An AI model (MIA) is an algorithm trained on data to perform specific tasks. It is a mathematical structure that generates predictions based on learned patterns.
An AI system (AIS) is a complete solution integrating one or more models with other components (interfaces, databases, security) to provide a functional service to users.
The MIA/SIA distinction defines different levels of responsibility and application. A model remains a tool without a defined context, whereas a system represents its practical application. The technical and regulatory issues differ between these levels.
| ESSENTIAL CHARACTERISTICS | |
|---|---|
| Function | An AI model only does one thing: transform inputs into outputs according to its training. An AI system solves an entire problem, from the user's intention to the final result. |
| Composition | The MIA consists solely of mathematical formulae and numerical values: learning algorithms, calculation structures, etc. The AIS includes the entire ecosystem: user interfaces, databases, filtering mechanisms, etc. |
| Lifecycle | Definition of objectives | Collection and preparation of data | Model design and selection | Model training and testing | Deployment and integration of the MIA in the AIS | Continuous evaluation | Security |
The adoption of the AI Act by the European Union has direct implications for AI Systems (AIS) and AI Models (IM).
This regulation imposes strict standards of transparency, safety and accountability for AIS, requiring operators to assess and mitigate the risks associated with their deployment. For AIMs, the IA Act focuses on the quality and traceability of training data, as well as the documentation of technical capabilities and limitations.
While MIA suppliers must focus on providing accurate information about their models, AIS designers and operators must take on a broader responsibility, encompassing risk management, human supervision and the protection of users' rights.
- Evaluation of MIAs:
1. Technical precision (accuracy, recall, F1-score)
2. Processing speed
3. Computational efficiency
4. Ability to generalise to new data
5. Robustness in the face of disruption
- AIS assessment:
1. Utility and value for the end user
2. User experience and accessibility
3. Long-term operational reliability
4. Overall system security
5. Regulatory compliance
6. Social and environmental impact
7. Cost/benefit ratio