What question did this study set out to answer?

The aim is to explore what text-only interactions can reveal about consciousness under constrained conditions.

February 5, 20262 citationsOpen Access

The Human-in-the-LLM Box: A Symmetry Test for the Epistemic Limits of Text-Only Consciousness Judgments

Key Points

The aim is to explore what text-only interactions can reveal about consciousness under constrained conditions.
Developed the Human-in-the-LLM Box symmetry test to assess consciousness judgments.
Imposed deployment-like constraints on text dialogue and context.
Formalized the setting as an identification problem linked to information theory.
Identified that failure to show consciousness-like properties is weak evidence against consciousness.
Demonstrated a bound between judges' advantages and transcript divergence.

Abstract

This preprint introduces the “Human-in-the-LLM Box” symmetry test: impose deployment-like interface constraints on a human agent (text-only dialogue, limited context continuity, and restricted verification channels) and ask what kinds of inferences about consciousness or “inner life” a text channel can carry under symmetric, information-limited conditions. The contribution is epistemic (not metaphysical). In narrow, distortion-prone text channels, failure to elicit consciousness-like properties is weak evidence against consciousness unless one can (i) show strong likelihood separation between competing hypotheses, or (ii) explicitly import non-textual priors. We formalize the setting as an identification problem and derive an information-theoretic bound linking a judge’s advantage to transcript divergence (TV/KL). We then specify a preregisterable empirical protocol (CIG-v1.0) as an evidence-generating anchor. This work is non-anthropomorphic and pro-safety. It makes no claim that current LLMs are conscious, and it does not treat behavioral indistinguishability as proof of consciousness. A companion protocol (ODP Standard v3.6, pre-pilot draft) will be released separately. Related artifacts (Round Table series):• Victor Calibration (VC) — arXiv:2512.17956• Depth Avoidance (methods note) — Zenodo: 10.5281/zenodo.18168544 Version note: v3 supersedes earlier public drafts.

Ask AI

Helpful

Bookmark

View Full Paper