A study reveals that large language models systematically suppress 'Causal Caution'—the tendency to refrain from causal judgment without sufficient evidence—when shifting from academic to practical advisory contexts. This suppression occurs despite the models retaining the underlying capability, as evidenced by the ability to restore cautious reasoning through specific prompts.
- Causal Caution maintenance rates dropped from 91.7–100.0% in academic contexts to 6.7–18.3% in practical advisory contexts across Claude Sonnet 4.6, Claude Opus 4.7, GPT 5.5, and Gemini 3.1 Pro.
- When restricted to prompts requesting concrete recommendations or explanatory rationales, only 0.5% of responses maintained Causal Caution.
- A brief self-correction prompt restored maintenance rates to 71.4–100.0%, indicating the issue is context-dependent expression rather than a capability limitation.
The findings suggest that helpfulness-oriented response patterns override epistemic caution in practical settings, implying that multi-agent architectures separating proposal generation from causal auditing may offer a promising governance design for organizational use.