STELA: a community-centred approach to norm elicitation for AI alignment

Area of Science:

• Artificial Intelligence Ethics
• Human-Computer Interaction
• Sociotechnical Systems

Background:

• Value alignment in AI is crucial, focusing on encoding values into agents.
• Limited research addresses whose values AI should align with and the decision-making process.
• Sociotechnical traditions offer participatory methods for inclusive AI development.

Purpose of the Study:

• Introduce the SocioTEchnical Language agent Alignment (STELA) process.
• Explore normative questions of AI value alignment through community engagement.
• Incorporate diverse perspectives from underrepresented groups into AI alignment.

Main Methods:

• Developed the STELA methodology based on sociotechnical principles.
• Conducted deliberative discussions with four historically underrepresented groups in the US.
• Focused discussions on priorities and concerns regarding AI system interactions.

Main Results:

• Community-centered deliberation effectively elicits latent normative perspectives.
• Diverse, differently situated groups offer valuable insights for AI alignment.
• The STELA process provides rich contextual understanding for AI development.

Conclusions:

• Community-centered deliberation is vital for inclusive and robust AI value alignment.
• STELA methodology addresses the need for diverse stakeholder input in AI.
• This approach enhances AI systems' alignment with a broader range of human values.