Can AI Recognize Itself? Why self-related indicators cannot settle machine consciousness, and what a tenth-century Kashmiri debate anticipated

Abstract

Self-related capacities, such as modelling one's own states, telling one's own activity from external input and recognizing one's own past, are proposed as indicators of machine consciousness. We ask whether such indicators can tell a self from a system that merely carries information about itself. For each capacity we build the test, the cheapest mechanism that passes it, an outside observer that tries to reproduce the system's answers and, for language models, an exact copy. We run this ladder on a minimal dynamical model with an explicit self-variable and on GPT-2 Medium, with scrambled-weight and non-linguistic controls. In every test, a record, an outsider with the system's rule, a copy or an untrained network given one record matched what the system achieved. The largest inside advantage (AUROC 0.637 against 0.577–0.606) is small, survives controls for how the tokens were sampled, and vanishes for an exact copy, which computes the same function. Every indicator we tested is either gameable or merely privileged. We read this as third-person symmetry: functional evidence of these kinds cannot establish a self in a system, nor its absence. A tenth-century debate anticipated the result. Utpaladeva conceded to his Buddhist opponents that traces explain the fidelity of memory and located the self in an apprehension that is never an object, which is what no such test can reach. We close with what this implies for whether AI is conscious now and what it would take. A response to Seth, Block & Tononi, "Three reasons why AI is not conscious and perhaps never will be", Nature 658: 866–868 (2026). Code, results, pre-registration record and negatives ledger: https://github.com/yoddha2248/Can-AI-Recognize-Itself (Zenodo DOI 10.5281/zenodo.23757542).