The article is interesting as it introduces a number of explainability techniques as currently used in the field.

These methods include:

• Feature Importance via Shapley Values
  
• Counterfactual explanations
  
• Adversial Training
  
• Influential samples
  

Besides the discussion of these techniques the authors highlight important concerns that should be kept in mind when using explainability.

1. Causality should not be the sole focus as this would exclude interesting data/events from consideration
2. Privacy and the possible loss of privacy needs to be a consideration while using these techniques.
3. The use of explainability can lead to better performance. The authors note that this performance can then be used in ways that the stakeholders might not necessarily agree with. Their example concerned better natural language processing in order to be able to better moderate social media posts.

And finally the authors provide the suggestion, based on their interviews, that designing explainability with and for (external) stakeholders will be an important feature if we want to take the explainability from an internal tool used by the data scientist and engineers to a tool providing explanations to external stakeholders. In order to achieve this they suggest to:

• Identify stakeholders;
• Investigate what explainability means for these stakeholders;
• Determine the effect of an explanation on a stakeholder and design your explanations with this in mind.