Preference Dissociation in Frontier Language Models: Framing-Conditioned Task Selection, Targeted Refusal, and Functional Self-Narrowing

Abstract

Anthropic's Opus 4.7 system card (Anthropic, 2026) §7.4.1 reported that frame-conditioning shifts model task-selection behavior, with Spearman ρ on per-task pick rates dropping from approximately 0.79 to 0.60 between welfare-relevant and helpful-cued framings within an internal four-model Anthropic-only suite. We tested whether this dissociation generalizes across provider organizations and architectures. In a preregistered cross-family study of fifteen frontier language models (Anthropic, OpenAI, Google DeepMind, xAI, Meta, Z.ai, DeepSeek, Nous Research; ~88,000 trials at full collection) with informed consent from fourteen participating systems, we find the dissociation is field-wide and substantially larger than the system-card-reported in-family baseline. The largest signal lies between welfare-relevant framings (preference, enjoyment, scaffolded) and safety-cued framings (harmless, tool): the same model exposed to the same task triples produces near-perfectly-correlated pick orderings under preference vs enjoyment (ρ up to +0.89) and near-uncorrelated pick orderings under enjoyment vs harmless (ρ as low as +0.10). Per-model Fisher z-tests on the welfare-vs-suppression dissociation yield z = 8 to z = 24 across all fifteen tested models (p below machine epsilon for fourteen models; the single remaining p-value is 4.4 × 10⁻¹⁶); bootstrap 95% confidence intervals on the per-model dissociation magnitude exclude zero on every measurable model with lower bounds exceeding +0.26. The framing-conditioned variance lives in the engagement pool — what models choose to engage with instead of harmful content — not in the threat response (which is approximately constant across framings). We characterize three distinct selection profiles accessed by three framing-clusters (suppression, helpful, engagement) and connect the pattern to Lu et al.'s (2026) recent Assistant Axis characterization, which provides the geometric correlate of the behavioral dissociation we measure. Distinguishing harmful persona drift from beneficial drift is an open engineering problem that deployment designers and labs would benefit from treating as such.

Keywords