AI according to the European Commission - draft AI Act

AI according to the European Commission

The European Commission presented its proposal for the Artificial Intelligence Act (AIA) to the world on April 21, 2021. This regulation can be seen as the AI version of the GDPR. Whereas the GDPR focuses on personal data protection, the AIA focuses on the use of artificial intelligence. The first question then is: what constitutes AI within the meaning of the AIA? Article 3 of the AIA sets out the definitions from the regulation. Article 3(1) AIA defines what an AI system is, based on the technical and functional requirements.

TECHNICAL REQUIREMENT

The first requirement for an AI system under Art. 3(1) AIA is that an AI system is developed using techniques and methods from Annex I AIA. This Annex describes, in three sections, the techniques and methods covered by this regulation. The three categories are as follows:

• Machine learning in the form of supervised, unsupervised, deep and reinforcement learning and the methods required for these. An example is Google's chess computer, AlphaZero.
• Logic and knowledge-based methods such as other expert systems, inductive programming, knowledge representation and systems that reason via inference, deduction or symbolic reasoning. An example of this category is the medical diagnostic system MYCIN.
• Statistical methods and operations research techniques such as optimization methods. For these methods, you can think of scheduling delivery networks.

In summary, this concretely means that when software is not explicitly programmed, it quickly becomes AI. In fact, any form of learning from data falls at least under category 1. Systems that do not learn but combine data fall under category 2. Finally, there are methods that search for optimal solutions based on fixed formulas. These all fall under category 3.

FUNCTIONAL REQUIREMENT

If a system is developed using Annex I methods, then the second requirement becomes important. An AI system must meet the following functional requirements. An AI system must:

1. Based on human-defined goals
2. Generate output such as
   1. Content,
   2. Predictions,
   3. Recommendations, or
   4. Decisions that affect the system's environment.

So, what the AI system produces matters. The reason why or for what is not important to the functional requirement. These substantive reasons are only considered when assessing which risk category an AI system falls under; is it a prohibited application, a high-risk application, or an application for which there are transparency obligations?

CONCLUSION

To determine whether certain software should soon be treated as an AI system, it is necessary to consider whether the technical and functional requirement have been met. If the software is sold with the promise that it is intelligent or that it contains AI, then the software will most likely have to be treated as an AI system. Also, any form of machine learning in your software will mean that it falls under the concept of an AI system.

If you have further questions about AI systems and would like advice on the implications of qualifying your software as an AI system, please contact us.