Essay · The values layer
Whose values should AI hold?
There is no view from nowhere. Every model that helps decide anything carries values, picked by someone, somewhere, usually without being asked out loud. The only honest question is whose, and how they were chosen.
When a model declines one request and grants another, places one answer above its neighbor, or flags a person and clears the next, it is applying values. Not in a mystical sense, in a plain one: a judgment about what matters more. Those judgments came from somewhere, a training set, a policy, a handful of choices made under deadline. The uncomfortable truth is that every AI system already holds values. We have just been letting them in through the back door and calling it neutrality.
Why "whose" is the real question
Alignment is usually framed as making AI follow human values, as if there were one settled set to follow. There is not. People who all act in good faith still disagree about fairness, risk, and what a system owes the person on the other side of it. So the question is never only "aligned to what," it is "aligned to whose," and the answer is a choice that someone is making whether or not they admit it. Pretending the choice is technical hides who is making it.
The values layer
If values are going in regardless, they should go in deliberately, in a layer you can see, question, and change, rather than baked invisibly into weights. That is what I mean by a values layer: the part of the system where what it should care about is made explicit and contestable, instead of inferred and hidden. A value you can point at is a value you can argue with. A value buried in a model is one you can only discover by being harmed by it.
Who gets a say
On matters of value, as opposed to matters of fact, there is a strong case that the people affected should weigh equally, rather than deferring to whoever built the system. Expertise belongs to questions with right answers. Values are not that kind of question. Building AI that takes this seriously means giving the governed a genuine say in what it holds, which is a harder problem than scale, and a more important one.
Read on
For the layer itself, see the values layer for AI. For why one model cannot hold one set of values for everyone, and what that implies, read on through the cluster. For the broader contest, see who sets the rules AI runs on.