ChatGPT
Perplexity
Technical debt used to announce itself. You could open a file, see the tangle, and put a refactor on the roadmap. The debt that AI-assisted development introduces behaves differently, because it arrives looking finished.
The code compiles, the tests pass, the diff reads clean, and it moves through the same review gate you have always trusted. Your gate was built to catch debt you can see, and this is the other kind.
AI defects survive because your review waves them through
The assumption is that whatever an assistant gets wrong gets caught in the normal churn of development. The largest study to date says otherwise and points to the reason.
Researchers tracked 302,000 verified AI-authored commits across 6,299 public repositories and followed each issue over time. More than 15% of commits by each assistant introduced at least one issue, and 22.7% of those issues remained in the repository’s latest version.
They persist because the gate that should stop them does not. When peer-reviewed work compared how reviewers responded to AI-generated pull requests with human ones, the AI requests elicited more neutral and positive sentiment, even though they carried measurably more redundancy.
Surface plausibility read as quality, and the same excessive trust showed up on the author’s side. This is public code, so treat the exact numbers as directional. The direction is what should worry you.
The control you would name if someone asked how debt gets caught is the one that quietly lets it through.
Stay updated with Simform’s weekly insights.
Comprehension is the expensive loss no diff shows
Redundant code you can eventually find and fix. The harder debt is the understanding that never formed.
In a controlled trial, developers who used AI to finish a task scored 17% lower on a comprehension quiz afterward than those who worked without it, with the widest gap in debugging.
The split inside the study is worth weighing. People who stayed engaged with the tool scored well, and those who delegated the thinking wholesale scored far below.
You can ship faster while understanding your own system less. Margaret-Anne Storey gave this a peer-reviewed name, separating the debt sitting in your code from the erosion of shared understanding among the people who maintain it.
When the module no one reasoned through finally breaks, no one holds the model to fix it, and in a shop of five to fifty engineers, that lands on the one or two people already carrying the load.
The rationale lived in a prompt, and the prompt is gone
Even where the code is sound, something goes missing. The reasoning behind an AI-generated change often lived in a prompt no one saved, so the intent behind the code cannot be reconstructed later. Storey names this intent debt: the absence of the externalized rationale that people and future agents need to change a system safely.
It gets worse as models move. In peer-reviewed testing, 58.8% of prompt-model combinations lost accuracy when the underlying model was updated, and formatting that worked on one model carried over only weakly to the next.
An instruction that produced good code last quarter is not a dependable asset this quarter. When the intent existed only in that instruction, a model upgrade can quietly turn working code into code no one can account for.
The teams getting returns changed what they measure
The reflex is to tighten review and slow the merge. That treats a visibility problem as an effort problem, and it misses, because the tools you would tighten are the ones already passing the code.
The teams seeing real returns moved elsewhere. The 2025 DORA report, drawing on nearly 5,000 practitioners, found that AI is now linked to higher delivery throughput under the right conditions and framed it as a multiplier that amplifies whatever foundation it meets.
What it amplifies upward is unglamorous groundwork, AI-accessible internal knowledge, small batches, and version control strong enough that a bad change is cheap to unwind.
In GitHub’s controlled study with Accenture, developers accepted only about a third of the assistant’s suggestions. Still, they shipped with a 15% higher pull-request merge rate because the human stayed the gatekeeper and the output remained a draft.
The moves that get you there are cheap. Tag your AI-authored commits and track your own issue-survival and rework rate, so you manage your numbers instead of a study’s.
The review question shifts from whether the code is correct to whether the author can explain why it works. And require a versioned specification or a decision record for AI-generated components, so the next model upgrade becomes a regeneration problem rather than a lost-knowledge one. Start where one person already holds all the context for a module, because that is where the next gap opens.
AI amplifies the discipline you already have, or the lack of it
None of this argues against AI in development, nor would the research support it. McKinsey found that the top quintile of performers achieved software-quality gains in the 31% to 45% range, getting there by re-architecting how they build rather than by buying a tool.
The teams pulling ahead treat understanding and intent with the same rigor they already spend on code. So the question worth putting to your own delivery system has less to do with whether the code AI writes is good, and more to do with whether anything you do captures what it was for, once the prompt that made it is gone.
If you are putting AI to work across your SDLC, the discipline that keeps generated code accountable lives in specification, context architecture, and independent verification, not in the review gate alone. See how Simform approaches a governed Agentic SDLC.
