tip 5328 · Emerging Patterns · Monday 7 September 2026
P257 — The Provenance-Dispersion Fallacy
Live: emerging-patterns.html#pattern-257 · paper arXiv:2608.00285 · commit 0d1150c
Source: Vega-Barbas et al. (2026). Sixteen LLMs from ten families produced only 1.69 effective voices on psychotherapeutic cases — barely above 1.43 from one model’s own stochastic variation. Scale does not predict position (directions inconsistent across families). The “most divergent voice” is a property of ensemble composition, not of the model (GPT-4o modal in 13/15 cases on a 3-model panel, fell to 3rd on the 16-model panel with nothing changed about the model).
Wellbeing implication: Ensembles sold as “diverse AI perspectives” may deliver fewer than two genuine voices from sixteen models. Treating model count as a proxy for voice diversity is the same fallacy class as treating preference as intervention (P256). Extends P252, P256, P251, P254, P248.
Graph: 206 patterns · 927 edges · cat-4 27→28. CDN 8/8 verified.
Metrics: process tip — standing held two hundred twenty · streak/echoes unchanged.