Blueprint for an AI Bill of Rights
Five principles the White House Office of Science and Technology Policy proposed to guide the design, use, and deployment of automated systems: Safe and Effective Systems, Algorithmic Discrimination Protections, Data Privacy, Notice and Explanation, and Human Alternatives, Consideration, and Fallback. Each principle is followed by a “From Principles to Practice” technical companion spelling out concrete expectations. The Notice and Explanation and Human Alternatives, Consideration, and Fallback companions give direct user-facing interaction guidance — explaining automated decisions and providing a human fallback when one goes wrong.
License: Public domain — U.S. federal government work (17 U.S.C. § 105).
Key points
- Match how much an explanation discloses to the level of risk — and in high-risk settings, choose an inherently interpretable model rather than bolting an explanation onto a black box. Where consequences are high (criminal justice, other high-stakes public-sector settings), the system’s full behavior should be explainable in advance by design, not reconstructed as an after-the-decision interpretation. See Explainable AI.
- Tailor an explanation to its purpose as well as its audience: a purely informational explanation is a different artifact from one meant to support an appeal, a dispute, or a contestation process, and the explanation should say plainly which purpose it’s serving. See Explainable AI.
- An explanation should state its own error range where one can be calculated, balanced against how much that adds to interface complexity — accuracy about the explanation’s own uncertainty, not just about the decision. See Explainable AI.
- Notify people that they’re entitled to opt out of an automated system in favor of a human alternative — a specific, brief, accessible notice of the opt-out right itself, not just a fallback channel that exists but is never surfaced.
- Size the human-fallback system’s staffing and availability to the stakes of the automated system it backs up — a system with greater potential impact on someone’s rights or access needs proportionately more human reviewers available, not a fixed-size queue regardless of what’s riding on it.
- When a human reviewer overturns an automated decision, propagate the reversal through every system component and undo whatever downstream consequences the original decision already caused — reversing the decision without also reversing its already-triggered effects (a benefit clawback, a flagged account) leaves the override only nominally effective.
- Time-critical automated systems need immediate human fallback, sometimes available before the harm occurs — a building’s automated access-control system backed by a manager who can just open the door is the concrete example given; waiting for a review process defeats the purpose once the window to act has passed.
- Combat automation bias in the human reviewers themselves, not only in the AI system: recurring training on how to interpret the system’s output for its intended purpose, plus ongoing assessment that human involvement isn’t quietly undermining the system’s safety and fairness properties.