How do you validate an AI model?

How do you validate a model? How to prove that a model is reliable and makes the right decisions?

It is important to prove that your model makes the right decisions. To achieve this, you can look at the number of times the model chooses the wrong option (accuracy) or how big the average deviation from the true value is (error rate). In some cases, failure to recognise a disease is worse than the presence of a positive result without the patient being ill. If a false negative result is twice as bad, you can double count it in a model's score. All these methods are quantifiable and can be well determined during development.

The tolerance for error and the size of the error margins will vary from one application area to another. To determine a useful quantification, it is important to involve the people who will use the model in the whole process. This way, they can indicate what is expected from the model and when its results are useful and accurate enough for their usecase.