The taxonomy distinguishes the following project stages in which a toolkit may be applied.

Design phase

Some methods for ethical AI are applied before the machine learning pipeline is developed, for example approaches to translating AI principles into design requirements. This stage is called the design phase.


The typical machine learning pipeline starts with a preprocessing stage. Within the scope of this project, this stage also includes data exploration.


After preprocessing comes learning, evaluation and prediction. In a typical training loop you also already do prediction and evaluation. To keep things simple, all steps of a typical training loop are summarized as in-processing.


Additionally, some tools are applied after a model is trained and predictions are made, which we can call post-processing or post-hoc. Post-hoc explainability methods are a good example of this. Model evaluation after the training loop is also categorized as post-processing.