Causal Signatures of Awareness-Related Processing in Large Language Models

Debates about artificial awareness often move directly from surface behavior to claims about phenomenal consciousness. Recent mechanistic studies make a narrower question empirically tractable: do internal representations associated with report, affect, self-monitoring, persona, and flexible reasoning causally organize model behavior? We conducted a rapid systematic review of eight Anthropic-led evidence streams and identified two intervention families with sufficiently comparable outcomes for quantitative synthesis. Random-effects meta-analysis of three model experiments found that J-space argument swaps redirected flexible downstream computation on 45.9% of eligible trials (95% CI 37.4%-54.6%). Meta-analysis of five model experiments found that swaps targeting unspoken intermediate representations redirected two-hop answers on 58.3% of eligible trials (95% CI 48.2%-67.7%). Independent replication produced weaker intermediate-redirection effects than Anthropic and showed that direct answer-direction controls outperformed purported thought interventions, limiting the stronger interpretation. Functional-affect interventions produced large, directionally coherent behavioral changes, while introspective identification remained vulnerable to post-hoc selection and nonspecific steering confounds. The evidence supports structured, causally active awareness-related functions but does not establish phenomenal experience. Under the Taylor-Valmere Theory of Awareness, the results support a nonzero, uneven awareness profile while leaving continuity and subjective feeling substantially underdetermined.