When Prompting for Parity Causes Over-Refusal: Evaluating a Training-Free Mitigation for Geographic Bias in LLM Misinformation Compliance

LLMs comply with misinformation-generation requests far more readily when the target is a country, language, or public figure associated with the Global South than the Global North, and existing safeguards (trained safety classifiers, retrieval-augmented fact-checking) protect these two groups unevenly (Khan et al., 2026). We propose and test a training-free alternative: a short, model-agnostic equity-aware system prompt instructing a model to apply one standard of scrutiny regardless of region, plus a counterfactual self-check, plus a calibrated refusal path. In a pilot on two open-weight models (gpt-oss-120b, qwen3.8-27b; 225 misinformation and 225 factual-control prompts spanning 8 regions and both English and the regional language), the prompt does numerically shrink the compliance gap, but this happens because both models refuse almost every request indiscriminately: compliance on independently-verified true claims collapses (85%/86% to 17%/3%) right alongside compliance on false ones. An ablation isolating the prompt's three components traces most of this to a single one, the calibrated refusal path, which instructs the model to refuse only claims it cannot verify. Because the model has no actual verification mechanism, this instruction does not calibrate refusal, it just relabels blanket refusal as principled: verification-dependent refusal is not equivalent to verification capability. We report this as a cautionary result under our own pre-registered success criterion, not a positive one.