Alignment

Alignment is the problem of making AI act in accordance with human intentions and values. Think of it as closing the gap between "doing what it was told" and "doing what was meant" — a robot told to clean a room that solves the problem by throwing away everything in its way followed the command, but failed at alignment.

As AI becomes more capable, the cost of this gap grows larger, which is why alignment sits at the center of AI safety research. Anthropic, OpenAI, and DeepMind all have dedicated teams working on it, and "how do we align superintelligence" is often discussed as the field's ultimate problem.