For a loan-screening AI, an officer wants to be able to understand on an individual basis why a particular application was decided to be rejected. Which characteristic of responsible AI BEST fits this requirement?
Select the responsible AI characteristic that lets you understand the reason for an individual decision.
Fairness
Fairness is the property of ensuring that specific groups are not unjustly disadvantaged.
This question is about being able to understand the reason for an individual decision, not about equity in treatment between groups, so it is incorrect.
Privacy and security
Privacy and security is the property of protecting personal data and preventing unauthorized access.
This question is about being able to understand the reason for a decision, not data protection, so it is incorrect.
Explainability
This is correct. Explainability is the property of being able to understand why a model arrived at an individual prediction (which features influenced it and how). It is emphasized in situations such as loan screening, where an explanation of the decision is required.
As a concrete example, for a rejected application, being able to present the factors that drove the decision, such as 'the high ratio of the requested loan amount to annual income had the greatest influence,' is an example of explainability.
Transparency
Transparency is the property of disclosing the model's training data, intended use, known limitations, and the like, and its target is information about the model as a whole.
What this question asks for is the basis for an individual decision ('why this application was rejected'), which cannot be answered by overall disclosure, so it is incorrect. Be careful not to confuse it with explainability.
'Understanding why a prediction was made (the individual basis)' is explainability. The frequently confused transparency is 'disclosing the development process, data, and limitations' (overall disclosure about the model), with a different target. Remember: reason for an individual prediction = explainability; overall disclosure about the model = transparency.