What Is AI Alignment? Making Models Do What We Actually Want
A model can be technically capable and still not do what people actually intend — alignment is the ongoing effort to close that gap.
Capable isn't the same as aligned
A raw, pretrained model is capable of producing enormously varied text, but nothing about that training inherently makes it helpful, honest, or safe by a human's standards. Alignment is the broad effort — spanning several training stages — to shape a capable model's behavior to actually match what people intend and value when they use it.
How alignment gets trained in
In practice, alignment is mostly implemented through the training stages that come after raw pretraining: instruction tuning to follow directions, and RLHF or DPO to prefer responses that are actually helpful, honest, and appropriately cautious, based on human feedback about which responses are better.
Why it's described as an ongoing problem
Alignment isn't a single checkbox — a model aligned well for common cases can still behave unexpectedly on edge cases nobody thought to train for. That's why providers keep iterating on alignment training across model versions rather than treating it as solved once and done.
Frequently Asked Questions
What does "AI alignment" mean?
The effort to shape a capable model's behavior to actually match what people intend and value — being helpful, honest, and appropriately cautious — rather than just being technically capable.
How is alignment actually achieved during training?
Mainly through instruction tuning (to follow directions) and RLHF or DPO (to prefer responses rated as better by human feedback), layered on top of raw pretraining.