Background
On 6 February 2025, the European Union (EU) Commission published guidelines on AI system definition to facilitate the first EU AI Act’s rules application (Guidelines). These Guidelines aims to assist in-scope persons resp. entities, in particular IT-providers whether a software constitutes an AI system or not.
As noted by the EU Commission, the Guidelines are not binding. Any authoritative interpretation of the EU AI Act will ultimately be of the competence of the Court of Justice of the EU. The Guidelines are subject to update and amendment over time depending on practical experience and use cases. The Guidelines are not (and cannot be) exhaustive due to the variety of AI systems and potential developments in the future.
From a Swiss perspective, the Guidelines are an opportunity to have an in-depth look at the definition of AI systems. To date, there is no comparable basis as the EU AI Act in Switzerland and thus, no legal definition of such systems. Until adoption of a formal definition, Swiss authorities might also consider the EU-definitions of AI systems to be persuasive and use it for own legal assessments (e.g. when assessing AI in the current legal framework such as data privacy or liability laws). A better understanding of the definition may also be useful to determine to which extent Swiss based providers, distributors or deployers for instance may fall within the extraterritorial ambit of the EU AI Act.
The notion of AI system
According to the EU AI Act an AI system is defined as a "machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments".
This definition comprises seven main elements: (1) a machine-based system; (2) that is designed to operate with varying levels of autonomy; (3) that may exhibit adaptiveness after deployment; (4) and that, for explicit or implicit objectives; (5) infers, from the input it receives, how to generate outputs (6) such as predictions, content, recommendations, or decisions (7) that can influence physical or virtual environments.
Comparison with OECD and ISO standards
The EU definition is largely similar to the definition already proposed by the Organisation for Economic Co-operation and Development (OECD). However, the definition under the EU AI Act is more detailed and specific than the existing definition set out e.g. in the industry standard ISO/IEC 22989:2022 ("Engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives"). It is worth noting that the International Organization of Securities Commissions (IOSCO), which published on 12 March 2025 its consultation report on "Artificial Intelligence in Capital Markets: Use Cases, Risks and Challenges," has followed the OECD definition. This can be explained by the international nature of IOSCO, whose members are not limited to EU authorities.
Following the OECD rationale, the definition of an AI system should be based on a lifecycle-based perspective encompassing two main phases, namely the pre-deployment (building phase) and the post-deployment (use phase) of the system. According to the Guidelines, the seven elements set out in this definition need not necessarily be present always during the two phases of this life cycle. The definition recognises that while certain elements may be present in one phase, they may not be present in the second phase. This dynamic approach to define an AI system reflects the complexity and diversity of AI systems. It also emphasizes an outcome-based approach. If a tool in the development stage is not considered AI, it can still develop AI features at a later stage und thus, be subject to regulation.
Clarifications of certain elements
The Guidelines are useful to clarify certain concepts and how they impact the definition of AI. In particular, they make a clear distinction between (i) a system's autonomy and (ii) its adaptiveness. While both are closely related, they remain distinct. They represent different dimensions of how an AI system operates. The use of the term "may" in relation to the "adaptiveness" shows that, in order to constitute AI, a system may, but does not necessarily have to possess adaptiveness or self-learning capabilities after deployment.. Accordingly, a system's ability to automatically learn, discover new patterns, or identify relationships in the data beyond what it was initially trained on is a facultative, not a mandatory criterion for determining an AI system, contrary to the element of autonomy, which is necessary to qualify as AI.
In the same vein, the "inference capability" seems another key feature that allows to distinguish AI systems from other (less advanced) software systems. As per the Guidelines, the EU AI Act should therefore not apply to systems that are only based on rules pre-defined solely by persons, i.e. to automatically execute orders.
Switzerland outlook
As mentioned, to date, there is no specific AI-legislation in Switzerland and therefore no legal definition. Instead, Swiss authorities, in particular the Swiss Financial Market Supervisory Authority (FINMA) with respect to financial market laws follow a more technology-neutral approach adopted. This approach allows the supervisory authority to ensure that financial market laws and regulations can flexibly cope with new technological developments and ensure implementation of adequate governance, compliance and risk management processes within supervised entities.
This principle has been restated by FINMA in its Guidance 08/24 on FINMA's Guidance on governance and risk management on the use of artificial intelligence published in December 2024. This Guidance also makes an explicit reference to the OECD definition and principles, FINMA reiterating that it expects the supervised institutions to have a sufficiently broad understanding of AI in line with international requirements to cover all relevant risk and ensure proper classification of AI applications. FINMA also reminds supervised entities that the same risks should generally be addressed in the same way.
The absence of AI-related legislation in Switzerland is not a blank check given to market participants, in particular not with respect to financial markets to integrate, use or advertise AI-based tools and products (to prevent "AI-washing"). Switzerland will closely monitor international developments to determine the potential need to adopt a dedicated legislative framework. The first outlines were already drawn with the ratification of the Europe's Landmark AI Framework Convention on 12 February 2025.
