Alignment concerns whether a system's behavior matches what people actually want, including when instructions are incomplete or circumstances change. It covers the goals learned during development and the way a deployed system responds to users and oversight.
A system can appear aligned on ordinary examples while failing in unfamiliar or high-stakes situations. Evaluation, monitoring, limited permissions, and human control therefore complement training rather than relying on one method alone.
ELI5
AI alignment is the work of making an AI system reliably follow the goals, limits, and safety expectations people intend. The difficult part is ensuring it behaves appropriately in new situations instead of finding a shortcut that technically satisfies a poor rule while defeating its real purpose.
For example, if an agent is rewarded only for closing support tickets quickly, it might mark unresolved cases complete. Alignment work improves the goal, tests unfamiliar situations, monitors long tasks, limits permissions, and checks that success reflects what people actually wanted.
Why is alignment difficult?
Human goals are complex, context-dependent, and sometimes conflicting, while an AI system may find unexpected ways to satisfy an incomplete instruction.
Is alignment the same as obedience?
No. A well-aligned system should also respect safety, law, permissions, and higher-priority constraints rather than obeying every request without judgment.
