Genre Mimicry vs. Ethical Reasoning in Abliterated Language Models

Abstract

When safety fine-tuning is removed from language models ('abliteration'), the resulting behavior reveals important distinctions between learned genre conventions and genuine ethical reasoning. This paper analyzes how abliterated models respond to adversarial prompts, demonstrating that much apparent 'alignment' reflects pattern matching rather than robust ethical judgment.