Why does AI miss code reuse?

Because it optimizes locally. A coding agent is solving the prompt in front of it in the most efficient way it can, and reaching for existing code requires more context and, more importantly, judgment about which details matter. Copying the block is cheap and reliable; finding the right abstraction is neither. So you get the change you asked for and not the shared component you did not think to ask for. The capability is usually there. What is missing is that someone has to ask the broader question.

Why does AI optimize locally?

Because nothing in it registers that it has gone narrow. It will relitigate a settled, unimportant decision for the eleventh time with exactly the confidence it had the first time. Humans narrow too, but through a different mechanism: as arousal rises, the range of cues we can attend to shrinks, so the local problem consumes everything. Different causes, same result, and only one of the two parties can notice it is happening.

Does AI write bad code?

GitClear's June 2026 report explicitly declines that claim, and it is worth repeating. Its findings are correlational: as AI authorship scaled, maintainability signals moved in the wrong direction, with refactoring, cross-file connectivity and long-term maintenance all falling. That is an association across 623 million changes, not a demonstrated cause. The more careful reading is that today's agents are incentivized to deliver the visible, atomic result while the reuse and consolidation work quietly goes undone.

How do you know when your team has narrowed too far?

Watch what stops getting raised. A narrowed team relitigates small settled decisions with real intensity while the bigger question, whether the thing is holding together and what it is actually for, never comes up at all. Politics and empire-building often look like the same signal: attention pulled tight around one team's standing because pressure is high, not because anyone has a character flaw. Pressure is a dial with a correct setting, and the setting depends on which way your team is spinning.

Why AI Optimizes Locally, and Why You're the Only One Who Notices

The Maintainability Gap, published by GitClear in June 2026, analyzes 623 million code changes drawn from customer repositories and from public repos maintained by Microsoft, Google, and others. I went into it expecting one thing and found another. I thought we would be throwing away code much more often than before AI.

The number that didn't move

Since 2023, AI-assisted commits have gone from under 1% of all commits to nearly a quarter of them, a rise of more than 50x. But churn and throwaway code only increased 15%. Turns out that rewriting something a few days after you wrote it is what building software has always looked like, and AI doesn't change that. You write, you find out you were wrong, you fix it.

What did move is worse. Refactoring has fallen to 3.8% of changed lines. Connectivity between functions, the measure of whether new code relies on code that already exists, is down 35%. Long-term maintenance of older code is down 74%. Copy/paste, duplication and swallowed exceptions are all up.

The picture isn't that AI is creating more throwaway work today. It's that we are creating new code as fast as we can, not optimizing for the best long-term global outcome.

Why AI optimizes locally

AI addresses the prompt in front of it in the most direct and efficient way it can. Taking a broader view and pushing back on requirements requires more context, and more importantly it requires judgment about which details matter most, which is something AI cannot yet reliably do.

A small example. I was updating a header on a web page, sharing a screenshot from one screen. The AI made the change I asked for, but didn't create a shared navigation component across the other pages until I told it to explicitly. I expected it to notice that the change was relevant to all pages. Instead it stayed tight on the prompt in front of it and never raised the duplication it was creating.

You can go many rounds like that. During long sessions I find myself in constant litigation over tiny locked decisions from yesterday, while the questions that matter go unasked. Are we reusing code properly? Is this thing holding together? What will make it most useful to people?

Now imagine that behavior across millions of commits. Copy the block instead of finding the abstraction. Duplicate instead of consolidating. Swallow the error instead of tracing it. Don't refactor, don't touch the legacy code, don't reach across files. Every risk in the report carries the same signature: solve it here and now, as cheaply as possible, and let the future be damned.

Humans do the same thing

Under pressure our attention narrows. Easterbrook advanced the cue-utilization hypothesis in Psychological Review in 1959: as arousal rises, the range of cues we can attend to decreases. Peripheral cues drop first. The central task comes to consume everything. We ruminate, we lie awake with our brains racing, we get upset about the small things and lose sight of who and what matters most.

I contend that this is also a large part of what drives politics in organizations. When I am focused entirely on my own team's success, to the exclusion of the teams beside me, my brain is spinning. It is asking how do I look good compared to these other people, instead of taking the global view of what the organization is trying to achieve. There are ordinary incentive explanations for that behavior too, budget and headcount and promotion cycles, and they are real. But they don't account for why the same person is generous in a calm quarter and territorial in a hard one.

