AI is built by people who assume their own ‘normal’ is the global standard. Then they ship it. Suddenly, a codebase written in one specific cultural bubble is being forced onto the rest of the world, regardless of whether it actually makes sense in the local context.
This is not really a data problem, and it does not show up on a bias-testing dashboard.
It shows up in the assumptions baked into the question an assessment asks in the first place. A great deal of ethical review, reasonably, starts from the individual:
- did this person consent,
- was this person’s data used fairly,
- was this person treated the same as everyone else.
That is a real and necessary question. It is also not the only system that can fail.
Some of the oldest and most carefully worked-out ethical traditions in the world start from a different unit of concern entirely, not the individual in isolation, but the relationship, the household, the community a person is answerable to and formed by.
Under that framing, the sharper question is not only “did this person agree,” but “what does this do to the web of trust the people around them depend on.” A system can pass every individual-consent test and quietly be corroding something none of those tests noticed.
We are not proposing that one tradition is right and the other is wrong! We are pointing at the gap between them, an assessment that does not ask whose definition of harm is doing the work is not answering the question it thinks it is answering.
This is one of the reasons a finding of runs through more than one ethical lens before it goes out, not as an academic exercise, but because a system built to pass one tradition’s test can still fail the other’s, and a client only finds that out in public if we do not check for it first…