alignment

Alignment is the broad name for making a model behave the way its makers intend. A raw model trained on text will happily continue any sentence you give it, including unpleasant ones. Alignment is the extra training on top that teaches it to be helpful, to decline dangerous requests, to admit uncertainty, and to avoid stereotyping. Refusal training, where a model learns to say no to certain requests, is one part of it.

The awkward part is that alignment is a blunt instrument. Teaching a model to avoid one failure often creates another: a study in 2026 found that models trained to avoid gender stereotyping ended up erasing female characters almost entirely, defaulting to “it” instead. Alignment is also not the same as security, since jailbreak prompts can bypass it, and it disappears completely if someone downloads a model’s weights and retrains them.