Developer Essay Argues AI Code Isn't the Problem — Knowing What It Does Is
A widely read blog post contends the real cost of AI-assisted programming is that engineers can no longer explain why their own systems behave as they do.
A short essay published on the personal site ssp.sh has struck a nerve among working engineers by reframing the debate over AI coding assistants. Its argument, bluntly stated in the title "The problem is not the AI code, but nobody knows anything anymore," is that the danger is not defective output but the quiet loss of shared understanding inside engineering teams. The post argues that generated software arrives without the mental model that used to accompany it, and that this gap is an epistemic problem rather than a code-quality one.
The mechanism the author describes is straightforward. When a person writes a function, the reasoning behind each choice lives in their head and travels to colleagues through code review and conversation. When a model produces the same function from a prompt, the artifact appears without that reasoning attached. The repository still compiles and the tests may still pass, but the explanation for why the code works the way it does has to be reconstructed from scratch by whoever touches it next.
That loss shows up in routine engineering work. Debugging a failure requires tracing logic that somebody has to understand first, and reviewing a pull request becomes a judgment about intent, not just syntax. If the engineer who merged a change cannot rebuild the model's path to it, review degrades into a check on whether the test suite is green rather than whether the design is sound. The essay's point is that this erosion is invisible in the metrics teams usually track.
The reaction around the piece reflects that tension rather than settling it. Readers split between dismissing the concern as overstated and recognizing it from their own teams, according to the discussion that followed publication. Similar friction has already appeared at large companies, where maintainers have pushed back on AI-generated pull requests that no one can fully account for.
What the post deliberately leaves open is what replaces the understanding that used to travel with code. Some teams are experimenting with requiring authors to record the prompt and the model's output alongside a change, or with keeping AI-generated edits small enough to review closely. None of that has become standard practice, and the author does not claim to offer a fix.
The question worth watching is whether this remains a blog-post complaint or hardens into process. If engineering organizations start writing explicit rules about how much machine-generated code can enter a repository without a named human who can explain it, that would be a real shift in how AI assistants are governed. Until then, the essay's warning sits as an open item for every engineering manager with a model in the loop.