The Patentability of Inventions Involving Artificial Intelligence

Patent practitioners and other stakeholders in Artificial Intelligence (AI) have recently had the opportunity to meet at a conference (“Patenting Artificial Intelligence”, Munich, May 2018) organized by the European Patent Office (EPO). A timely event, judging by the way AI is currently driving technology and society into the 4th industrial revolution.

The patentability of AI inventions directed to specific hardware (e.g., resistive processing units) is not an issue: such inventions are in principle patentable, just like any other machine. More questionable, however, is the patentability of AI innovations implemented on usual computer hardware. A typical example is the use of a machine-learning model for email filtering, where the model is trained and run in software on a classical computer. But as expected, speakers confirmed that such inventions are subject to the same regime as computer-implemented inventions (CIIs) at the EPO. That is, software implementations of AI inventions must have “technical character” (e.g., have a technical impact on the physical world) in order to be patentable. And even if the concept of “technical character” is often not easy to reconcile with the concept of “software implementation”, the EPO has established clear and stable criteria to assess the patentability of such inventions. In comparison, the current situation for CIIs is still blurred in other jurisdictions, starting with the United States Patent and Trademark Office.

Of particular importance, however, is the clarity of the patent claims. Clarity is needed to enable relevant comparisons with the prior art (is the invention novel, inventive?) and a sound appreciation of the extent of protection conferred by the patent, once granted. There, two comments are in order. First, AI inventions are typically claimed using abstract terms (as usual with CIIs), which often gives rise to clarity objections from the patent offices. Second, AI is still a relatively new field, in which new terminologies swarm. The lack of stable, universally accepted terminologies necessarily impacts the interpretation of the claims and their clarity. Thus, the terms of the claims should all the more be carefully defined in the description. 

Thus, if AI inventions are in principle patentable, the patent documents should be very thoughtfully drafted in the first place, especially for software implementations. Now, this conclusion evokes nothing more than a saying that is always applicable to patents. Namely, the very right thing to do for applicants and patent attorneys alike is to do the things right (and, this, right from the beginning).

Sébastien Ragot