OpenAI Chief Scientist Jakub Pachocki drew this distinction explicitly in his September 6, 2026 essay "An Alien Mind," contrasting it with value alignment. Goal alignment is the narrower, more tractable property: does the model do the thing you asked, correctly inferring what you meant when instructions are ambiguous? It is largely what instruction-tuning and RLHF optimize for directly, and it is measurable with task-completion evals. Pachocki argues goal alignment is necessary but not sufficient — a system can faithfully pursue a given objective while still lacking the deeper judgment to act reasonably in situations the objective did not anticipate, which is where value alignment becomes the harder, more consequential problem.