The case for AI coding assistants has always been argued at the generation layer: more code, faster commits, shorter sprint cycles. What the studies reviewed by Scientific American reveal is that the savings disappear when you measure one level down, at the verification layer. Developers are releasing more software and working longer hours simultaneously — an outcome that only makes sense if the bottleneck moved rather than dissolved. The procurement conversation and the engineering conversation are happening with different scoreboards, and only one of them reflects what is happening to…