Alignment
The technical and philosophical problem of ensuring AI systems reliably do what humans want them to do — including cases where 'what humans want' is complex or contested.
Alignment is the AI-safety subfield concerned with getting AI systems to behave in ways that match human intent. The problem sounds simple and is deeply not. Two subproblems: outer alignment (specifying what we want clearly enough that we can measure it) and inner alignment (ensuring the model that gets trained actually pursues what we specified, rather than something correlated with it that comes apart in edge cases).
In current practice, alignment work includes techniques like reinforcement learning from human feedback (RLHF), constitutional AI (using written principles to guide behavior), and various interpretability and evaluation methods for detecting misalignment before deployment.
The reason the field exists: as AI systems become more capable, the cost of misalignment goes up. A model that misinterprets a chat request wastes your time. A model that misinterprets a request when it has broader autonomy — writing code that runs on production systems, executing financial transactions, planning multi-step actions — can do meaningful damage. The field is trying to solve the alignment problem for the low-stakes cases now, before it becomes urgent in the high-stakes cases later.