Over the past year, the way I work has changed. I still build systems and write code, but I rarely start from nothing. I use large language models to generate draft code and suggestions, then spend most of my time shaping, correcting, and refining what they generate. Outputs arrive quickly, often good, sometimes excellent. Productivity is up. And yet, something feels off. It feels, vaguely, like cheating.

To counter this, I rarely accept any LLM solution without checking it carefully. As long as I verify every line, I tell myself I am still in control, I am just an unusually fast developer using modern tools. I do not want to resist technology out of pride or nostalgia. Tools are inevitable, and refusing them feels pointless. Still, the discomfort persists. If I am thinking deeply while reviewing the output, what exactly is being lost?

The answer lies in the distinction between generative and evaluative thinking.

Generative thinking creates structure where none exists by framing problems, proposing explanations, and deciding which questions are worth asking. Evaluative thinking operates on existing structure by judging correctness, refining outputs, and choosing among alternatives. Both are non-trivial but they are not same. Evaluation presupposes generation. There is nothing to critique unless something has first been created.

Evaluative thinking feels sufficient because it works. It is fast, scalable, and visibly productive. Reviewing strong AI output can feel indistinguishable from having generated it yourself. In most everyday situations, nothing breaks. The system delivers acceptable results with minimal friction. So the temptation is rational. If refinement reliably produces outcomes, why struggle to generate at all?

The problem is not performance; it is atrophy. Evaluative thinking is not shallow, but it is derivative. It operates within an existing space of possibilities. It improves, selects, and combines, but rarely expands that space in a meaningful way.

· · ·

One might argue that this is not new. We have always relied on books, reference implementations, and prior work. No one truly starts from scratch, nor should they. Progress has always depended on “standing on the shoulders of giants.” So why should it matter if the giant is now an LLM?

The difference is subtle but important. Before LLMs, external sources rarely provided complete solutions. They offered fragments—ideas, techniques, and examples. One still had to assemble them, adapt them, and bridge gaps across contexts. Generative effort was unavoidable. With increasingly capable AI, that gap is shrinking. The system now proposes coherent solutions from start to finish. What remains for the human is judgment. We are moving from generative thinking, to assisted generation, and potentially to pure evaluation in the future.

That transition is what worries me. To see why, it helps to look at physical effort as an analogy.

For most of human history, physical effort was embedded in daily life—walking, lifting, carrying. When physical movement disappeared from daily life, we did not immediately notice the loss. Life became easier and more efficient. Nothing seemed broken. What we lost was the ability to function healthily without movement. Then we invented “workouts” as a proxy for healthy living. Something similar may be happening to thinking.

· · ·

The deepest loss is not creativity in the artistic sense. It is problem discovery. If generative thinking disappears from everyday work, intelligence will continue to operate, but mostly within what is already known or only slightly beyond it. Some argue that writing the prompt is the new generative act. But even prompting is often just specification. True insight lies in noticing that something does not fit before it can be articulated. If the machine handles the pathfinding, our ability to notice when the map itself is wrong begins to erode.

If this trend holds, the future of thinking may resemble the history of physical fitness. Intelligence will continue to function, but the habit of generation will no longer be required by default. It will survive as a conscious exercise, practiced deliberately, not for immediate output, but to preserve the ability to think outside the box.