IP Protection of Artificial Intelligence in Europe: Tailor-Made Solutions Required

Introduction

IP institutions around the world are addressing a variety of issues associated with AI (e.g., for a U.S. perspective, see "Protecting Artificial Intelligence IP: Patents, Trade Secrets, or Copyrights?"; "When Innovation Invents: Artificial Intelligence Issues at the U.S. Patent and Trademark Office"; "When Innovation Creates: Additional Developments in Artificial Intelligence at the U.S. Patent and Trademark Office"; and "AI and the Biopharmaceutical Industry"). European stakeholders, i.e the EPO, the European Commission, national governments and IP offices have launched initiatives that may impact IP protection policies of AI innovations. The latest developments and tactical considerations are summarized below.

Choice of IP Rights 

1. What May Be Protected?

The general workflow of constructing an AI model solution for a defined problem includes the acquisition/preparation of training data, the design of an appropriate model architecture (e.g.choosing suitable AI algorithms, setting initial parameter values), model training, evaluation, and optimization. Additionally, large high-quality, representative training datasets are extremely important for reliable performance of an AI model when processing new data. 

Business value may therefore be found in protecting (i) AI models and/or algorithms; (ii) software in which the models/algorithmsare embedded; (iii) training, evaluation and/or optimization strategies; (iv) training data; and (v) result data (i.e.work product). IP protection may be sought for all or a subset of these potential assets. 

While copyright provides different options for protection of AI such as source code written by a programmer, data base structures, data collections and data models to be used for AI (which we will address in another Commentary), further IP rights suitable for other aspects of an AI innovation are discussed below. 

2. European Patents

A. General Principles

Although mathematical models and algorithms are not patentable under the European Patent Convention ("EPC"), AI inventions are generally patentable as a subgroup of computer-implemented inventions ("CIIs") (see "Patenting Artificial Intelligence and Machine Learning Innovations in Europe").

According to established EPO case law, inventiveness can be assessed by considering only those features that contribute to a technical character of an invention (G-VII, 5.4 of the Guidelines for Examination at the EPO). Howevera non-technical feature (e.g.algorithm), which interacts with technical features to solve a technical problem, should also be considered (COMVIK, T 641/00). Protectable subject matter differs from country to country depending on what it is directed to (e.g.product vs. method) and the technical field.

B. Latest EPO Case Law on AI Patenting

In a recent decision (T 697/17 in October 2019), the EPO's BoA "dissected" existing case law and emphasized that to identify technical contributionit is necessary to examine if a non-technical feature under dispute was chosen based on technical considerations aimed at achieving a technical effect. If going "beyond 'merely' finding a computer algorithm to carry out some procedure" such technical considerations may result in technical contribution. In T 731/17 in January 2020, the same BoA further addressed the assessment of technical contribution in line with its previous opinion. 

When drafting a patent application in the CII or AI field, it is important to identify technical considerations and motivations behind non-technical features that contribute to the solution of a technical problem.

C. AI as an Inventor

Earlier this year, the EPO published reasoned decisions on the refusal of two patent applications (EP18275163; EP18275174) designating AI as an inventor. The EPO concluded that for the mandatory designation of an inventor for a European patent application the inventor must have a legal personality. 

3. Utility Models

In a plurality of countriestechnical inventions can also be protected by utility models, sometimes called Utility Certificates (see "French PACTE Act: Provisional Patent Application and the Evolution of the Utility Certificate"). Protectable subject matter differs from country to country depending on the technical field and the type of subject matter (product vs. method). The requirements for acquiring a utility model arein most countriesless stringent. Lacking substantive examination in most countries, the registration and publication process is simpler, faster, and cheaper. Utility models should be considered as an expeditious, less costly way to obtain IP protection for often "quickly" outdated AI inventions. 

For example, in Germany it is possible to "branch-off" a utility model from an earlier patent application while maintaining priority and the filing date of the patent application. The "branch-off" utility model may be registered in a few weeks. This strategy may be considered to obtain an enforceable IP right while the patent application is examined. 

4. Trade Secrets 

Directive EU 2016/943 has widely harmonized trade secret protection across the EU. An owner is entitled to protection against unlawful acquisition and use of a trade secret and is entitled to damage claims in case of misappropriation.

Trade secrets may provide the broadest scope of IP protection. To rely on it, the directive requires an owner to implement internal trade secret policies and adequate measures to protect know-how. Relevant technological information should be documented to prove possession of the know-how in case a litigation arises. 

Mariana Zaichuk of the Munich Office is a coauthor of this Commentary.