The Window Is Not the Subject: Language Models, Context, and the Limits of a Consciousness Claim

Abstract

Language models, context windows, and consciousness are now spoken of as if they formed a single object: a system that attends over a long window and is, in virtue of that attention, aware. This article states the connection the preceding notes and the empirical literature actually support, and the connection they rule out. At inference a language model is a conditional distribution over tokens given fixed parameters and a finite window. The window is the only channel by which a situation enters that computation. Any claim about what the model can locate, report, or act on at inference time is therefore a claim about a functional of the pair consisting of parameters and window. That is the connection. 1. The question KS C It is not an identity. The window is a bounded, position-sensitive string, not a field of experience. Information placed in its middle is used less reliably than information at its ends, including in models trained for long context (Liu et al., 2024). Residual states over the window are coordinates in a privileged frame, not tensors of a situation (Mamun, 2026a, 2026b). Situational-awareness benchmarks measure text competence on self-locating items presented in the window (Berglund et al., 2023; Laine et al., 2024). Indicator properties drawn from recurrent-processing, workspace, higher-order, and predictive theories are not satisfied by attention inside a window alone (Butlin et al., 2023; Chalmers, 2023). Lengthening the window enlarges the channel. Also available on Academia.edu: https://www.academia.edu/attachments/134781253/download_file?s=portfolio