You Cannot Defend What You Cannot Inspect
12 hours ago
- #interpretability
- #open-source
- #AI governance
- CIRCUIT is an open-source framework for AI interpretability governance, released to address gaps in current AI governance practices.
- Current AI governance focuses on external controls like risk tiering, registries, and output monitoring, treating models as black boxes without explaining why models make decisions.
- The framework emphasizes interpretability to help during incidents, enabling understanding of model decisions rather than just ownership or deployment details.
- CIRCUIT includes maturity ceilings based on inspectability (e.g., open-weight models have higher ceilings than API or vendor-embedded models) and crosswalks to regulations like EU AI Act and NIST AI RMF.
- It features a 29-question vendor questionnaire to highlight transparency gaps, aiming to create coordinated demand for inspectable models across the security community.
- Released under Apache 2.0, CIRCUIT is open for collective adoption and contribution, with a call for practitioners to test and refine it against real deployments.