When Every Prompt Feels Personal: How the Barnum Effect Distorts Perception and Trust in Large Language Models

This paper examines how large language models can feel highly personal even when they have no user-specific information, due to the Barnum (Forer) effect—a cognitive bias that makes broad, mildly flattering statements seem uniquely accurate. It disentangles this illusion from genuine, data-grounded personalization and argues for an evaluation-first approach: (i) it explains how low-information prompts bias decoding toward hedge-heavy, mid-valence language that “fits anyone,” and (ii) it proposes replacing impressionistic prompts with precise, narrowly scoped, falsifiable queries tied to pre-specified items, making the presence (or absence) of relevant memory explicit and preventing Barnum effects by construction. The paper also outlines a memory-centric framework for more transparent personalization and reports that a precision-first evaluation protocol substantially reduces Barnum-style phrasing while improving the perceived relevance of responses.