually reproduce this training. Whether feedback into the physical world will also become mandatory in the future for the technical character and sufficiency of disclosure of a biotech AI invention remains to be seen. In other jurisdictions, this is already the case.