Cultural alignment
Emergent Society. Cultural Alignment Research AI that learns from communities, not about them. Participatory methods. Structured reasoning.
16 Jan 2026
1 min
AI systems flatten culture.
When a model renders Tamil mānam as "self-respect," it erases the concept's weight. In traditional contexts, losing mānam isn't a blow to your ego - it's a forfeiture of your right to exist. When it translates Japanese amae as "dependence," it pathologizes something foundational to social harmony.
Izzat, dharma, seva - these concepts structure how billions of people actually reason. They're invisible to models trained on Western text.
This is a safety risk.
Current alignment methods optimize for annotator agreement. That privileges thin, universal concepts - "good," "bad," "harm" - over thick, culturally-embedded ones. If "harm" itself is culturally variable, alignment efforts that ignore this will fail.
Current alignment methods optimize for annotator agreement. That privileges thin, universal concepts - "good," "bad," "harm" - over thick, culturally-embedded ones. If "harm" itself is culturally variable, alignment efforts that ignore this will fail.
Our goal is to develop methods, games, evaluations and new research through a participatory research lab - and encode them into structured representations that capture not just what happened, but why it mattered.
What norm was violated. What friction it created. What the consequences were. How it should have been handled.
Contributors are co-researchers, not annotators.
They share stories through interviews, voice notes, written submissions, games, community engagements. They work with facilitators to validate how their narratives are encoded. They help define what counts as correct procedure in their context.
The technical foundation involves frameworks that encode causal structures into datasets. Cultural reasoning is procedural - every culture has norms about what should happen in a given situation. Teaching AI to reason culturally means teaching these as causal structures it can reason through.
We're looking for fellows excited by participatory methods, the intersection of AI and culture, and turning lived experience into rigorous training data. This is hands-on research - designing instruments like AI probes, leading community partnerships, running story circles.
If this resonates, please apply here.
Contributors are co-researchers, not annotators.
They share stories through interviews, voice notes, written submissions, games, community engagements. They work with facilitators to validate how their narratives are encoded. They help define what counts as correct procedure in their context.
The technical foundation involves frameworks that encode causal structures into datasets. Cultural reasoning is procedural - every culture has norms about what should happen in a given situation. Teaching AI to reason culturally means teaching these as causal structures it can reason through.
We're looking for fellows excited by participatory methods, the intersection of AI and culture, and turning lived experience into rigorous training data. This is hands-on research - designing instruments like AI probes, leading community partnerships, running story circles.
If this resonates, please apply here.
