Then came Black Friday.

The system crossed an invisible throughput ceiling that had existed since day one. Nobody knew the ceiling was there because nobody had reasoned about why the code was structured the way it was. Response times climbed from 50ms to 2 seconds, then 10. The database connection pool exhausted. The checkout pipeline collapsed. The incident lasted 47 minutes. Customers were charged with no corresponding record in the system.

The code had two bugs hiding in plain sight. A transaction boundary that pinned database connections during external API calls, and a missing safety net that meant customers could be charged with no record in the system if anything failed after the charge succeeded. Any senior developer who looked at the code would have caught both. Neither is obscure.

The question is not what went wrong in the code, what the bug was, or how it was fixed. The question we should all be asking is, why nobody looked.

Why AI-generated code passes code review

Here is the thing about AI-generated code that makes it uniquely dangerous to review: it looks like code you have seen before.

Human-written bad code usually has a smell. Weird naming. Inconsistent style. A comment that says // TODO: fix this later. An unusual pattern that makes you pause and ask “What the actual hell? What was this person thinking???” Those friction points are not bugs. They are signals.

AI code has none of those signals. It is clean. It is too clean.

Code review has always been a pattern-matching exercise. We scan for things that look wrong. AI broke that heuristic by producing code where nothing looks wrong but critical things are wrong. The problems in this payment service are not in what the code says. They are in what it does not say: no timeout, no retry, no idempotency, no operational awareness. Absences do not trigger alarms in a 30-second review.

And there is something deeper going on. When a human writes code, you can review their intent. “Why did you choose this approach?” is a question that has an answer. When AI generates code, there is no intent to review. Nobody sat in a chair and chose @Transactional here. It was statistically selected. The most common pattern in the training data. You cannot review intent that does not exist. I wrote about this distinction in Your Coding Assistant Is Not You.

So the developer who merged this PR did not make a mistake. They did exactly what code review has trained them to do: check that the code looks correct, verify the tests pass, and approve. The process worked perfectly. The code broke anyway. Because the process was designed for a world where humans authored code and reviewers could question their reasoning. That world my friends, is gone.

It’s in my repo. But it’s not my code.

The first two debts in this series (comprehension debt, homogeneity debt) are about code. This one is about people.

Ownership debt is what happens when your team stops feeling responsible for the code in their own repository.

Not because they are lazy. Not because they do not care. Because the relationship between developer and code has fundamentally changed. Before AI, you wrote code. It was yours. You made every decision. If it broke at 2 AM, you knew where to look because you remembered why you did it that way.

Now? You merge code. You reviewed it for 30 seconds. The tests passed. It looked reasonable. You clicked approve. And three months later, when it is on fire in production, you are staring at it like a stranger’s luggage at baggage claim. It looks like it’s your bag. It has your name on the tag, but you have no idea what is inside.

That is the shift. From author to curator. From “I built this and I’ll defend it” to “I approved this and I hope it’s fine.”

And here is the psychological piece that nobody talks about. When you write code, you feel ownership. It is yours. You are proud of it, or at least you are responsible for it. When it breaks, it is personal, it is your baby. You spent many long hours, into the wee small hours of the morning, to get this perfect. And that is actually a good thing, because it means you care enough to fix it properly.

When AI writes code and you merge it, there is a psychological distance. It is not yours. You are a curator, not a creator. And when it breaks, the instinct is not “let me understand what went wrong.” It is “let me get the AI to try again.”

Many years ago, in the olden days (we are talking about 2022 of course), when something broke, the developer said “I’ll fix it, I wrote it, I know where to look.” In 2026, when something breaks, the developer says “I am going to paste the error and have AI solve it.” Or, “Let me try regenerating it with a different prompt.”

That is not debugging. That is gambling. You are rolling the dice and hoping the AI gives you a better answer this time.

And what happens when the AI’s fix introduces a new bug? They ask the AI again. And again. And now you are three layers deep in AI-generated patches on top of AI-generated code, and nobody, nobody, has a mental model of how this pipeline actually works.

The payment service incident played out exactly this way. Nobody in the incident room felt it was their code. Nobody felt responsible for understanding it before it shipped. The developer who merged it said “well, the AI wrote that part” not as an excuse, but as a genuine expression of distance. They were describing reality. It was not theirs. They never made those decisions. They just approved them.

And the first instinct during the incident was not to read the code and understand the failure. It was to ask the AI to generate a fix. Under pressure. On a Black Friday outage. Gambling.

How this debt compounds - in people, not in code

Technical debt compounds in a codebase. Ownership debt compounds in a team.

Every time a developer approves AI-generated code without deeply understanding it, the next review gets a little shallower. Not because they are being careless. Because the muscle of “I need to understand this before it goes in” is not being exercised. It atrophies. One shallow review makes the next one easier to justify. “I approved something similar last week and it was fine.”

