The Decisions Nobody Made
Using AI to build software does not remove the requirements or the specification. Every decision still gets taken — by the model, by the prompt, by the default of a library nobody chose. What goes missing is any signal that nobody wrote them down, because an absent decision is the one failure in software with no detector: no test fails, since a test is written from a requirement, and no review objects, since review reads what is in front of it. The measurements come from a product built with AI and deployed for months: its own record held 198 entries in four registers where 664 in nine were needed to describe it, and 162 of those sat in categories the old record had nowhere to put — among them 118 named errors a consumer's code branches on, not one of which had been written down. One of the figures you can check yourself.