The Glass Dome Of AI Productivity

July 23, 2026

My dad taught me how to program. He taught me how to think about problem solving. One of his key lessons, still with me today after 20 years, was to teach oneself by reading out loud. A lesson, when read out loud, becomes much more understandable, almost magically.

Reading my own code, just like reading a chapter, or this piece, makes the topic clearer in my head. I often used to review my own code on Github before giving it to another human. I’d see improvements that I missed in my own editor. This fundamental aspect of how I used to write code has now changed with AI.

Writing code by hand has the implicit side-effect of spending more time with the code - thereby allocating more mental bandwidth to it. The classic book “Thinking Fast and Slow” talks about this: the cognitive process of complex thoughts is necessarily slow. That’s just how we are wired. Having to type the code, write the test, refactor the source, we implicitly spent more time thinking about the topic. With AI, the time we spend on any piece of code is shortened, if not entirely eliminated. It works really well for a while, for the repeating, mundane aspects of coding, but for every non-trivial change, our brain has to keep up with a much faster stream of incoming information now. We just aren’t capable of processing it all at that speed. I can’t even read every line of generated code, let alone think about it.

That would have been fine. After all, I don’t read the final executable binary in machine code either. Except, there is one tiny caveat: AI is probabilistic. There is a false dichotomy here that we have come to accept - either AI is always right or it is sometimes right. If it’s the first, we don’t need to bother at all about the code. But if it’s the second, we need to understand what AI writes, and that requires us, at the very least, to read the code.

AI’s probabilistic productivity can reliably be grounded by having a human in the loop, which would require making AI work at the slower speed of human cognition. If we want to review AI-written code with the same diligence as we did with human-written code, the whole system needs to become slower and the selling points of AI would be severely diminished. Instead, I tend to trust (hope) that AI is right. I find myself using AI to test AI-generated code and to summarize AI-generated code. The flaw in that way of working is the glaring single point of failure.

Because AI is right most of the time, we expect to rely on it all of the time. If we don’t understand how AI is changing our systems, soon the systems we own will become unrecognizable to us. Is productivity worth the price of clarity? Can we keep going at a speed that our cognitive processes cannot keep up? I reckon there will be a limit that we will hit soon. Whether we aim for the stars or not, we are living under a glass dome of productivity.

See anything to improve? Edit this post!