Generative AI: Can data models be protected?
Cristina Mesa, socia de Propiedad Industrial e Intelectual de Garrigues
Data models, which are the “brain” of artificial intelligence systems cannot be protected under intellectual property law. In principle, they are not patentable either, but they could perhaps be protected as trade secrets. Let’s take a look at what is the angle in Spain.
In the process of adopting new artificial intelligence systems, the focus as far as intellectual property is concerned, is now on the contents used for training (usually accessible online) and on whether unauthorized use can be considered a massive infringement of IP rights (see our post here). Without wishing to play down the importance of this issue, today we will be addressing a different matter: the possibility of protecting data models, the “brain” behind these systems.
The first thing that we need to bear in mind is that although generative AI uses software to a greater or lesser degree, it is by no means its most valuable part. When talking about deep learning, data models and training data are usually more important, with code taking a back seat. This change in the value change raises new questions as to how to protect the investment made in developing these systems on the one hand and the results obtained through their use on the other. Let’s take it in stages.
A data model is NOT a database
Data models are the outcome of training and fine-tuning through the massive use of content or data). They act as the “brain” of the system, and are what allows it to generate quality content. The key to an effective system is the massive ingestion of data. It is this data ingestion that enables the model to identify and learn patterns and characteristics of the information and based on this learning, to make predictions, classifications or generate new content.
We should remember that according to the Directive on the legal protection of databases (1996), databases are collections of data which are systematically arranged and can be individually accessed. Although it is true that you can feed data models with data from a database, they should not be confused with a database. Data models do not store the training data as such, but rather retain an abstract representation of that data. The model contains a series of weights, parameters, and connections, but not the original data themselves. Just as we learn as children, what a cat is for example, we do not remember every single cat we have seen, but rather a general idea or the concept of a “cat”, based on the common characteristics we have observed through our experience.
For the same reasons, the protection of databases by the sui generis right, which protects the investment effort in creating the database, from a quantitative and/or qualitative standpoint, is not applicable either. In other words, this sui generis right could be applied, where appropriate, to the data sets we use to train AI systems, since there are strong arguments to maintain that an investment has been made in the compilation, including, as required by intellectual property law, the use of economic, material and human resources in obtaining, verifying or presenting the contents.
Data models are not (in general) patentable either
The Patents Law (2015) provides that for a patent to be patentable it must meet three requirements: (i) novelty, (ii) inventive step and (iii) industrial application. The fact that it is practically impossible to patent a data model is due to two main factors. On the one hand, a data model can be defined as a mathematical method, a structure or scheme that defines how data is organized, stored, and processed. As such, it cannot be protected by patent law which does not consider "discoveries, scientific theories and mathematical methods" to be inventions. In addition, data models, in their pure form, do not have a specific application that solves a technical problem and, therefore, would not meet the "industrial application" requirement either.
However, it is true that where the data model forms part of a broader technical solution, which solves a specific problem, provided that it meets the novelty and inventive step requirements, its patentability could be assessed.
A data model COULD be a trade secret
In the absence of protection through intellectual property law, it is possible, in my opinion, to turn to trade secrets. The Trade Secrets Law (2019) is an extremely useful tool to protect the truly valuable assets of companies that focus strongly on digital innovation, which are increasingly moving away from lines of code.
The law provides a very broad definition of trade secrets, extending it to any information or knowledge, including technological, scientific, industrial, commercial, organizational or financial information or knowledge that (i) is secret, (ii) has commercial value because it is secret, and (iii) has been subject to reasonable steps to keep it secret. It is, in short, the regulatory positivization and independence of the principles already required by the World Trade Organization (WTO) via the Agreements on Trade-Related Aspects of Intellectual Property Rights, and applied by the Spanish courts through unfair competition legislation.
For example, although we have not found any case law that deals with the specific problem of data models, case law on protection of algorithms as trade secrets is useful. An example of this is order number 229/2018 issued by Barcelona Commercial Court no. 4 which recognized the possibility of protecting algorithms as trade secrets.
However, protecting data models through legislation on trade secrets has its pros and cons, which can be summarized as follows:
Conclusion
The growing value of data models in deep learning-based AI leads us to reflect on the possibility of gaining a competitive edge by protecting these models. Intellectual property law does not appear to facilitate obtaining exclusive rights, but data models could, in principle, be protected as trade secrets. However, these models must have an economic value because they are secret, and we must also be able to prove that we have used reasonable steps to keep them secret. It is time to rethink the trade secrets protection policies we use in our companies, including reviewing employee, supplier and other third-party clauses, security measures and forms of marketing.