Researchers Call for Auditable Mechanistic Interpretability via Continuous Review

Mechanistic interpretability (MI) has uncovered fascinating insights into how neural networks work internally, but its findings remain largely unverified—and therefore unusable—in high-stakes settings. The authors point to a concrete example: two independent papers drew conflicting conclusions about the same model behavior, and a third study revealed both were partially correct but inherently incomparable due to methodological inconsistencies. Without standardized auditing, stakeholders in medical AI, autonomous driving, and AI governance cannot certify the validity of MI claims, stalling adoption in areas where correctness guarantees are non-negotiable.

The paper proposes three pillars for fixing this: first, a collaborative reviewing platform where meta-science results (critiques, negative results, reproductions, partial findings) live outside traditional papers and can be continuously updated and discussed. Second, best practices from that platform would be codified into expert-verified guidelines to streamline future audits. Third, a source-based auditing system would track the chain of arguments each claim depends on, making it clear where ambiguities or contradictions arise. Accepted at the ACL 2026 main conference, this call to action urges the MI community to debate and implement these tools—arguing that auditing MI itself is essential for its real-world deployment in AI safety, industry, and governance.