Position: We Must Proactively Address AI Safety Debt

Peter Wallich, Raymond Douglas

Keywords: AI safety, technical debt, safety debt, risk management, responsible scaling, frontier AI, governance, agentic AI systems

TL;DR: We introduce AI safety debt as a framework for tracking how safety gaps accumulate in frontier AI systems, and propose a structured debt register for managing them.

Abstract: This is a position paper. We argue that AI safety debt — the cost of closing the accumulated gaps between an AI system's actual safety approach and the approach it needs — is accumulating rapidly in frontier AI systems. In the race to unlock near-term capabilities, practitioners often implement safety interventions that do not scale to more advanced, less transparent models. The concept extends the established software-engineering notion of technical debt, but four structural properties make AI safety debt harder to manage: capabilities and contexts shift unpredictably, closing gaps may require solving open scientific problems, harms largely fall on third parties, and adversaries and AI systems may actively exploit gaps. Our position is that the AI community must explicitly track and manage this debt rather than continually deferring it. We propose the AI safety debt register, a practical approach using structured "debt cards" that connect safety claims, supporting evidence, and organisational decisions. We argue that this framework complements existing governance approaches by providing bottom-up aggregation of safety gaps, proactive assessment of how evidence degrades over time, and an improved treatment of uncertainty.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 243