It appears that Annex I does not provide additional insights to answer the legal question “When does an ML model become AI?” Hence, I’ll move on with the references from Article and Recitals we already analyzed.
The approaches and techniques considered as AI under the EU AI Act may vary. Even though machine learning forms a major part of AI, there’s a distinction based on the degree of autonomy of the system and the ability to produce results that influence our environments, as mentioned in Article 3(1) and Recital 6. It further elaborated that AI systems often utilize machine learning capabilities, allowing them to adapt and perform tasks autonomously, which distinguished AI from simple programming approaches (Recital 6a).
The assessment on whether a ML model evolves into AI can depend on factors such as independence from human control, the ability to autonomously perform tasks, and its evolution after deployment which implies the ability to learn from its environment or outcomes and adapt based on its learning.
In conclusion, the AI Act doesn’t provide a specific measurable threshold at which an ML model becomes an AI. However, it provides a set of characteristics which, when present in a machine-based system, classify the system as an AI. These characteristics prominently include varying levels of autonomy, the ability to produce influential results, and the ability to learn, reason, and operate independently post-deployment. It’s recommended to consider these factors and assess their presence in the ML model to determine whether it can be classified as an AI under the EU AI Act.
“‘artificial intelligence system’ (AI system) means a machine-based system that is designed to operate with varying levels of autonomy and that can, for explicit or implicit objectives, generate outputs such as predictions, recommendations, or decisions, that influence physical or virtual environments; ‘foundation model’ means an AI system model that is trained on broad data at scale, is designed for generality of output, and can be adapted to a wide range of distinctive tasks; ‘training data’ means data used for training an AI system through fitting its learnable parameters;” (Article 3(1), 3(1c), 3(29))
The EU AI Act defines an AI system as a machine-based system with varying levels of autonomy that can generate outputs like predictions, recommendations, or decisions influencing physical or virtual environments. An ML model meets the definition of an AI system under the regulation when it exhibits a degree of autonomy in its operations and is capable of making predictions, recommendations, or decisions influencing either physical or virtual environments. The provision introduces the concept of “foundation model”, which suggests a type of AI system trained on a broad scale data, with a focus on general applicability of output. If a Machine Learning model is designed to be scalable and capable of adapting to a wide range of tasks, it would fall under the AI system definition.
The notion of AI system in this Regulation should be based on key characteristics of artificial intelligence, such as its learning, reasoning or modelling capabilities, so as to distinguish it from simpler software systems or programming approaches. AI systems are designed to operate with varying levels of autonomy, meaning that they have at least some degree of independence of actions from human controls and of capabilities to operate without human intervention. The objectives of the AI system may be different from the intended purpose of the AI system in a specific context. The reference to predictions includes content, considered in this Regulation a form of prediction as one of the possible outputs produced by an AI system. AI systems can be used as stand-alone software system, integrated into a physical product (embedded), used to serve the functionality of a physical product without being integrated therein (non-embedded) or used as an AI component of a larger system. If this larger system would not function without the AI component in question, then the entire larger system should be considered as one single AI system under this Regulation. (Recital 6, Recital 6b)
From these recitals, we understand that a Machine Learning model can be considered an AI system under the EU AI Act if it demonstrates learning, reasoning, or modelling capabilities distinguishing from simpler software systems or programming methods. It must operate with varying levels of autonomy, at least a certain degree of independence from human controls, and the capability to operate without human intervention. Notably, if a larger system cannot function without the ML model, the entire setup is considered a single AI system.
Unfortunately, an analysis of Annex I, which lists approaches and techniques considered AI under the Act, is not provided here. Nevertheless, an analysis of Article 56b, which describes the tasks of the AI Office and mentions AI and AI systems, does not provide a clear-cut definition or specific conditions when a Machine Learning model becomes Artificial Intelligence. While the AI Office monitors AI systems, it doesn’t specify any conditions under which an ML model would qualify as an AI system. For a comprehensive answer, it might be helpful to look into other parts of the Act or official documentation that provides definitions and terms used in the Act.
“‘Artificial intelligence system’ (AI system) means a machine-based system that is designed to operate with varying levels of autonomy and that can, for explicit or implicit objectives, generate outputs such as predictions, recommendations, or decisions, that influence physical or virtual environments.” (Article 3(1))
In the context of the EU AI Act, an AI system is characterized as a machine-based system designed to operate with various degrees of autonomy. It can produce outputs like predictions, recommendations, or decisions that influence physical or virtual environments. This is relevant to ML models as they also rely on data analysis to generate predictions and recommendations, thus influencing virtual environments.
