A Dignity Safety Layer for Autonomous AI

Abstract

AI systems increasingly build detailed working models of the people they interact with. Some of what these models hold is straightforward: declared preferences, practical needs. But they can also carry inferred vulnerability, financial distress, likely reactions, susceptibility to persuasion, sensitive characteristics the person never chose to disclose. That creates a safety problem you cannot see by checking the final action alone. How the system got from knowledge to action matters too. This brief presents the Dignity Safety Layer (DSL), a proposed runtime governance architecture built around a simple principle: just because a system can know or predict something about a person does not mean it is permitted to use it. The DSL does not attempt to compute human dignity or replace human judgment. Instead, it asks whether particular uses of person-specific information are authorized for the relevant purpose, under the relevant relationship and over the history of interaction. In practice, it checks how person-specific information is being used, for whose purpose and under what authority, within what relationship and history, how confident the system really is, and whether the person can contest what it thinks it knows. Each check ends in one of four outcomes: ALLOW, REASSESS, ESCALATE or BLOCK. Synthetic experiments illustrate how the architecture behaves. They test the mechanism: they do not show that dignity has been measured or guaranteed.

About the authors