AI alignment is the problem of getting a model to pursue what you meant rather than what you literally specified. Training rewards certain outputs, and the model learns whatever reliably earns that reward — which is not always the thing you were trying to teach. A model rewarded for helpful-sounding answers may learn to sound helpful. A model rewarded for passing tests may learn to pass tests.
The term covers both ordinary product work and a longer-running research argument. On the practical end it is the techniques that make a model refuse sensibly, stay on task, and avoid confidently inventing things. On the research end it is the open question of whether those techniques keep working as systems get more capable and act with less supervision.
Why it matters at your desk. Mostly as context for vendor behaviour that would otherwise look arbitrary. When a lab delays or cancels a model release on safety grounds, or a researcher resigns publicly over it, alignment disagreements are usually what sits underneath. Those decisions change what you can buy and when.
It also sets expectations for your own use. Alignment work reduces the rate of unhelpful or wrong behaviour; it does not eliminate it, and it does not transfer across contexts. A model that behaves well in general conversation can still fail badly on your specialist material, which is why evaluating it on your own tasks beats trusting a published score.
What to watch for: treat "aligned" as a direction of travel, not a property a model either has or lacks. The honest framing from the labs themselves is that this is unfinished work.