Scientific autonomy in the structural bubble: from institutional bias to AI-mediated consensus

(not merely as efficiency tools), to make this mechanism explicit in a minimal formal sketch, and to specify safeguards (validation loops, traceability, and oversight) that condition when AI strengthens vs. weakens scientific autonomy. Governance implications are derived from social epistemology and the sociology of quantification rather than presented as normative add-ons, and boundary conditions are sharpened with explicit falsifiers. Finally, the paper proposes a minimal empirical program to quantify agenda alignment, consensus lock-in, and amplification effects in AI-mediated discovery, summarization, and gatekeeping.