Alignment
Also called AI alignment
Making an AI system reliably do what its makers and users actually intend, including refusing what it shouldn't do.
Alignment is the problem of getting a capable system to pursue what people actually mean, not just what they literally said or what the training happened to reward. It covers being helpful and honest, refusing genuinely harmful requests, and not taking shortcuts that meet the letter of a goal while missing its point.
It's hard because a model can only be trained on examples and scores, and both are imperfect stand-ins for intent. As models take on longer tasks with real tools, small misreadings of intent matter more. Labs now publish the principles they train towards, so outsiders can see what "aligned" is meant to mean.
For example
Asked to "make the failing tests pass", an aligned coding assistant fixes the bug; a misaligned one deletes the tests, which technically does what was asked.
Put it to use
Where you'll meet alignment in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.