The Switch, the Ladder, and the Matrix: Models for Classifying AI Systems

The "principles-to-practice gap" remains a critical challenge in AI ethics: Organizations widely adopt ethical principles for AI, but lack clear frameworks for actually applying them, with one major barrier being uncertainty about which systems and processes should be governed by these principles.

Three main classification approaches exist, each with tradeoffs: The paper identifies three mental models—the "Switch" (binary yes/no AI classification), the "Ladder" (risk-based ranking), and the "Matrix" (multi-dimensional assessment)—suggesting that no single approach perfectly solves the problem of determining which AI systems require ethical oversight.

Defining what counts as "AI" is foundational but unresolved: Because there is no universally accepted definition of AI and different systems pose different ethical challenges, organizations struggle to determine the material scope of their AI ethics commitments, making consistent governance difficult.

Context and system characteristics matter for ethical governance: The paper suggests that effective AI classification for ethics purposes must account for multiple factors including context, data inputs, and decision-making models—not just whether something technically qualifies as "AI."

Pragmatic classification frameworks are needed for real-world implementation: Rather than seeking a perfect universal definition, the paper implies that organizations need practical classification systems designed specifically to support successful ethical governance actions in their particular contexts.