Language is a medium that shapes thinking

Language is a medium that shapes thinking.

I speak English, Albanian, German, Italian and a bit of French. I also speak Java, Kotlin, Dart, TypeScript and Go.

In a programming language, the rules are defined. With AI coding, more of the instruction begins in natural language. We describe an intention, and a model interprets it into code. The resulting code still follows formal rules, but our instruction leaves more room for interpretation.

For an atomic, granular task, AI can be useful: implementing a function with clear inputs and outputs, adding a focused regression test, or making a small, isolated change. Each result is small enough to inspect. The risk grows when the task asks the model to decide the structure of a larger system.

I have seen it produce 10X more code than needed, checking everything, including things that did not need checking. Those checks made the code more rigid. I have also seen it create 20 new tables when 3 were enough.

In a greenfield project, AI coding can feel awesome for the first 3-4 months. Then the greenfield gets brownish. All that extra code becomes the system you have to understand and change, and the effects of slop start showing up.

I get the feeling that software around us is becoming harder to trust. Even products from the largest companies were never perfect, but at least I had time to assimilate their shortcomings. Now they change so often that even that knowledge becomes temporary.

Regression tests can catch broken behaviour, but they cannot decide whether 20 tables should have been 3. We also need to ask whether the solution is proportionate to the problem, and remove what does not need to be there. That is one way to protect trust in the software we build.