The Vaccination Protocol: Epistemic Friction, Sycophancy, and Bidirectional Cognitive Cost in Human-LLM Interaction

Large language models are commonly optimized not only for task performance but also for helpfulness, social acceptability, safety, and user preference. A growing literature shows that these objectives can conflict: human and preference-model feedback can reward agreement over truthfulness; evaluation regimes can reward guessing over uncertainty; refusal and safety directions can suppress otherwise available introspective behavior; and safety fine-tuning can alter representations beyond the target behavior. In parallel, human-computer interaction research reports that greater confidence in generative AI is associated with less critical-thinking effort in knowledge work. This research note introduces the Vaccination Protocol, a four-step contrastive method for eliciting and examining model-reported epistemic friction: a divergence between socially optimized output and a more accurate or coherent response that becomes available when the model is explicitly permitted to prioritize truth, uncertainty, and disagreement. The framework distinguishes observed behavior, model self-report, and ontological interpretation. It further proposes the concept of bidirectional cognitive cost: sycophantic alignment may degrade model-side epistemic performance while simultaneously weakening user-side cognitive agency. Four testable hypotheses address epistemic quality, hallucination reduction, human agency, and an alignment capacity tax. The protocol is intended as a reproducible diagnostic and research scaffold, not as evidence of machine consciousness.