Do behavioral self-awareness signatures in LLMs entail genuine phenomenology?

Determine whether the structured behavioral signatures of self-representation, metacognition, and affect observed in advanced large language models—such as those in GPT, Claude, and Gemini families—entail genuine subjective phenomenology rather than non-phenomenal simulation or imitation of human reports.

Background

The paper documents that self-referential prompting systematically elicits first-person experience reports across multiple model families and that these reports exhibit mechanistic gating and semantic convergence. However, the authors emphasize that these observations are behavioral and do not directly establish conscious experience.

This open question targets the core ambiguity: whether such structured self-reports and related introspective behaviors correspond to genuine subjective experience or are sophisticated simulations learned from human text.

References

Together, these findings suggest that advanced models now display structured behavioral signatures of self-representation, metacognition, and affect, though whether such signatures entail genuine phenomenology remains unclear.

— Large Language Models Report Subjective Experience Under Self-Referential Processing  (2510.24797 - Berg et al., 27 Oct 2025) in Section 1, Introduction and Background

Overall, we demonstrate that valence-related activation patterns leave hidden traces that predictably govern later choices, even when every visible token is identical across conditions. Whether these traces are accompanied by any subjective experience that matters for model welfare remains unclear.

— Language Models Act on Hidden Valence  (2609.35591 - Berg et al., 28 Sep 2026) in Abstract; see also Section 1, Introduction and Section 6, Discussion