What are the ethical considerations in using surrogate models for explainability?
Instruction: Discuss the ethics of using simpler surrogate models to explain more complex AI models, including potential risks and benefits. Calculate overall and slice agreement for the supplied surrogate and decide whether overall agreement justifies using it to explain critical cases.
A surrogate agrees with a fixed target model on 891 of 900 routine examples and 60 of 100 critical examples. The two sets are disjoint and make up the full 1,000-example evaluation sample. Correct target predictions are not supplied.
Updated
Official answer available
Read the opening below, then unlock the full answer and practical guidance.
The way I'd explain it in an interview is this: The main ethical concern with surrogate models is that they can create an illusion of understanding...