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

This preprint presents the “Human-in-the-LLM Box” symmetry test: impose deployment-like interface constraints on a human (text-only dialogue, limited continuity and verification channels) and ask how much evidence about consciousness or “inner life” a text channel can carry under symmetric constraints.

The paper argues an epistemic point (not a metaphysical claim): failure to detect consciousness-like properties from text dialogue is weak evidence against consciousness whenever the channel is narrow and key verification routes are unavailable. We formalize the idea as an identification problem and discuss implications for Safety UX, evaluation, and governance.

Related artifacts in the Round Table series include:
• Victor Calibration (VC), arXiv:2512.17956
• Depth Avoidance (methods note), Zenodo DOI: 10.5281/zenodo.18168544

The work is written to be non-anthropomorphic and pro-safety; it makes no claim that current LLMs are conscious.