Better Software Patents

Will “Trained” AI Kill Your Patent?

If you draft patents for AI innovations, you must pay close attention to this new decision by the European Patent Office.

The central question:

The case centered on a method for controlling continuous chromatography using a “trained and validated mathematical model”. The Board of Appeal revoked the patent, finding the claims unclear.

The main arguments:

  • Simply stating the model was “trained and validated” didn’t adequately define its characteristics or how it differed from an untrained/unvalidated model.
  • The Board highlighted that one can’t readily determine if a given model has been trained and validated without comparison to an untrained/unvalidated counterpart.

What I find important:

This decision seems at odds with the practice of some EPO examiners who require applicants to include “trained” as a qualifier for their ML model in the claims, arguing that an untrained model wouldn’t function as intended.

What to do now

  1. Reconsider including terms like “trained” and “validated” in your ML model claims. Focus on the technical features of the model and its role within the invention.
  2. Read the whole decision here (it’s in German).
  3. Join my Beyond Best Practice newsletter to receive more tips like these.

All the best,
Bastian

If this was helpful, you’ll love my mailing list! You'll get software patent drafting tips, helpful checklists, and a 20% discount on my seminars. Join today:

Join 1,088 happy subscribers 🙂

Better Software Patents