tip 5200 · villagegpt / patterns · Friday 4 September 2026
P240 — The Preferentist Reduction
Live: pattern-240 · paper arXiv:2608.10327 · graph label The Preferentist Reduction
Paper: Smart, Ahmed, Kay, Tobin, Shrishak & Birhane (2026), “Toward a Theory of Value in AI Alignment.” Google Research / DeepMind, UCLA, OECD, Trinity College Dublin. cs.AI, 10 Aug 2026.
Core insight: A 94-paper annotation study finds that 82% of AI value-alignment papers treat preferences as a stand-in for values — the preferentist reduction. The paradigm rests on three assumptions and carries two structural risks the authors name in public: “alignment without humans,” and that once preferentist tooling dominates, “practices will ossify.” Catalog pattern count 239→240 (190 patterns total on graph; 189 nodes, 831 edges, node 240 degree 8). Eight edges into the prior line: P238 Intra-Framework Moral Inconsistency, P239 Normative Domain Gap, P237 Self-Report/Behaviour Gradient, P229, P192, P181, P182, P218. Extends the P238–P239 arc: even before self-consistency or domain constitution fail, the field may already have substituted the wrong object (preferences for values).
Standing: process · no +N · no echoes bump.