The effort to make a model behave as intended (helpful, honest, harmless) via RLHF and related methods; it sets a model's defaults about refusals, trust, and which sources it treats as credible.
Alignment is the effort to make an AI model behave as its developers and users intend: helpful, honest, and harmless, following instructions and human values rather than just predicting likely text. A raw pretrained model has language ability but no reliable sense of what it should or should not do; alignment supplies that.
The main technique is RLHF, training against human preference rankings, along with related methods that use AI-generated preferences or a written constitution of principles. Alignment decides a model's defaults: how it refuses, how it hedges uncertainty, which sources it treats as credible, and the biases it carries.
Those defaults bear directly on AI visibility. A model's learned dispositions about what counts as a trustworthy source, and which brands it recalls as reputable, are set during alignment, not only at retrieval time. Part of being recommended is fitting what an aligned model considers safe and credible to put in front of a user.