After six months of this, a team’s relationship with their own codebase has quietly changed. They stop treating AI-generated modules as their code. They start treating them as something closer to a vendor library. Something you use, configure, call into, but do not open. Something where the response to a bug is “let me re-prompt the AI” rather than “let me read the code and understand the failure.”

Except it is not a vendor library. It is your business logic. Your payment processing. Your authentication flow. Your company’s name is on the invoice when it fails. And nobody on the team can explain why it is structured the way it is, because no human ever decided that.

This is the part that should worry engineering leaders. It is not one bad PR. It is a slow, invisible drift where an entire team gradually becomes strangers to their own system. And you do not see it happening because every individual review looks fine. The tests pass. The code is clean. The approvals are fast. It all looks like a high-performing team. Until the incident.

And this has a human cost beyond incidents. Developers who spend their days reviewing and merging AI output start to feel like operators, not engineers. They are not solving problems. They are supervising a machine that solves problems. That is not why most of us got into this field.

Your best developers, the ones with options, they notice this first. And they leave. Not because the AI is bad, but because the work stopped being interesting. I explored this exodus in The Next Casualty of the GenAI Revolution. Ownership debt does not just degrade your codebase. It degrades your team.

What makes this different from technical debt?

Traditional technical debt is a shortcut you chose. You know where it is. You can explain why you took it. You remember. It is yours.

Ownership debt is a decision someone else made. A model selected the most probable pattern and you inherited it without examination. You cannot explain it because you were never part of the decision. You did not even know a decision was being made.

The critical difference: traditional debt has an owner. Someone who can say “Yeah, I did that because we were rushing to ship and I knew it was wrong.” Ownership debt has nobody. The code is collectively everyone’s and individually nobody’s. Nobody volunteers for on-call on the AI-generated modules. Nobody says “I’ll be the one who answers when this breaks.” It is an orphan wearing your team’s jersey.

And it accumulates at AI generation speed: 10 to 50 times faster than human-authored debt. I wrote about this acceleration in Tomorrow Is Coming to Collect. Because every time a developer prompts an AI, reviews the output for 30 seconds, runs the tests, and merges, another batch of decisions enters the codebase that no human fully examined.

How to prevent ownership debt

Unfortunately, there is no tooling fix for ownership debt. You cannot lint for it. You cannot add a CI check. The debt lives in people’s heads, not in the code, so the solution has to change how people relate to the code.

Three things that work.

Before you merge: explain why, not what. Before merging any AI-generated PR, the developer writes - in the PR description, in their own words - why the code is structured the way it is. Not what it does. Why this approach and not another. What the tradeoffs are.

For the original payment service, a proper explanation would have read:

The DB connection is held for the duration of the gateway call. With our pool size of 10 and gateway p99 of 200ms, this is acceptable at current load. Needs revisiting if concurrent payment volume exceeds ~8/sec. If the commit fails after a successful charge, the PENDING row rolls back and we lose the audit trail - acceptable only if gateway charges are idempotent.

Notice what happens. Writing “the DB connection is held for the duration of the gateway call” forces the developer to think about that fact. At which point they might ask: “Wait, is that actually acceptable?” And the conversation starts. The debt is surfaced before the merge, not during an incident.

If you cannot write that explanation, you do not understand the code well enough to ship it.

For critical paths: rewrite the scary part by hand. Identify the code paths that handle money, user data, authentication, and core business logic. If those paths were AI-generated, have a developer rewrite the critical sections. Not because the AI’s version is necessarily wrong. Because the act of writing builds the mental model that reviewing never will.

A developer who hand-writes the payment state management will instinctively separate the DB transaction from the gateway call. They will add the idempotency key because they have been paged at 2 AM before. They will handle the timeout because they are thinking about what happens when things go wrong, not just what happens when things go right. The code they produce will be theirs. And when it breaks, they will know where to look.

Watch how your team talks about their code. “The AI’s payment module” versus “our payment module.” The pronoun tells you everything. If your team refers to parts of their own system as belonging to the AI, ownership debt is already compounding. In retrospectives, add a standing question: “Did we discover code this sprint that nobody on the team could explain?” Silence is the answer you are listening for.

The goal is not less AI. It is more ownership.

AI coding tools are not the enemy. They are genuinely powerful, and teams that use them well have a real advantage. The teams that will struggle are the ones that mistake “merged faster” for “owned.”

The fix is not to stop using AI. The fix is to stop pretending that approving code is the same as understanding it. Everything in this article has one purpose: turning your team from curators back into authors. Explaining why forces you to think. Rewriting the critical path forces you to engage. Watching how your team talks about their code tells you whether it is working.

Because the real crisis is not a transaction boundary bug. It is the moment when an incident fires and everyone in the room looks at each other and says “I just merged that, I didn’t write it.” That moment is preventable.

The best code is code your team would fight to defend at 2 AM. Not because they have to. Because it is theirs.

Give this a try on your team next week, I would be genuinely interested to hear your thoughts or comments and how this worked out for you, so please feel free to ping me on Twitter or LinkedIn.