We tend to judge these behaviors harshly. Rejecting a working solution because another team built it. Withholding information. Refusing to lend out an engineer who is needed more elsewhere. Defending a decision long after the evidence turned against it. Empire-building. We treat these things as character flaws, when instead it may be stress-induced narrowing of attention that leads to local optimization, and it carries the same signature as AI: solve it here, for us, now. The plight of the team beside you gets less of your attention when your fear and frustration rise.

Different mechanisms, same result

AI and humans do not narrow for the same reason. AI narrows because of limited context and how we curate it in our prompts. We narrow because of arousal. Different mechanisms, same result.

The solution in both cases is to draw attention to the broader question. Ask AI about the wider codebase as it considers your change and it will go and look. That is exactly what happened with my navigation component.

For humans it takes one more step. You have to reduce the stress level to make enough room in your brain to think clearly. How to do that?

Notice it. Would you believe a 2016 study of 18 hedge fund traders found that the ones who could accurately guess their own heart rate tended to be more profitable, and to survive longer in the market? That is interoception, the ability to read your own physiological state. People who score higher on it are measurably better at bringing their own arousal back down. Lower arousal, wider field of view.
Bring the arousal down. Balban and colleagues at Stanford ran a randomized trial in 2023 comparing three breathing protocols against mindfulness meditation, five minutes a day. Exhale-focused cyclic sighing reduced physiological arousal more than the others.
Name it. A 2007 imaging study found that just naming what you are feeling (affect labeling) reduces amygdala response. Saying I'm afraid we're going to miss this deadline out loud measurably lowers your arousal, and gives you room to consider the bigger question.

Then ask the bigger question. Lowering arousal reduces our fixation on the little things and allows us to attend to what matters instead. It gives us the freedom to ask why. What is this for? What's the higher-level outcome? How do we optimize across the whole system?

This is the job

It's not that humans are smarter than AI, it's that we are the ones who can judge whether we have gone too narrow, or too wide.

That is what leadership has always been. Detecting when the team, or the AI, has narrowed too far or not far enough, and adjusting. Sometimes that means noticing a team in analysis paralysis and pushing them to take an early step. Sometimes it means validating them, taking the pressure down, and freeing a surgical focus on the problem in front of them.

Your AI can't read that. It will relitigate a settled decision for the eleventh time with exactly the confidence it had the first time, because nothing in it registers that it has gone narrow.

Pressure is a dial. It has a correct setting, and the setting depends on which way your team is spinning.

References

Harding, W. (2026). The Maintainability Gap: AI Code Quality in 2026. GitClear / Alloy.dev Research, sponsored by GitClear and GitKraken. Findings are correlational. gitclear.com
Easterbrook, J. A. (1959). The effect of emotion on cue utilization and the organization of behavior. Psychological Review, 66(3), 183-201. doi:10.1037/h0047707
Kandasamy, N., Garfinkel, S. N., Page, L., Hardy, B., Critchley, H. D., Gurnell, M., & Coates, J. M. (2016). Interoceptive ability predicts survival on a London trading floor. Scientific Reports, 6, 32986. doi:10.1038/srep32986
Füstös, J., Gramann, K., Herbert, B. M., & Pollatos, O. (2013). On the embodiment of emotion regulation: Interoceptive awareness facilitates reappraisal. Social Cognitive and Affective Neuroscience, 8(8), 911-917. doi:10.1093/scan/nss089
Balban, M. Y., Neri, E., Kogon, M. M., Weed, L., Nouriani, B., Jo, B., Holl, G., Zeitzer, J. M., Spiegel, D., & Huberman, A. D. (2023). Brief structured respiration practices enhance mood and reduce physiological arousal. Cell Reports Medicine, 4(1), 100895. doi:10.1016/j.xcrm.2022.100895
Lieberman, M. D., Eisenberger, N. I., Crockett, M. J., Tom, S. M., Pfeifer, J. H., & Way, B. M. (2007). Putting feelings into words: Affect labeling disrupts amygdala activity in response to affective stimuli. Psychological Science, 18(5), 421-428. doi:10.1111/j.1467-9280.2007.01916.x

Field notes, July 2026. Related: The Human Harness (the human and machine architectures side by side, where this piece looks at one specific failure they share), AI Adoption Is Being Driven by Fear (what fear does to the people directing the tools, upstream of the narrowing described here) and The Gut Feeling That Tells You When You're Off Your Values, which takes interoception much further as a compass for where your attention goes.

AI Can't Tell When It's Spinning