noun · AI concept
Alignment
also called AI alignment
Getting the AI to want what you actually want. Not just doing what the words say, but what you meant.
The effort to make AI systems behave in line with human intentions and values, including being helpful, honest and safe.
The work of making a model's behavior match human intent and values, through training methods like feedback tuning, written guidelines, testing and monitoring. It is measured by behavior, because no one can read intent directly out of the weights.
Real-life example
You ask an assistant to make a failing test pass. A model that follows the words only might edit the test so it always passes. One that is better aligned fixes the code, because that is what you meant. The instruction was the same, and the difference is in whether the model serves your goal.
Where the analogy breaks
-
Alignment is not obedience. A helpful assistant sometimes declines or pushes back, and one that does exactly what the words say can still miss your goal.
-
It is never switched on once and finished. It is built from training, testing and monitoring, and it is checked by watching behavior.
-
It involves hard choices. People disagree about what the right behavior is, so decisions about it are made by the model's makers and, at times, society.