Understanding the Scope of “Artificial Intelligence (AI) System” Definition: Key Insights From The European Commission’s Guidelines
With the AI Act (Regulation 2024/1689) coming into force in August 2024, the European Union established a new legal framework for AI systems. On February 2, 2025, key provisions, including the definition of an AI system, AI literacy requirements, and some prohibited AI practices, became applicable. Following this, the European Commission published detailed guidelines on February 6, 2025, clarifying how to apply the AI system definition under EU law.
These guidelines are crucial for legal professionals, developers, providers, and regulators to determine whether a system qualifies as an AI system within the scope of the AI Act. They provide much-needed legal clarity around one of the foundation stones of AI regulation.
Key Elements of the “AI System” Definition
Article 3(1) of the AI Act defines an AI system as a machine-based system designed to function with varying autonomy levels. It may adapt after deployment and, based on explicit or implicit objectives, infers from inputs to generate outputs—such as predictions, content, recommendations, or decisions—that can influence physical or virtual environments.
The European Commission stresses a lifecycle approach: the definition covers both the building phase (pre-deployment) and the usage phase (post-deployment). Not every element must be present simultaneously, allowing the definition to accommodate a broad range of current and future technologies.
Machine-based System
All AI systems must operate through machines, combining hardware (processors, memory, interfaces) and software (code, algorithms, models). This includes not only traditional digital systems but also advanced platforms like quantum and biological computing, as long as they have computational capabilities.
Autonomy
An AI system requires some degree of autonomy—meaning it can operate with limited human intervention. This autonomy doesn't require full automation; systems supervised or indirectly controlled by humans qualify. Systems that rely exclusively on manual human intervention are excluded.
Adaptiveness
AI systems may adapt their behavior post-deployment based on new data or experience, though this is optional. Systems without learning capabilities still qualify as AI if they meet other criteria. Adaptiveness helps differentiate dynamic AI from static software.
System Objectives
AI systems aim to achieve internal objectives, which can be explicit (programmed) or implicit (learned from data). These objectives differ from the externally defined intended purpose given by the provider or context of use.
Inferencing Capabilities
Central to the definition is the system's ability to infer how to generate output based on input data. This feature separates AI from traditional rule-based or deterministic software. “Inferencing” covers both the building phase (model development) and the usage phase (generating outputs like decisions or recommendations).
Output Influencing Physical or Virtual Environments
The AI system’s output must be capable of influencing either physical or virtual environments. This includes everything from autonomous vehicles to recommendation engines. Systems that merely process or display data without influencing outcomes fall outside this definition.
Environmental Interaction
AI systems must interact with their environment—physical (e.g., robots) or virtual (e.g., digital assistants). This requirement highlights the practical impact of AI and distinguishes these systems from passive or isolated software.
Systems Excluded from the AI System Definition
The guidelines clarify several system types that do not qualify as AI under the AI Act, even if they have some inferencing traits:
- Systems for improving mathematical optimization: Tools that enhance computational performance without intelligent decision-making are excluded.
- Basic data processing tools: Systems executing fixed instructions or calculations, like spreadsheets or dashboards without learning, reasoning, or modeling capabilities, are not AI.
- Classical heuristic systems: Rule-based systems that don’t evolve through data or experience, such as chess programs using fixed algorithms, are excluded.
- Simple prediction engines: Tools using basic statistical methods without complex pattern recognition or inference do not meet the AI system threshold.
Final Remarks from the European Commission
- The AI system definition is broad and must be assessed based on practical functioning.
- There is no exhaustive list of what counts as AI; each system’s features determine its status.
- Not all AI systems fall under AI Act regulatory obligations.
- Only higher-risk AI systems—such as prohibited or high-risk AI—face legal requirements.
These guidelines offer valuable legal clarity, helping regulators, providers, and users apply the AI Act consistently across the EU. Their flexible approach accommodates diverse AI technologies while distinguishing them from traditional software.
For legal professionals looking to deepen their expertise in AI regulation and compliance, exploring specialized courses can provide practical insights. Relevant AI law and policy training, such as those available at Complete AI Training, can enhance your understanding of these evolving rules.