Analyzing AI for harms by examining its pieces in isolation misses the dangers that emerge from how those pieces interact.
A team of researchers has adapted System Theoretic Process Analysis (STPA) - a framework borrowed from industrial safety engineering - to evaluate AI systems as a whole rather than auditing training data, models, and deployment pipelines separately. They tested the approach, which they call Process-oriented Hazard Analysis for AI Systems (PHASE), across three case studies: a linear regression model, a reinforcement learning system, and a transformer-based generative model. The core idea is that safety is an emergent property of the full system, including the company processes surrounding it - not just a checklist property of each component. The researchers found that STPA's core concepts transfer to AI, but need targeted adjustments for problems specific to the field, such as model opacity, uncertain capability boundaries, and the sheer complexity of model outputs.
The practical payoff is significant. PHASE gives safety analysts four concrete tools: a way to spot hazards that only appear when disparate problems accumulate, a method for accounting for social factors that feed algorithmic harm, a traceable chain of accountability from a harm back to who can fix it, and a structure for ongoing monitoring rather than a one-time audit. That last point matters in a field where models evolve and deployment contexts shift constantly.
Most current AI safety reviews still treat the model as the unit of analysis - red-teaming outputs, auditing datasets, or stress-testing APIs. PHASE pushes the boundary out to the full development and operation process, which is closer to how aviation and nuclear industries think about safety. Whether AI labs will voluntarily adopt a framework that makes accountability chains explicit - and therefore visible to regulators - is a separate question.