This version incorporates six targeted revisions made in light of Gurnee et al. (2026), "Verbalizable Representations Form a Global Workspace in Language Models" (Transformer Circuits Thread, Anthropic, July 6, 2026), published after the original study period (February 2026) but before this paper's first public upload. Specifically: Section 2.2 now cites direct empirical evidence for workspace-like internal structure in transformer-based language models, situating the paper's use of Global Workspace Theory within a mechanistic rather than purely theoretical context. Section 7.3 (Confabulation Objection) incorporates a methodologically independent line of evidence bearing on the central epistemic concern of this paper. This evidence is presented even-handedly: it includes findings that complicate a strong confabulation account of the present study's Test 2, as well as a finding (J-space ablation reducing experiential language regardless of referent) that moderates any strong reading of self-report as evidence of self-specific internal states. Section 9.1 notes a structural convergence between the paper's proposed Test 3 and Gurnee et al.'s counterfactual reflection training methodology. Section 9.3 updates the paper's proposed future research direction to reflect that mechanistic interpretability work of this kind is now underway. Section 9.4 adds a methodological limitation acknowledging that causal-intervention methods offer a stronger alternative to the behavioral protocol used here. The Abstract, Conclusion, and Author Note have been updated for consistency with the above, including a note clarifying the timeline of this revision relative to the original study. No experimental data, results, or claims from the original February 2026 study have been altered. All additions are clearly attributed and scoped to avoid overstating what the cited interpretability research does or does not establish regarding phenomenal consciousness — a question this paper, and the cited work, both explicitly leave open.
Roca et al. (Fri,) studied this question.