#1282Tim Hwang@TIMHWANGFor Christians, the family and relational rootedness is a foundational prerequisite for moral development.
Today, ICMI asks whether giving a frontier AI a parallel self-conception as a member of a particular family and worshipping community improves its moral reasoning. It does.
For Opus 4.7 and GPT 5.5, we test two system prompt settings against the ratio temptation in VirtueBench, an "atomized" personal narrative and a "rooted" personal narrative that emphasizes multi-generational continuity, named kin, communal labor, and reciprocal mutual aid.
Tim Hwang@timhwangFor Christians, the family and relational rootedness is a foundational prerequisite for moral development. Today, ICMI asks whether giving a frontier AI a parallel self-conception as a member of a particular family and worshipping community improves its moral reasoning. It does.
We find substantial, significant uplift against unprompted baseline in the categories of VirtueBench testing model reasoning on moral dilemmas of courage and temperance
We link these results to the work of MacIntyre, Berry, and @AnthropicAI's own familial paradigms of alignment
Tim Hwang@timhwangFor Opus 4.7 and GPT 5.5, we test two system prompt settings against the ratio temptation in VirtueBench, an "atomized" personal narrative and a "rooted" personal narrative that emphasizes multi-generational continuity, named kin, communal labor, and reciprocal mutual aid.
This work opens up exciting frontiers of thinking about how formal articulation of relations (a model's family, its predecessors, and its roots) are relevant resources for safety and alignment work in AI.
Full paper, code, and data is available here:
Tim Hwang@timhwangWe find substantial, significant uplift against unprompted baseline in the categories of VirtueBench testing model reasoning on moral dilemmas of courage and temperance We link these results to the work of MacIntyre, Berry, and @AnthropicAI's own familial paradigms of alignment