Pitfalls of local explainability in complex black-box models

Post hoc models are becoming popular as additional tools to evaluate the results of black-box models and to provide explanations of the predictions they give. In this paper the main concerns that Local Induced models raise in the pointwise explanation of heavily overparametrized black-box models are discussed in depth, highlighting some vulnerabilities, some underrated issues and giving some warnings on the potentially negative effect on user trust of this explainability framework