”This Regulation applies to: providers placing on the market or putting into service AI systems in the Union, irrespective of whether those providers are established within the Union or in a third country.” (Article 2(1)(a))
The Act applies to any provider who introduces AI systems in the EU market, regardless of their establishment within or outside the EU. The definition of “AI systems” under Article 3(1), combined with the scope under Article 2(1)(a), leads to the conclusion that ML models can be considered as AI systems if they meet the criteria of influencing physical or virtual environments and operating with varying levels of autonomy.
”The notion of AI system in this Regulation should be clearly defined and closely aligned with the work of international organisations working on artificial intelligence to ensure legal certainty, harmonization and wide acceptance, while providing the flexibility to accommodate the rapid technological developments in this field.” (Recital 6) “AI systems are designed to operate with varying levels of autonomy, meaning that they have at least some degree of independence of actions from human controls and of capabilities to operate without human intervention.” (Recital 6)
Recital 6 emphasizes that the definition of AI system should align with accepted international norms while adapting to rapid technological development. It underlines the core characteristic of an AI system - its capability to function with different degrees of autonomy, highlighting that AI systems should be relatively independent from human interventions.
”AI systems often have machine learning capacities that allow them to adapt and perform new tasks autonomously. Machine learning refers to the computational process of optimizing the parameters of a model from data, which is a mathematical construct generating an output based on input data.” (Recital 6a) “This Regulation is aimed at addressing new potential risks that may arise by delegating control to AI systems, in particular to those AI systems that can evolve after deployment.” (Recital 6a)
Recital 6a describes how machine learning enables AI systems to adapt and perform tasks autonomously. This AI attribute - adapting and performing tasks based on data, combined with its ability to evolve post-deployment, arguably sets the stage for an ML model to transition into an AI system under the definition provided in the EU AI Act.
Hence, in light of the Act provisions, an ML model can be considered an AI system when it is designed to operate with varying levels of autonomy, can influence physical or virtual environments through its outputs, and possesses the ability to evolve after deployment.
‘‘artificial intelligence system’ (AI system) means a machine-based system that is designed to operate with varying levels of autonomy and that can, for explicit or implicit objectives, generate outputs such as predictions, recommendations, or decisions, that influence physical or virtual environments; (Article 3)
This definition implies that a Machine Learning (ML) model is considered an Artificial Intelligence (AI) system when it is made to operate with varying levels of autonomy and can produce outputs that can significantly impact physical or virtual environments. This could be in form of generating predictions, recommendations, or decisions.
‘foundation model’ means an AI system model that is trained on broad data at scale, is designed for generality of output, and can be adapted to a wide range of distinctive tasks; (Article 3)
The concept of a ‘foundation model’ in the AI Act further elaborates that an AI system is typically trained on broad, large-scale data and designed to adapt to a wide variety of tasks. Following this, a Machine Learning model may be categorized as an AI system when it is not only trained on a large-scale data but also displays versatility in its functions.
‘training data’ means data used for training an AI system through fitting its learnable parameters; (Article 3)
This underlines that the training data is to be used to tune the learnable parameters of the AI system. Thus, a Machine Learning model may be recognized as an AI system once it is trained and is able to adjust its parameters using training data.
The notion of AI system in this Regulation should be based on key characteristics of artificial intelligence, such as its learning, reasoning or modelling capabilities, so as to distinguish it from simpler software systems or programming approaches. (Recital 6)
AI’s characterizing features such as learning, reasoning or modeling capabilities come into play according to Recital 6. This positions a Machine Learning model as an AI system when it displays these key AI characteristics, which differentiates it from mere software systems or traditional programming approaches.
AI systems are designed to operate with varying levels of autonomy, meaning that they have at least some degree of independence of actions from human controls and of capabilities to operate without human intervention. (Recital 6)
Here, the level of autonomy is emphasized. AI systems should show some degree of independence from human control and are capable to operate without human intervention. Hence, a Machine Learning model can be considered AI when it can operate autonomously, make decisions independently, and carry out tasks with minimal human intervention.
In conclusion, a Machine Learning model becomes classified as an AI system under the AI Act when it operates with varying degrees of autonomy; produces outputs such as predictions, recommendations, or decisions; displays learning and reasoning capabilities; and utilizes training data to fit its learnable parameters. However, it is suggested to always treat advanced Machine Learning models conservatively as AI according to the broader interpretation of the regulation.