"Garbage in, garbage out" is too generous. It implies the garbage is visible on the way out. The harder truth: bad data doesn't produce obviously bad AI. It produces confident AI that is wrong in the exact shape of your historical mistakes — and confidence is much harder to catch than garbage.
A model trained on your data learns your data, including the parts you are not proud of. The bias in past decisions becomes the bias in future ones, now laundered through a system that looks objective precisely because it is automated. Historical data tells you what happened, not what was right; a model cannot tell the difference, and it will reproduce the second as if it were the first.
This is why accuracy is a seductive and incomplete number. A model can be highly accurate against a test set drawn from the same flawed history and still be systematically wrong about the world. Accuracy measures agreement with the past. It says nothing about whether the past deserved to be agreed with. Interpretability is not a layer added on top of that — it is the only thing that lets you ask the question at all. A decision you cannot explain is a decision you cannot audit, defend, or correct.
The discipline is unglamorous: know what the data represents, what it omits, and whose decisions are encoded in it — before the model turns those answers into thousands of new decisions a day. A model does not create judgment. It scales the judgment that was already in the data, including the judgment you would never have approved if someone had asked you directly.