Taking over someone else’s codebase has always meant inheriting decisions you weren’t part of. What’s changed recently is what some of those decisions actually were, because a growing share of the code sitting in a handover was written by an AI tool rather than a person, and the developer data on how well that code holds up isn’t reassuring.
Stack Overflow’s 2025 Developer Survey found 84 percent of developers now use or plan to use AI tools in their workflow, up from 76 percent the year before. That’s close to standard practice now. But the same survey found trust in that code’s accuracy sitting at 29 percent, down from 40 the year before, and separate research from Google’s DORA program found 39 percent of developers reporting little to no trust in AI generated code specifically.
The practical effect shows up as duplication. Every AI generated shortcut that solves a problem in a slightly different way than the rest of the codebase is a place a future fix has to be applied twice, or missed entirely because nobody remembers it’s there. Industry analysis tracking this pattern has tied rising AI code generation to measurable increases in exactly this kind of churn, the raw material technical debt is made of.
None of this means AI generated code is worse than the code it’s replacing. Legacy codebases built entirely by hand have always accumulated their own version of the same problem, inconsistent patterns, undocumented decisions, workarounds nobody wrote down. What’s different now is the volume and the speed at which it accumulates, because a tool that can generate a plausible looking solution in seconds can also generate a subtly wrong one in seconds, at a scale no individual developer previously could.
What actually breaks first when we take over a handover is rarely the big, obvious thing. It’s the small inconsistency nobody flagged, three different approaches to the same problem sitting in the same file, each one looking confident enough on its own that nobody questioned it until something downstream stopped working. That pattern predates AI tools, but AI tools that generate confident, plausible, almost right code at speed make it easier to accumulate more of it faster.
The honest takeaway from the developer data isn’t to avoid AI assisted code. It’s to treat a codebase’s provenance as something worth actually knowing before you take it over, not because AI written code is automatically worse, but because knowing which parts were generated quickly and which were reasoned through carefully tells you where to look first when something doesn’t add up. Debt doesn’t care who or what wrote the line, only that someone eventually has to pay it back.
Debt doesn't care who or what wrote the line, only that someone eventually has to pay it back.