AI Model
An AI model is a rule-based algorithm or a mathematical algorithm with parameters that are trained on data. It is the AI within the system and part of the AI subsystem of model and harness. The AI system uses the model to derive from inputs how to generate outputs. What a trained model can do comes from its data. That data is therefore managed with the same care as the code itself (GAMP 5 Second Edition, Appendix D11). The EU AI Act defines the AI system (Art. 3 No. 1) and – for general-purpose models only – the GPAI model (Art. 3 No. 63); the AI model as such it does not define: it is a component, not a legal object in its own right. Because a model is in motion and subject to change (training, drift, retraining), it is verified against its specification – not qualified; validation takes place at the level of the overall computerized system in the process (QFINITY working convention). Once verified, the model is fit for purpose for the task its specification defines in the Context of Use; validation in the process shows whether the overall system is fit for intended use.


