AI alignment is the work of making a model's behavior match what its developers and users intend. In practice it has two parts, teaching the behavior in training and enforcing it at run time.
What it means
An aligned model does what it is meant to do, including where nobody wrote a rule: it follows instructions, declines what it should, says when it does not know, and uses tools only for the task.
Misalignment is the gap between the behavior you intended and the behavior you get: a support bot that gives legal advice, a coding agent that edits a failing test so that it passes.
The gap exists because behavior is learned from data, not written as code. Pre-training teaches a model to continue text. It says nothing about what your product should do.
How it is done
- Teaching. Post-training shapes the behavior: supervised fine-tuning on examples of what you want, preference tuning on better and worse answers, reinforcement fine-tuning with a reward for the right behavior.
- Enforcing. AI control checks the behavior at run time from outside the model, with policies on inputs, outputs and tool calls.
Training makes the intended behavior likely, not certain, and a jailbreak or an injected instruction can still push the model off it. Control covers what training misses.
With White Circle
Halo is our open-source framework for post-training open models. Control enforces policies written in plain language on every input, output and tool call.
