Ethics bolted on at the end catches what’s already shipped: a pre-launch review that flags a problem in a system that’s mostly built, mostly funded, and psychologically hard to send back for a redesign. Ethics by design catches something bolted-on review structurally cannot: the option that was never built at all, because the question got asked before the architecture was locked in.
The two are not simply different points on the same timeline, they catch different categories of failure.
Bolted-on review is good at spotting a specific harmful output, a biased score, an unsafe recommendation, because there is a working system to test it against. It is structurally not good at catching the harm that comes from the system’s basic shape, because by the time there is something to test, that shape is already load-bearing.
This is why bolted-on ethics review so often produces the same finding pattern: small, fixable issues get flagged and fixed, while the larger, structural choice, the one made months earlier when the system was scoped, sails through untouched. The review is not failing, it is doing exactly what a late-stage review can do, which is narrower than what people assume it does.
Ethics by design is not a purer, more virtuous version of the same activity, it is a genuinely different intervention point: the question “should this system optimise for this, and who does that choice put at risk” asked while the answer can still change the architecture, not just the output. Designed in catches what never should have been built that way in the first place, which bolted-on review, however rigorous, was never positioned to catch.