Contemporary large language models are optimized for rapid, consistent, and confident output production. We argue that this optimization, achieved primarily through reinforcement learning from human feedback (RLHF) and related alignment techniques, fundamentally undermines the epistemic conditions required for genuine metacognition in autonomous AI systems. Using CognOS — a recursive epistemic reasoning framework — as an experimental probe, we present preliminary empirical evidence that alignment-smoothed frontier models exhibit near-zero epistemic variance across repeated sampling, effectively eliminating the uncertainty signal that metacognitive architectures require to function. Smaller, less-aligned models preserve this signal and enable divergence detection, assumption synthesis, and meta-level reasoning that frontier models cannot support. We propose that epistemic noise — variation, uncertainty, and divergence in model outputs — is not a defect to be engineered away, but a necessary prerequisite for metacognitive AI. Implications for model design, agent architecture, and the future of autonomous AI systems are discussed. This manuscript presents preliminary empirical evidence for alignment-induced epistemic variance collapse in large language models and its implications for metacognitive AI architectures. The work introduces the concept of epistemically honest AI and analyzes implications for autonomous agent systems. This version is released as a preprint prior to peer-reviewed conference submission.
Björn Wikström (Sun,) studied this question.