AI patent strategy: wHY YHE ‘black box’ iS NOT ENOUGH

Simply stating that “an AI model has been trained on a dataset” often leads to rejection. A successful patent strategy requires the disclosure of a concrete ‘recipe’. The applicant must provide insight into what exactly the model optimizes (such as target functions or loss functions), what input data is required (features and labels), the model configuration, and how the training procedure works exactly. If the applicant cannot disclose their specific training data in full in order to protect trade secrets, it is generally important to provide reproducible instructions on how an expert can collect, filter, and label comparable data himself.

Another common pitfall is claiming overly broad protection. If the applicant claims that the AI works for images, text, and audio, the technical description must support this full scope of the claim. Broad claims are commercially very valuable, but require at least one detailed practical example and a technical justification as to why the method can be successfully generalized across that entire breadth. If this broad justification is lacking, the claims will most likely have to be narrowed down in the granting or opposition proceedings.

Purely conceptual claims often do not hold up. Concrete evidence is typically crucial: describe empirical results or at least a plausible, testable argument that demonstrates that and why the invention works.

In short, a robust AI patent strategy requires a smart balance between maximum commercial protection and in-depth technical transparency. Therefore, work with your patent attorney to forge a robust and reproducible foundation for your AI patent in a timely manner.

Want to know more?

Read the dossier ‘Patenting AI inventions’