Position: An Inner Interpretability Framework for AI Inspired by Lessons from Cognitive Neuroscience

Inner Interpretability faces similar challenges to Cognitive Neuroscience: The paper identifies that critiques questioning the usefulness of inner interpretability research parallel longstanding debates in cognitive neuroscience about how to build meaningful mechanistic explanations, suggesting solutions from neuroscience could address current skepticism about AI interpretability work.

A conceptual framework can guide more productive research: The authors propose a general conceptual framework and concrete methodological strategies for building mechanistic explanations in AI systems, positioning this approach as a way to move inner interpretability research beyond current debates toward more useful outcomes.

Mechanistic explanations require careful methodology: Rather than pursuing interpretability research without clear direction, the paper emphasizes that developing rigorous mechanistic theories of AI systems—informed by how neuroscience has tackled similar problems—is essential for the field to demonstrate its value to responsible AI goals.

Cross-disciplinary lessons strengthen AI safety work: By drawing connections between cognitive neuroscience and AI interpretability, the paper demonstrates that responsible AI development benefits from learning how other fields have addressed the challenge of understanding complex systems, potentially accelerating progress in understanding and controlling AI behavior.