Making AI systems pursue the goals their operators and society actually intend, and the open question of whether that's achievable.
The definition itself is disputed: what this term refers to depends on who is using it. Read any argument that leans on it twice.
At its narrowest, alignment is product work: tuning a model (largely via RLHF) to be helpful, follow instructions, and refuse harm. At its broadest, it is a research agenda about whether the goals of systems more capable than their evaluators can be specified, verified, or trusted at all.
The word is contested because both camps use it and mean different things. Practitioners disagree about whether alignment is an engineering problem being solved incrementally, a hard open research frontier, or unachievable in principle for sufficiently capable systems, and, beneath that, whose intentions the system is supposed to be aligned with. An argument that "alignment is going well" and one that "alignment is unsolved" may not be about the same thing.