Definition of AI (Art.3)

The definition of AI is important as it will dictate the products that fall under the scope of AIA, as well as filtering out products using the term AI incorrectly. A global concern on the definition of AI has been raised, leading to a broad range of definitions in official documents. Some of which are presented in Table 1 - Definitions (TC260, 2021) that make use of the term “intelligence”. These are ambiguous however, as there is no globally accepted definition of intelligence (Shane Legg, 2007) (Wang, 2019).

AIA has approached, in the initial 2021 release (EC, Artificial Intelligence Act, 2021), the definition in a more straightforward manner by stating specific techniques for developing AI in Annex I. One could argue that some of the Annex I AI techniques and approaches are broadly defined (e.g., Statistical approaches, inductive (logic) programming), allowing space for methods not widely accepted as AI to fall under the AI definition. However, this approach is straightforward and specific in comparison to more general AI definitions found elsewhere (Table 1). As Wang (Wang, 2019) comments, the use of an AI definition that only considers what is established methodology would exclude new techniques and might be an obstacle to innovation.

The recent EU Council General approach (Council, Nov 2022), defines AI more broadly (see Table 1). However, in the proposal there are additional statements refining (paragraph 6) on what is considered AI. Paragraphs 6a and b further refine AI approaches:

• “In particular, for the purposes of this Regulation AI systems ..., using machine learning and/or logic and knowledge based approaches …”
• “A system that uses rules defined solely by natural persons to automatically execute operations should not be considered an AI system”
• “Machine learning approaches include for instance supervised, unsupervised and reinforcement learning, using a variety of methods including deep learning with neural networks, statistical techniques for learning and inference (including for instance logistic regression, Bayesian estimation) and search and optimisation methods.”
• “Logic- and knowledge based approaches include for instance knowledge representation, inductive (logic) programming, knowledge bases, inference and deductive engines, (symbolic) reasoning, expert systems and search and optimisation methods.” 

It is important to note that, irrespective of the proposed definition, it is the responsibility of the conformity assessment body, when a third-party assessment process is required, to judge and challenge whether a proposed AI product falls under the definition. The ambiguity in definitions has been encountered in other sectorial legislations and additional guidelines were issued to provide more clarity; Medical Device Guidance documents, endorsed by the Medical Device Coordination Group (MDCG), are an example; MDCG 2022-5 (MDCG, 2022-5) provides guidance on borderline products between Regulation (EU) 2017/745 and Directive 2001/83/EC.

This blog post is an excerpt from our forthcoming whitepaper: Ethical and trustworthy Artificial Intelligence. To download our other medical device white papers, please visit the Insight page on the Compliance Navigator website.