The technical debt bill for AI-generated code is coming due
Google says 75% of its new code is AI-generated. Seventy-five percent of technology leaders expect moderate to severe technical debt this year. Those two numbers are not a coincidence — they are the same trend measured from two ends.
Seventy-five percent of technology decision-makers expect moderate to severe technical debt by year end. Google has disclosed that 75% of its own new code is now AI-generated. Those numbers describing the same moment from opposite ends is not a coincidence — it is the same shift in how code gets written, showing up first as velocity and now as a bill.
Debt with a new shape
More than 90% of organizations report significant hurdles merging AI-generated code with existing infrastructure, and the debt taking shape is not entirely the kind engineering teams know how to manage. Epistemic debt — code too complex or unfamiliar for the team maintaining it to safely change — is emerging as a distinct category, alongside a rise in code smells that do not break anything immediately but accumulate into real maintenance cost over time. The uncomfortable distinction from classic technical debt: the old kind was usually created by a rushed human who could explain the shortcut later. The new kind is sometimes created by a model whose reasoning nobody on the team ever actually reviewed.
Why "AI writes better code" missed the point
The debate over whether AI-generated code is better or worse than human-written code was always the wrong axis. The more consequential question is whether the team maintaining it understands it well enough to change it safely six months from now, under a different engineer, without the original prompt or context available. A codebase can be functionally correct today and still be a liability if the person who has to modify it in Q3 cannot reason about why it works. That is what "faster code, deeper debt" is actually describing — velocity purchased against a maintainability balance nobody is tracking on the same dashboard as delivery speed.
The fix is process, not a better model
AI tools already exist for automated code comprehension — parsing legacy and AI-generated code alike, extracting structural documentation, flagging duplication. The gap is not tooling. It is that most teams generating AI code faster than ever have not embedded that generation into a disciplined engineering framework: mandatory human review proportional to complexity, documentation requirements that do not get waived because the code shipped fast, and an explicit test for whether a reviewer can explain the change, not just approve it.
Track epistemic debt as its own metric — code nobody on the current team can confidently explain — separately from raw code volume or PR velocity. The 75% figure on both sides of this story is the same debt, measured as speed on one end and as a bill on the other.