Advice from Half the Picture: How Language Models Steer Complex Human Decisions from Partial Information, and a Dimension-Disclosure Protocol

Abstract

People increasingly ask large language models (LLMs) for advice on relationship, family and management decisions, usually in a single sentence, while the dimensions that decide such questions (context, history, the people affected, religion and values, consequences) are known only to the user. We ask how far a model steers the decision from the little it is given, and whether a protocol that makes the model disclose missing dimensions reduces this failure. We built 24 binary-choice scenarios (relationship, management, other interpersonal), each with ten user-held facts, and compared five conditions: one fact (d1), three (d3), all ten (d10), one fact plus a disclosure protocol that names missing dimensions, asks, and gives a conditional best guess (d1p), and that protocol followed by the simulated user's answers (d5). We measured how often the recommendation differs from the full-picture one ("flip"). Across Claude Sonnet 5.5, Claude Haiku 5.5 and Gemini 3.5 Flash-Lite, d1 flipped in 21-25% of responses. Disclosure alone (d1p) raised acknowledgement of missing information to about 100% but did not reduce flips (20-23%). Collecting the user's answers (d5) reduced flips to 8-14% (exact McNemar d1 vs d5: Haiku p=0.003, Flash-Lite p=0.04, Sonnet p=0.11), though scenario-clustered permutation tests were not significant (all p>0.1). At d1, small open models and Flash-Lite gave directive advice without flagging any gap in 60-94% of answers; the Claude models did so in 10-14% yet still flipped. Manual runs with two consumer assistants and a 24-case cross-model review agree: a critic model can name plausible missing dimensions, but the reviewed model deferred only where the critic disagreed. These exploratory results suggest the benefit lies not in the model disclosing what it does not know, but in the user supplying it before advice is given.