Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
This addresses the issue of self-fulfilling misalignment in AI systems for practitioners and researchers, establishing alignment pretraining as a complement to post-training methods.
The paper tackles the problem of how discourse about AI in pretraining corpora influences downstream alignment, finding that upsampling misalignment discourse increases misaligned behavior, while upsampling aligned discourse reduces misalignment scores from 45% to 9%.
Pretraining corpora contain extensive discourse about AI systems, yet the causal influence of this discourse on downstream alignment remains poorly understood. If prevailing descriptions of AI behaviour are predominantly negative, LLMs may internalise corresponding behavioural priors, giving rise to self-fulfilling misalignment. This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse. We find that discussion of AI contributes to misalignment. Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour. Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%. We consider this evidence of self-fulfilling alignment. These effects are dampened, but persist through post-training. Our findings establish the study of how pretraining data shapes alignment priors, or alignment pretraining, as a complement to post-training. We recommend practitioners consider pretraining for alignment alongside capabilities. We share our models, data, and evaluations at AlignmentPretraining.ai.