Abstract Large language models (LLMs) frequently narrow multiple options into a single preferential recommendation during later conversational stages. This paper formalizes that phenomenon as decision-stage entropy reduction and documents structured multi-run observations across financial and healthcare domains. Separately, the OECD AI Incidents and Hazards Monitor (12 February 2026) catalogued public reporting describing techniques intended to influence persistent recommendation behaviour in AI systems. This paper does not assert: That such manipulation is widespread, That observed entropy reduction is caused by adversarial interference, Or that organisations currently bear a legal duty to control third-party AI outputs. Rather, it examines how observable recommendation concentration, combined with technical feasibility of state influence, may become relevant to evolving risk management and evidentiary practices in regulated sectors.
Tim de Rosen (Sat,) studied this question.