Who will the AI system belong to?

What may the AI/ML developer do with the data? What should a customer expect from the AI/ML developer?

How can you shape the relationship between a developer of machine learning ('ML') or artificial intelligence ('AI') models and its customers? That is what we will discover in this case study. For now, the customer has their own data(set) and a specific question they would like to see answered. The developer wants to answer the question in exchange for payment. How can they arrange their collaboration? In this blog, I set out three possible relationships that suit different goals of the customer and the developer.

In this article, model means a chosen ML/AI model with enough settings to be trained, and AI system means a trained model that can be put into action.

1. AI SYSTEM AS A SERVICE

The client comes to the developer and provides her data and the question she wants to see answered. The developer goes to work. He cleans up the data, clarifies the question and then chooses a model. With this model and the preliminary steps, he can make new AI systems with more data in the future. The developer now has an AI system which can start making predictions based on new data. This AI system is put on a server by the developer, through which the customer can send new data via a website or an API and get the results.

So, in this case, the customer has only bought a service and the developer has a copy of the data(set). The customer does not have to worry about the systems or technology as this is all under the care of the developer. Furthermore, it is advantageous for the developer that he only has to explain his AI system and does not have to transfer it in its entirety, which reduces questions regarding copyright and other intellectual property rights.

Features for the customer:

• Subscription to the model.
• Access to the AI system.
• Hosted by developer.

2. AI SYSTEM AS SOFTWARE

The customer comes to the developer in the same way as before, but with an additional requirement. She wants to be able to inspect the AI system herself and run it on her own servers. The reasons for wanting to inspect the AI system may stem from a legal requirement or an ethical requirement related to accounting for the results of the AI system. As a result, the developer loses the revenue from offering the service of option 1. Here, the developer still needs access to the data and is also likely to do the pre-processing of it. After this, he will again choose a model and with this he can train an AI system using the data. Now, however, he will have to give this AI system to the customer under some kind of licence. The customer can then use that AI system herself for predictions and also do the analysis and explanation of the AI system, thus reducing later risks for the developer.

Features for the customer:

• Training as a service.
• License on the AI system.
• Self-hosted.

3. AI DEVELOPMENT CONSULTANCY

As the most extreme case, the customer may want the developer's expertise but wants to retain control over her data. In this case, the developer will have to go to the customer physically or digitally. There, the developer will go through the development process with the customer and then the customer will keep control over the data(set), the model and the AI system. Here, the customer will have to bear even more of the responsibility (and liability) for the AI system. However, the customer does know for sure that she has full control, and this can be of great importance in certain sectors.

Features for the customer:

• Access to expertise.
• Control over the data, the model and the AI system.
• Full in-house hosting and development.

CONCLUSION

The parties will need to consult together on which setup is most convenient for them and how things like access to data will be arranged. Furthermore, the developer should be aware that even if he cleans, indexes, organises and then edits the client's raw data to make the data ready for training the model, he will not be allowed to simply use this data for other projects. The client might have a database right to this (edited) data.

Tips for clients

• Decide how much access you need to the AI system for things like legal requirements of explaining the analyses and other compliance aspects.
• Consider how much responsibility you want to bear yourself as an organisation.
• Decide whether you want a one-off AI system or whether you want to keep upgrading the AI system with new data(sets).
• Check whether you have received sufficient safeguards to work ethically with the AI system.

Tips for developers

• Discuss with the client how they want to use the AI system.
• Determine the division of responsibilities with your customer(s).
• Consider what services you want to offer to your customer(s), from expertise to full hosting or in between.
• For possible copyright protection of your AI systems and models, it matters whether personal choices have been made and whether the model is sufficiently complex.

This article was written by Robin Verhoef and Jos van der Wijst and previously published on BG.legal.