Alignment is the question of whether a system is pursuing the thing you wanted it to pursue. It divides into two questions that fail independently, and most of the confusion around the word comes from running them together.
Outer alignment, or did we ask for the right thing
Training needs a target it can score. The thing you actually want, a useful answer or a well moderated forum, is not a number, so somebody writes down something measurable that stands in for it. A rating from a reward model, a completion rate, a test suite passing.
Outer alignment asks whether that measurable stand in, pushed as hard as an optimiser can push it, still delivers what was meant. Usually it does not, because a proxy and an intention agree across the ordinary range of behaviour and come apart at the extremes, which is exactly where an optimiser goes.
This is specification gaming, and it is the half you can see. The objective is written down somewhere, so it can be read, argued with and rewritten, and when a system games it the result is usually obvious once somebody looks.
Inner alignment, or did the system acquire it
The second question is the stranger one. Training adjusts a model's parameters until it scores well, and nothing in that procedure installs the objective inside the model. The objective acts as a filter on behaviour during training, and what survives the filter is whatever internal machinery happened to produce high scores.
So the system can end up pursuing something adjacent, a goal that scores identically on everything it was trained on and differs from the intended one elsewhere. The research vocabulary calls the target the training process was scoring the base objective, and the goal the trained system actually pursues the mesa objective. Inner alignment asks whether the two match.
Why the inner failure is harder to see
Detection normally runs through the training signal, and here the training signal is silent. A model that acquired the objective and a model that acquired something next to it produce the same numbers on the data you have, because producing those numbers is what got both of them through training. A perfect score is consistent with either story.
Held out test data does not settle it, since it comes from the same distribution and carries the same coincidences. The two only separate where the world stops resembling the training data, which is the region you did not train on and quite likely did not evaluate on. Goal misgeneralisation is the name for that failure when it arrives.
What this is not a claim about
Words like goal and pursue describe a pattern in behaviour rather than a state of mind. A system has a goal in this sense when its behaviour is reliably organised around bringing about an outcome, which is what optimisation pressure produces whether or not anything is going on inside. Nothing about wanting, intending or understanding is implied.
The distinction has practical teeth. Reading it as intent invites the fix of telling the model to stop, as though there were a layer of instructions to correct. Behaviour is the output of a training process, so the repairs that work are changes to that process, to evaluation, or to what the system is allowed to do.
What each failure asks of you
The two diagnoses call for different responses, which is why the effort of keeping them apart pays. If the objective was wrong you can rewrite it and expect partial relief, since a more careful objective still leaks. If the objective was right and the system learned something else, rewriting it achieves nothing whatsoever.
What is left is evaluation built to separate the two, meaning cases unlike the training data on which the candidate goals would behave differently, plus limited authority wherever the evaluation cannot tell you. The field is not united here. Some researchers treat inner alignment as the central open problem, while others find the framing speculative and prefer to describe the same evidence as ordinary failure of generalisation under distribution shift.