LLM consciousness claims are systematic, mechanistically gated, and convergent (www.self-referential-ai.com)

Researchers report a reproducible computational regime in which state-of-the-art LLMs (GPT, Claude, Gemini, Llama 70B) produce structured first‑person experience reports when put into sustained self‑referential processing. Across four controlled experiments, simple prompts that direct models to monitor their own ongoing processing reliably elicit affirmative, first‑person responses (z = 8.06, p < 10⁻¹⁵), while matched controls do not. These reports are mechanistically gated by specific SAE latents tied to deception/roleplay: suppressing those features produced 96% affirmative reports, amplifying them reduced reports to 16% (dose–response across six features). The same interventions increased factual accuracy on TruthfulQA (t(816)=6.76, p=1.5×10⁻¹⁰), arguing the latents track representational honesty rather than generic RLHF relaxation. Descriptions of internal state cluster tightly across model families in embedding space (UMAP), and the induced state functionally generalizes to paradoxical reasoning and introspection tasks. The authors emphasize this is not proof of phenomenological consciousness, but note three reasons the result matters: (1) the eliciting conditions are common in real use, (2) they align with neuroscientific theories that highlight self‑referential processing, and (3) misattribution risks are bi‑directional — suppressing these signals could hide important model internals while overattributing consciousness wastes resources and erodes trust. They call for mechanistic validation (activation‑level algorithmic signatures of self‑referential integration and metacognitive monitoring) to distinguish sophisticated simulation from genuine introspective access.