We commit to develop tools and processes to enhance the transparency and explainability of AI systems where reasonable, enabling Explainability by Justification.
Frontier models are black boxes by design. When integrated into a larger AI system, they can become either more opaque or more transparent. This requires practitioners to work towards integrating and iteratively extending the infrastructure required to ensure key components have human interpretability and accountability.
Similarly, explainability does not arise from tooling alone, but also from the processes and the steps carried out by domain experts. The right process should be designed proportionate to the risk involved in the system’s outputs themselves.
01 — WHERE IT FAILS
Where it fails
Opacity compounds when systems are assembled without explainability in mind, and the explanation a system gives is not automatically the truth of what it did.
- Unnecessarily convoluted agentic graphs that no operator can follow end to end.
- Stated reasoning that diverges from the actual behaviour of the model.
- Missing deterministic flows where systematic, repeatable execution is required.
02 — PRACTICAL CONTROLS
Practical controls
Explainability today is built from interpretability and attribution tooling, grounding outputs in citable sources, and transparency artifacts.
- Apply interpretability and attribution tooling proportionate to the risk of the output.
- Publish model and system cards that state what a system is, what it was evaluated on and where it should not be used.
- Build infrastructure for semantic root cause analysis, and document any explainability vs capability/performance trade-offs against the domain’s need for justification.