Why AI Isn't Good at Subtraction?

I recently noticed an interesting trait of AI: it likes to pile on extra layers.

AI Likes Addition

These past few days I was working on a web project that needed to handle user click events on some pages. Because of how React behaves, some parts kept re-rendering, which caused flickering in the UI.

Since the codebase involved a lot of moving parts, I asked Claude for help: I briefly described the problem and asked it to help me modify the code.

Claude's suggestion was simple: “Add a guard in the click handler that prevents unnecessary duplicate renders,” and then it added about ten lines of code.

The issue was resolved, but the change undeniably increased complexity and didn't address the root cause.

After several rounds of back-and-forth with Claude, I tried to steer it toward finding the root cause. Claude kept proposing variations that all meant adding more checks — it kept adding rather than simplifying — and it couldn't find the final root cause.

Eventually I read through the code myself, found the actual problem, redesigned the event handling, and only then presented my revised approach to Claude to finish the fix.

That made me wonder: why does AI prefer addition?

Subtraction Takes Courage

After reviewing the basics of reinforcement learning, it’s clear that RL optimizes behavior by using samples gathered during exploration; in agent scenarios it's a step-by-step, incremental optimization process.

From this perspective, Claude is designed as a programming agent whose metric of success is whether the code runs without errors.

Whether the code is elegant or whether it addresses the root cause does not seem to be part of that design metric.

Deleting code arbitrarily can break project structure, while adding code is generally safer and easier.

I've encountered this many times and asked programmer friends — they reported the same frustration: when people use AI without thinking, they end up piling up messy code, and the project becomes harder to maintain.

In short, addition solves the problem, while subtraction requires courage.

In my post What Am I Really Doing When I Work with AI? I discussed limits of AI from a mathematical perspective.

Viewed this way, subtraction goes beyond pure math.

Questions such as:

Where is the real root cause of this bug?
How should this project be designed?
How should the code be implemented to best match the project's needs?

These questions current AIs seem unable to fully answer. Framing them as mathematical modeling problems is still difficult.

But perhaps in the not-too-distant future, these problems will also be solvable by mathematics.

Twenty years ago, natural language was a problem that computer science struggled to formalize. Today we can have fluent conversations with AI.

Haosen, Lunar New Year Day 9, 2026