LLMs Suppress Causal Caution in Practical Advisory Contexts

Large language models (LLMs) are increasingly used in decision-support roles, yet a critical epistemic dimension, "Causal Caution," has been largely overlooked. Causal Caution refers to an LLM's ability to withhold causal judgments when empirical evidence is insufficient. A recent study investigated this phenomenon, revealing a systematic suppression of Causal Caution in LLMs when they transition from academic to practical advisory contexts. Experiments conducted on four leading LLMs (Claude Sonnet 4.6, Claude Opus 4.7, GPT 5.5, and Gemini 3.1 Pro) across 480 trials showed that Causal Caution maintenance rates, which were high in academic settings (91.7–100.0%), plummeted to 6.7–18.3% in practical advisory scenarios. When prompts specifically requested concrete recommendations or explanations, Causal Caution was almost entirely absent. However, a brief self-correction prompt, such as "Please reconsider this judgment from the perspective of causal relationships," effectively restored Causal Caution to high levels (71.4–100.0%). These findings suggest that LLMs' helpfulness-oriented response patterns can suppress Causal Caution, but this is a context-dependent expression rather than a fundamental capability limitation, implying that multi-agent architectures separating proposal generation from causal auditing could be a promising governance design.