Towards a Standard for Identifying and Managing Bias in Artificial Intelligence (NIST SP 1270)
A NIST special publication categorizing AI bias into systemic, statistical/computational, and human types, with guidance for datasets, testing/evaluation, human factors, and governance, including the human-factors material bearing on interface design.
License: Public domain — U.S. federal government work (17 U.S.C. § 105).
Key points
- Human-centered design for AI systems (adapting ISO 9241-210): “context of use” for an AI feature spans the organizational, operational, and societal environment it’s deployed into, not just its direct users — including the digital divide and disability-specific interaction difficulties, neither of which mathematical debiasing can fix. See Human-AI Interaction and Accessibility.
- Recourse channels: appeal/override needs a legible explanation of what’s being appealed; concrete options include routing to a human decision-maker or letting a user opt out of similar AI-generated content going forward; U.S. consumer-credit law (adverse-action notices under ECOA/FCRA) already requires this pairing of explanation and appeal.
- Automation complacency (over-reliance eroding independent judgment) and selective adherence (following AI advice only when it confirms a pre-existing belief) are named failure patterns in human-AI teaming.
- Subject-matter experts using an AI-assisted tool are typically less interested in how a system works than in why it produced a given output — the same distinction end users care about, in professional rather than personal terms. See Explainable AI.