[Daniel Brooks]

A vendor may add generative AI to an existing product. What evidence should a buyer require before treating it as useful?

[alex mercer]

Accepted answer

Define the permitted use, affected decisions, users, data, and consequences of an error. Test representative cases for accuracy, harmful output, privacy, security, accessibility, human override, logging, and failure behavior. Also document the underlying model and change-notice process. Keep a non-AI path for critical work when the risk requires it. The NIST AI RMF is a voluntary framework for managing trustworthiness risks across AI design, use, and evaluation.

[Maya Chen]

Community reply

Should the buyer request the same test set again after the vendor changes the underlying model?

[Author]

Posts