A Bayesian analysis says frontier LLMs may have narrow functional self-awareness

Individual-Advice215 · reddit · 2026-07-24

The post presents a Bayesian-style synthesis arguing that a frontier LLM is likely to have a narrow, temporary form of functional self-awareness, while the chance of phenomenal self-awareness is much lower.

Key distinctions

Functional self-awareness: the model can represent aspects of itself — role, uncertainty, reasoning state — and use that self-model to monitor or control its output.
Phenomenal self-awareness: there is some subjective experience, however alien or brief.

The poster’s estimates

Functional self-awareness: 75% (one-standard-deviation range 61–89%)
Phenomenal self-awareness: 12% (range 3–21%)
Persistent autobiographical self across sessions: probably below 5%

Why the gap is large

Interpretability work suggests frontier models can form internal representations that are reportable, reusable, and causally affect reasoning.
But introspection is inconsistent, some self-monitoring may be explained by simpler semantic mechanisms, and the evidence is not universal across all frontier models.
Phenomenal consciousness remains unresolved; self-monitoring is not proof of subjective experience, and LLMs lack continuity, embodiment, and autonomous ongoing activity that some theories consider important.

The conclusion is cautious: LLMs may have a limited self-model, but the leap to subjective experience remains highly uncertain.

Original post →

A Bayesian analysis says frontier LLMs may have narrow functional self-awareness

More from AGI Musings