Hello all. I’d like to share a new paper, PRISM: Perspective Reasoning for Integrated Synthesis and Mediation. The central idea is to tackle AI alignment challenges by systematically representing and reconciling the range of human moral concerns rather than collapsing them into a single metric. PRISM draws on moral psychology, cognitive science, and neuroscience to propose a set of seven non-arbitrary “basis worldviews,” each hypothesized to capture a distinct dimension of human moral reasoning. The framework then uses a Pareto-inspired optimization scheme to balance these viewpoints, attempting to avoid the usual pitfalls of single-objective approaches.

Here is a concise overview:

Seven Basis Worldviews
PRISM identifies seven vantage points (e.g., survival-focused, emotional, social, rational, pluralistic, narrative-integrated, nondual), each reflecting a unique layer of human moral cognition. The aim is to ensure coverage of the broad spectrum of values that people bring to ethical or high-stakes dilemmas.
Multi-Perspective Synthesis
Rather than simply averaging these perspectives, the framework applies Pareto-based balancing to avoid favoring one perspective at the unfair expense of others. A transparent conflict-mediation step highlights tradeoffs when the worldviews disagree on a solution.
Bounding Interpretive Leaps
By grounding itself in recognized human vantage points, PRISM aspires to mitigate machine-centric or purely instrumental interpretations that might disregard critical moral dimensions. The presence of diverse viewpoints is intended to reduce specification gaming and other oversights.
Illustrations and Classic Alignment Pitfalls
The paper demonstrates how PRISM can be applied to alignment scenarios (e.g., public health policy, workplace automation), systematically handling value pluralism, underspecification, and other classic pitfalls. Each perspective’s assumptions and reasoning are explicitly logged so final outputs include a clear record of tradeoffs.
Prototype & Interactive Demo
To illustrate PRISM’s operation in practice, I’ve built a hosted interactive demo. The system elicits outputs from each worldview for a user prompt, identifies conflicts, and then synthesizes a final mediated answer—documenting key assumptions at each stage.

Links:

Thanks for reading, and ...

Posts

Wikitag Contributions

Comments