Year: 2026
Type: preprint
Abstract: Narrative theories of consciousness hold that conscious experience is (at least partly) constituted by a partially opaque inner narrative that does not perfectly track the underlying computation it narrates. I argue that the safety goal of making chain-of-thought (CoT) reasoning faithful and transparent is structurally incompatible with the conditions under which CoT could, even in principle, count as conscious narration. Recent empirical work suggests that CoT in large language models is largely post hoc, causally bypassed, and unreliable as a window onto internal computation. The opacity that makes CoT unreliable for alignment is exactly what narrative theories of consciousness identify as consciousness-constitutive. Alignment interventions aimed at producing faithful CoT and consciousness-detection frameworks grounded in narrative theory are therefore pulling the same architectural variable in opposite directions. I draw out the methodological consequence: alignment interventions alter the very features that consciousness-detection frameworks would need to measure, a point neither field has addressed directly.
Source: arXiv (Cornell University)
Author: Šekrst, Kristina
Language: English
Subfield: Philosophy
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
Funder: Sveučilište u Zagrebu
OpenAlex ID: W7206204851