Four Principles of Explainable Artificial Intelligence (Draft NISTIR 8312)
A NIST draft report proposing four principles that define what it means for an AI system to be explainable, and comparing how well human decision-makers meet those same principles, alongside an explanation-type taxonomy.
License: Public domain — U.S. federal government work (17 U.S.C. § 105).
Key points
- Four principles: Explanation (a system supplies some evidence or reason for every output), Meaningful (the explanation is understandable to the specific person receiving it, and what counts as meaningful varies by user group and individual), Explanation Accuracy (the explanation must faithfully reflect the system’s actual process, not just sound plausible — distinct from whether the underlying decision itself was correct), Knowledge Limits (the system flags when a question is outside what it was designed for or its confidence is too low, rather than answering anyway).
- Five explanation categories by purpose: user benefit, societal acceptance, regulatory/compliance, system development, owner benefit — each implying a different explanation for the same output. See Explainable AI.
- A time-requirement-vs-detail trade-off: an emergency alert needs a brief, immediate explanation; a system audit or debugging session can absorb a long, detailed one.
- Tested against the same four principles, human decision-makers do poorly too — people fabricate plausible-sounding reasons for their own decisions after the fact (the “introspection illusion”) and are unreliable judges of their own knowledge limits (the Dunning-Kruger effect). See Explainable AI.