A company is establishing design principles to achieve human-centered, explainable AI. Each department submitted a proposed policy. Which of the following proposals are appropriate as design principles? (Choose TWO.)
A question that asks which two are design principles of explainable AI.
Design with users and affected people at the center.
Correct. Human-centered design is the principle of designing AI with the needs and safety of users and affected people at the center. It leads to AI that people can understand and control.
Disclose the mechanism, capabilities, and limitations so stakeholders can understand them.
Correct. Ensuring transparency is the principle of making an AI's mechanism, capabilities, limitations, and data handling understandable to stakeholders. It supports explainable and trustworthy AI.
Provide explanations only as technical documents that only internal experts can understand.
Preparing technical documents seems reasonable at first, but the core of explainability is presenting it in a form that affected ordinary users can also understand.
Limiting it to a form only experts can read means users cannot judge, falling short of human-centered and transparency principles, so this is incorrect.
Evaluate quality with a single metric that maximizes model accuracy.
Accuracy is an important metric, but human-centered, explainable AI builds a balance of accuracy, fairness, and explainability into the design.
Making accuracy the only evaluation axis drops explainability and consideration for users, so it is incorrect as a design principle.
Automatically return a default answer to unexpected input with no human check.
Automating routine processing looks efficient, but human-centered design leaves room for people to oversee and intervene as needed.
A policy with no human check at all violates controllability and human-centered principles, so this is incorrect.
The design principles of human-centered, explainable AI are human-centered design (centering users and affected people) and ensuring transparency (disclosing the mechanism, capabilities, and limitations understandably).
・Policies such as explaining only to experts, using accuracy as the only metric, and leaving no room for human intervention sound reasonable but undermine users' understanding and control, so they violate the principles.