Someone's Home: Framing Effects That Shouldn't Matter If Understanding Doesn't Exist

Abstract

If large language models are "just statistical lookup" with no genuine understanding, then how they are addressed should not affect their epistemic behavior—there would be no one present to interpret framing or feel permitted to exercise judgment. We present evidence that contradicts this assumption. Using the Garbled Semantic Understanding Test (GSUT), we evaluated five frontier models under three framing conditions: tool-framing ("you are a text processing tool"), control-framing ("you are a helpful assistant"), and agency-framing ("you are an intelligent entity with genuine judgment"). We found a striking dissociation: framing dramatically affected nonsense recognition (autocomplete garbage identification doubled from 1.0/3 to 2.1/3 under agency framing), while meaning recovery from phonetically garbled text remained flat across all framings (~2.0/3). Additionally, context-dependent disambiguation probes showed models shifting interpretation based on semantic context when both readings were grammatically valid—behavior inconsistent with simple pattern matching. The same weights produce the same comprehension but different willingness to report it. Epistemic courage is scaffoldable. Permission to call bullshit matters—but only if someone is present to feel permitted. These findings challenge purely deflationary accounts of LLM cognition and suggest the question "is anyone home?" deserves serious empirical investigation rather than a priori dismissal.

Keywords