Hi Bjoern and other members of the group, 

The EPO guidelines will only be made public in October. 

Regarding obviousness: taking an "AI" as a reference for assessing the inventive step (in the established "problem-solution approach") would lead to a series of unsolvable issues: how to define such AI? (an SVM? a random forest? a deep learning/NN? but which architecture, then? how many layers? Which training set, which paramters, which loss function? etc...), what would be considered a metric of "obviousness" for such a system? 

This lack of proper definition could lead to assuming that all solution that could be reached by an "AI" system would be, as such, rendered obvious in the sense of the patent law (take the example of the molecule you highlighted in your post).

This is a shortcut that some have taken. But this disregards the existing legal framework (as well as the purpose of the patent protection).

Indeed, for reasons of due process and legal certainty, the establishment of obviousness relies on the demonstration of a causal link between a closest prior art and a (claimed) solution: the question to answer is: "would the skilled person arrive at the claimed solution based on the teaching of the closest prior art (and the common general knowledge in the field) while attempting to solve the objective technical problem".

One (the patent examiner, the opponent or the judge) must demonstrate that such a causal link exists. It is not sufficient to claim that a skilled person (or an "AI" or a "skilled person+AI") *could* arrive at the claimed invention but one must provide sufficient convincing elements to demonstrate that the skilled person *would* indeed arrive at the claimed solution.

If such a demonstration entails the use of existing automation tools, this is fine (there is abundant case law on this matter, in various technical fields), but the existence of such automation tools (AI being one of these tools) cannot used as an general argument to argue against an inventive step (and it would make an easy, and rather arbitrary, argument to prevent granting any invention and to invalidate many). 

Practically speaking, the existence of automation tool to supplement the skilled person knowledge is not novel and the skilled person is always considered aware of the available tools, be they mechanical or algorithmic. A lack of obviousness only arises if it can be shown that the skilled person (+ "AI") would necessarily arrive at the invention. In practice (again, referring the existing case law from the boards of appeals) if, for example, the combination of parameters in which the "AI" system expresses the solution is known and if the number of parameters being optimized is small (e.g. below 4), or if the optimization problem is deemed "convex" (which means that it can be proven that the system does converge to a single, unique, solution), this could constitute an argument in favour of obviousness. 

However, the non convexity of Deep NNs (and the potentially huge dimensionality of both the input space and their internal degrees of freedom) renders the argument moot. 
The present legal toolkit is actually quite well adapted to coping with the latest development in automation with regard to the assessment of inventive step (again, this has been discussed rather extensively in the past). 

To come back to the example of the molecule: if a drug that cures malaria is identified using a combination of in silico screening, machine learning generative processes and biochemical simulation, a patent may be obtained as long as the requirements of novelty, inventive step and industrial applicability are met. If it cannot be demonstrated that such a molecule would be (necessarily) obtained (and not just "could" be obtained)  based on the available prior art, the condition of art/ 56 EPC (inventive step) will be deemed met.

  I hope this provides some clarification...  There is obviously much more to say (but I have already been quite long!) and this does not mean that AI does not open new questions wrt IPRs (including patent law), but obviousness is neither the most pressing one nor the most essential. In any case, I hope this will lead to more fruitful and insightful exchanges!  

Dear Bjoern,

I read this article, on that topic, sometime ago and found it somehow useful: The inventor’s new tool: artificial intelligence – how does it fit in the European patent system?, Peter Blok, E.I.P.R. 2017, 39 (2), 69-73. However, in my opinion, the questions remain unanswered.

​As mentioned already, my understanding is that the EPO guidelines will be made public in October this year.

Best regards,

​Barbara

Thanks for your answer Marina, it is very interesting to keep in touch for these topics and see how, as seen in the article you mention, AI art-works are more and more present.