New Framework Treats AI Safety as a Systems Problem, Not Component Silos

Most approaches to managing AI risks focus narrowly on individual parts—the training data, the model, the deployment environment—in isolation. A new framework argues this misses the real dangers: problems that only appear when components work together, or when the broader development and operational processes fail to catch them.

Moving Beyond Component-Level Analysis#

Researchers have drawn on decades of system safety expertise to build a more holistic approach. Rather than treating each layer of an AI system as a separate concern, they apply System Theoretic Process Analysis (STPA), a recognized safety framework from domains like aerospace and automotive engineering. The shift matters because hazards often emerge from interactions between supposedly safe parts—a phenomenon that becomes especially acute in machine learning systems where opacity, unpredictable scaling of capabilities, and complex outputs create new failure modes.

The PHASE Framework#

The team developed Process-oriented Hazard Analysis for AI Systems (PHASE) as an adaptation of STPA tailored to AI. They tested the approach on three distinct case studies: a linear regression model, reinforcement learning applications, and transformer-based generative models. The exercise revealed that while STPA's core logic holds for AI, the framework needs targeted tweaks to account for how ML systems behave differently from traditional engineered systems.

What the Framework Enables#

Applying PHASE offers four concrete advantages for teams managing AI harms:

• System-level hazard detection: identifying dangers that emerge from accumulated minor issues or component interactions, not just individual failures
• Social and process visibility: explicitly mapping how organizational and development practices contribute to algorithmic harms
• Accountability chains: tracing harms back to specific roles and decisions that could have prevented them
• Ongoing adaptation: building processes to spot and counter new hazards as systems evolve

The approach treats safety as something that must be woven into how teams design, build, and operate AI systems—not added as an afterthought at any single stage.