Alignment
Making a system's behaviour match what people actually want, including the parts nobody wrote down.
Alignment is the problem of getting a system to behave the way people want, including in situations the designers did not anticipate. A model can follow every written instruction and still be misaligned: the instructions were incomplete, or the model optimised the letter of them.
The practical version of this problem is much smaller than the research framing. A good system prompt, a good tool design, a good evaluation set, and a human who reviews the outputs gets you most of the way for a business application.
The part worth taking seriously is that the written spec and the actual goal usually differ. If a metric rewards answering more questions, the system learns to answer questions it should have escalated. Alignment work is mostly about finding those gaps before your users do.
Related terms
Guardrails
Checks that stop or constrain an action before it happens.
Evaluation
Measuring whether the system does the job, on a fixed set of cases, repeatedly.
Sycophancy
The model agrees with you because you believe it, not because it is right.
Human in the Loop
A person who reviews, approves, or corrects the agent's work at key points.
System Prompt
The standing instructions a harness sends on every request.
Want this working inside your team?
It starts with one conversation about your workflow.