Looking across the pond, the Intellectual Property Organisation has put together a whitepaper on best practices in protecting inventions relating to AI. While it focuses on US law, many sections span jurisdictions and are equally applicable to patenting AI in Europe. They are worth looking at if you are involved in preparing patent applications for AI inventions.
The whitepaper is organised into sections on preparing AI patent applications, prosecuting them, and enforcement. The section on preparing applications explores how to extract information from inventors, and it was great to see that it asks all the right questions for Europe, as well. There is even a section on international considerations, including the EPO. For those who want a bit more background on how patentability assessment works for Europe, you can take a look here for the general approach and here for a review of recent case law.
The section on prosecution is based on US law and practice. Still, it is always interesting to see the convergence with the EPO approach, at least in the "technical purpose” route to patentability at the EPO goes. The approaches at the USPTO and EPO are somewhat difficult to compare since there is a greater focus on inherent subject matter eligibility in the USPTO. On the other hand, this hurdle is vanishingly low at the EPO. The "action" is happening at the EPO when assessing inventive step based only on those features that contribute to the all-important "technical effect".
Nevertheless, there are large areas of technology where the two approaches are likely to produce similar effects. One notable exception relates to improvements in technical fields that are considered to be non-technical by the EPO, such as much of natural language processing or information retrieval. Another exception concerns improvements residing in an improved computer per se, where the EPO will discount algorithmic improvements, such as increased speed or reduced compute resources, unless they are motivated by considerations of the computer's internal functioning. On the other hand, the EPO's approach is typically favourable when assessing biomarker or diagnostic assay-based inventions. Interestingly, however, the authors' recommendations to help with US prosecution are equally helpful for prosecution at the EPO, be it that the law is different.
One huge consideration in drafting claims for AI inventions is the detectability of the features of the claims. Irrespective of the jurisdiction you are looking at, if one of the features of the claims cannot be detected in competitor activities, enforcement can be difficult, if not impossible. The whitepaper has an excellent section on how to draft detectable AI claims under the headings of visible features; features that can be reverse-engineered; features likely to be disclosed in competitor’s public documents; and features that are likely to be published in competitor's public documents or are part of a standard. Many applications of AI technology can be run in the cloud and be provided as Software as a Service, so detectability is a genuine concern. This is something on my list to write about more in the future, but for now, the detectability section in the whitepaper is a great starting point.