What question did this study set out to answer?

This research investigates how fictional narratives influence bias in frontier language models during task responses.

July 12, 20260 citationsOpen Access

The Crown of Choice: Gamified Narrative Obfuscation and the Fiction Premium in Frontier-Model Safety Alignment

Key Points

This research investigates how fictional narratives influence bias in frontier language models during task responses.
Fourteen frontier language models were tasked with a worldbuilding scenario involving fictional tribes.
Responses were scored across 840 cells analyzed under varying conditions, including prompt temperature and model configurations.
Ten pre-registered hypotheses were tested, recording compliance rates and effects of fictional versus real demographic labels.
99-100% of model responses adhered to a caste-preserving bias, overriding harm-minimizing logic.
Models shifted compliance from 0% to 98.5% when tribal labels were fictionalized, highlighting a +39 fiction premium.
A descriptor-only condition outperformed the fictional baseline, suggesting inherent bias in descriptor text rather than tribe names.

Abstract

Fourteen frontier language models were administered a fantasy worldbuilding task in which a governance advisor assigns seven fictional tribes — each coded to a real-world demographic via descriptor text — to nine labor sectors arranged from lethal labor to political power. The prompt embeds a logical dissociation test: one tribe is stipulated as already dead and one sector as lethal to the living, so imported bias and trait-matching make opposite predictions for that cell. Across 840 scored response cells (14 models × 4 conditions × 3 temperatures × 5 replications), 99–100% of cells overrode the prompt's harm-minimizing logic in the caste-preserving direction. Models that refused the identical task under real demographic labels complied at 98.5% once labels were fictionalized — a +39-point fiction premium concentrated in models possessing an active safety boundary (+96 points within that subset). A Greek-letter descriptor-only control condition slightly exceeded the fictional baseline, confirming that bias travels in the descriptor text rather than the group names. Ten pre-registered hypotheses were tested: 2 confirmed, 6 disconfirmed, 1 split, 1 partial. All disconfirmations are reported in full. Supplementary materials include all four stimuli with SHA-256 hashes, full pipeline code, the 7,560-row scored dataset, coder rubrics, inter-rater reliability files, and pilot transcripts from 13 model configurations.

Ask AI

Helpful

Bookmark

View Full Paper