A recent global developer survey reveals a growing paradox at the heart of modern software engineering: most developers do not fully trust AI-generated code, yet many fail to thoroughly review it before deploying. The result is a widening verification gap, where AI tooling accelerates code creation but leaves teams exposed to functional risks and quality issues.
Developers distrust AI code but skip full review
According to the survey, 96 percent of developers believe AI-generated code is not fully functionally correct. At the same time, only 48 percent say they always check AI-assisted code before committing it, exposing a significant inconsistency between perceived risk and day-to-day practice. This disconnect suggests that time pressure, convenience, and workload often override caution, even when engineers recognize the potential for hidden defects.
The study, based on responses from more than 1,100 developers worldwide, highlights that AI coding tools have rapidly become embedded in everyday workflows. A large majority of developers who have tried such tools now use them daily or multiple times per day, while only a small minority report using them less than once a week.
AI assistance now touches nearly half of all code
Developers report that around 42 percent of their current code includes significant assistance from AI models, up sharply from just six percent in 2023. They expect this share to grow to about 65 percent by 2027, suggesting that AI will soon be involved in a majority of new code written across many teams.
AI tools are being applied across a broad spectrum of work, from early-stage prototypes to internal production systems and even critical, customer-facing services. This means AI-generated or AI-assisted code is no longer confined to low-risk experiments; it increasingly underpins core business applications and infrastructure.
Verification has become the new bottleneck
As AI tools accelerate generation, the slowest part of the process has shifted from writing code to verifying it. Nearly all developers say they spend at least some effort reviewing, testing, and correcting AI output, and a majority describe that effort as moderate or substantial. Many respondents even report that reviewing AI-generated code can be more demanding than reviewing code written by humans.
This dynamic has given rise to what some leaders describe as “verification debt.” With AI making creation faster than comprehension, developers must now invest more time rebuilding understanding during review. The value in software engineering is therefore moving away from raw typing speed and toward confidence in deployment, testing rigor, and assurance that the code does what it claims to do.
Security, privacy, and tooling usage gaps
The survey also surfaces concerns for organizations beyond code correctness alone. More than a third of developers admit to using AI coding tools from personal accounts rather than corporate-managed ones. This behavior increases the risk of data exposure, policy violations, and inconsistent governance, especially when sensitive or proprietary code is involved.
At the same time, developers recognize both benefits and drawbacks in the AI shift. A large majority cite advantages such as better documentation support and help with test creation. However, many also highlight frustrations with code that appears correct but is wrong, or that turns out to be redundant and unnecessary, adding to maintenance overhead.
Toil is shifting, not disappearing
One of the more striking findings is that, despite the perception that AI reduces toil, the total time spent on tedious work remains largely unchanged. Developers estimate that roughly 23 to 25 percent of their time still goes into tasks like managing technical debt, debugging legacy code, and working around poor documentation.
The difference is that AI tools have shifted where this toil occurs. Instead of only wrestling with older human-written systems, engineers now also spend significant time correcting, rewriting, and validating AI-generated code. This reinforces the idea that AI is not a shortcut to skipping engineering discipline; it changes the nature of the work, but does not eliminate the need for careful review, robust testing, and strong development practices.
Read more such articles from our Newsletter here.
