Between Accurate Prediction and Justifiable Decision

A high predictive accuracy of artificial intelligence does not directly translate into justifiable decisions. In areas that affect life—such as lending, hiring, and healthcare—it is also important to consider which information has been used and how errors are corrected. Law and economics provides a starting point that simultaneously considers performance and rights protection, examining the costs and incentives associated with each choice.

Information That Varies Depending on the Explanation's Audience

Developers need to understand the causes of model failures, and users need to comprehend why outcomes unfavorable to them have occurred. Supervisory agencies must verify not only individual results but also whether the entire decision-making process is managed consistently. Therefore, it is challenging to assume that a single technical explanation can satisfy all purposes. Distinguishing the audience and purpose of the explanation first is necessary to appropriately determine the scope of information.

How Responsibility Alters Behavior

Who bears the cost of errors also affects prior investments. Imposing all responsibility on developers might overlook issues in the usage environment, while making only the user institution responsible may hamper control of internal model risks. It is important to examine concretely who can observe and mitigate which risks during data collection, model development, implementation, and operation. This is a research question comparing actual controllability rather than asserting specific responsibility allocations.

Linking Explanation to Procedure

Comprehensible explanations are a starting point. To make explanations helpful for actual exercise of rights, there must be avenues to correct faulty inputs, reconsider judgments, and allow human intervention. Research can observe not only the provision of explanations but also the time required for corrections, costs of objections, and outcomes of reconsiderations. Evaluating technical performance and procedural indicators together enables a more multidimensional assessment of technology's impact.

Questions to Consider Together

  • Who among users, supervisory agencies, and developers should receive explanations?
  • Do explanations actually lead to procedures where decisions can be contested?

Recommended Literature

This is a newly written educational commentary based on publicly available literature. It does not replace the original author's text and does not represent the official views of the Center or the professors.

View All Research Commentaries