As AI systems grow more advanced, their decision-making processes often become inscrutable "black boxes," making it difficult to ensure they behave as intended. While interpretability research aims to make AI more transparent, it remains unclear which approaches actually help align AI with human values. This creates a dilemma for researchers and funders trying to prioritize efforts in AI safety.

Mapping Interpretability to Alignment

One way to address this could be through a systematic analysis that evaluates different interpretability techniques based on their practical alignment value. The approach might involve:

• Cataloguing existing methods from mechanistic interpretability to concept-based explanations
• Developing criteria to assess each method's usefulness for detecting misalignment
• Interviewing alignment researchers about which techniques they find most actionable

The output could be a framework that helps distinguish between interpretability research that's academically interesting versus what actually moves the needle on safety.

Practical Applications for the AI Community

Such an analysis could serve multiple stakeholders in different ways. For research teams, it might highlight which interpretability approaches are worth incorporating into their development pipelines. Funders could use it to identify high-impact areas for grants, while policymakers might find guidance on which transparency methods warrant standardization.

The project could start with a lean version focusing just on literature review and preliminary interviews, producing a discussion paper that frames the key questions. This MVP could then evolve based on community feedback into a more comprehensive evaluation framework.