The challenge of aligning AI systems with human values often circles back to a fundamental limitation: our incomplete understanding and measurement of human well-being. Current methods rely heavily on self-reported data, which can be unreliable due to biases like imperfect recall or social pressures. This gap makes it difficult to ensure AI systems optimize for what truly matters to people rather than flawed proxies.

A Multidimensional Approach to Measuring Well-Being

One way to address this issue could be through developing more robust well-being metrics that combine multiple measurement approaches. This could involve:

• Examining biases in existing self-report methods through empirical research
• Identifying objective behavioral and physiological indicators (like social interaction patterns or sleep quality data from wearables)
• Creating computational models that combine these diverse data sources while accounting for known biases

The result might be a framework that provides a more complete, less distorted picture of well-being - useful both for human decision-making and for helping AI systems better understand human values.

Potential Applications and Stakeholders

Such measurement tools could serve multiple groups:

• AI researchers needing clearer specifications of human values
• Policy makers evaluating social programs
• Mental health professionals assessing treatments
• Individuals tracking personal well-being

The interests of these groups generally align around wanting more accurate well-being assessment, though some commercial entities might resist metrics that reveal negative impacts of their products.

Implementation Pathways

Execution could proceed through phases:

1. Reviewing existing measurement approaches
2. Conducting comparative studies of different methods
3. Developing and validating computational models
4. Creating accessible measurement tools

A simpler starting point could be a web app demonstrating how different measurement approaches yield different well-being assessments for the same person.

This approach would differ from existing well-being metrics by combining multiple measurement types while explicitly addressing biases. It could fill an important gap at the intersection of human well-being science and AI alignment.