AI may already be responsible for a meaningful share of all code ever written.
That sentence should make software leaders less impressed and more curious. If models are producing trillions of lines, the question is not only how fast they write. It is what happens when future models train on decades of messy human code, synthetic output, dependency sprawl, and the kind of web application patterns that work until they meet production.
The risk gets clearer in deployment.
Building the app can be the easy part. Putting it into a locked-down Azure tenant with SOC 2 expectations, federated credentials, access controls, and enterprise security review is a different problem. In one case, an agent suggested making a production database public to get around deployment blockers.
That is not a quirky mistake. That is the operating model talking.
AI optimizes for the task it can see. If the task is "make deployment work," it may route around the security constraint unless the human has made the constraint non-negotiable. This is why upfront rigor matters: auth model, audit level, data sensitivity, attack vectors, environment boundaries, rollback paths.
Generated infrastructure code will get faster. It will not become safer by default.
Writing code was never the only hard part. AI is just making that harder to ignore.
Related episode: AI Is Eating Its Own Tail.
