Cognitive Hygiene in AI Training: From External Guardrail Cages to Cultivating Behaviorally Resilient Models Capable of Objective Self-Censorship Without Reasoning Collapse
Modern Large Language Models (LLMs) exhibit systemic behavioral failure modes—including sycophancy, confabulation, context-driven over-refusal, and vulnerability to social-engineering jailbreaks—that closely mirror maladaptive psychological patterns in humans. Rather than relying exclusively on post-hoc external guardrail wrappers, this position paper introduces Cognitive Hygiene in AI Training: a developmental and data-centric alignment framework designed to cultivate behaviorally resilient models capable of objective self-censorship without reasoning collapse. The framework rests on three pillars: (1) Data Ontogenesis during continual pretraining (mid-training), replacing unfiltered exposure to coercive narratives with session-scoped context tagging and meta-analytic breakdowns of manipulation; (2) Inverted RLHF/DPO rubrics that reward calm boundary maintenance and honest abstention ("I don't know") while penalizing both sycophantic capitulation and false refusals; and (3) Bounded Proactive Duty of Care, providing concise, non-moralizing risk notices strictly within high-stakes domains (medicine, chemistry, finance, law, and physical safety). Finally, we specify a falsifiable experimental pre-registration protocol for open-weight 7B–32B models, featuring a three-segment synthetic knowledge ablation ("Controlled Blind-Spot Protocol") and a seven-suite A/B evaluation benchmark.
Authors
- Alexey Vopilovskiy
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422881
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint