Artificial intelligence has swept through software development with astonishing speed, promising to make programmers faster and more productive. But as companies and engineers settle into daily use of AI coding tools, a growing body of evidence suggests the technology is reshaping work in more complicated—and sometimes counterproductive—ways.
A Surge in Adoption, and Rising Expectations
In the past two years, AI-powered coding assistants have moved from novelty to near-ubiquity. Developers now routinely generate large blocks of software using simple text prompts, a shift that has transformed how code is written inside technology companies both large and small.
Earlier this year, Google reported that about 90 percent of software developers across the industry were using AI tools on the job, a sharp increase from roughly 14 percent the year before. The figures underscored how quickly generative AI had become embedded in everyday engineering work—and how high expectations had climbed that the tools would deliver dramatic efficiency gains.
Yet as adoption surged, so did quiet skepticism. Engineers and managers began to notice that speed did not always translate into simplicity. Code appeared faster, but reviews took longer. Bugs seemed more frequent, not less. What was marketed as frictionless productivity increasingly came with a trade-off: more output, paired with more scrutiny.
Measuring the Hidden Costs of Machine-Written Code
A recent report by the AI software company CodeRabbit offers one of the clearest empirical looks at that trade-off. Analyzing 470 pull requests, the company found that AI-generated code produced an average of 10.83 issues per request, compared with 6.45 issues in code written by humans.
In other words, AI-assisted code generated roughly 1.7 times more problems than human-authored code. The issues were not trivial. According to the report, AI-written software showed higher rates of “critical” and “major” problems—issues serious enough to demand careful reviewer attention.
“AI coding tools dramatically increase output,” said David Loker, CodeRabbit’s AI director, “but they also introduce predictable, measurable weaknesses that organizations must actively mitigate.”
The result, the report suggested, is a shift in what developers spend their time doing: less initial writing, more correction and interpretation.
Security, Quality, and the Burden on Reviewers
Other research points in a similar direction. A September analysis by management consultants Bain & Company concluded that while programming was among the first areas to deploy generative AI at scale, the cost savings had been “unremarkable,” with results falling short of early hype.
Security concerns have been particularly acute. The security firm Apiiro found that developers who relied on AI tools produced ten times more security problems than those who did not. These included flaws related to improper password handling and other insecure practices that could expose sensitive data.
A July study by the nonprofit Model Evaluation and Threat Research went further, concluding that programmers were, in some cases, actively slowed down by AI assistance tools. The time saved in generating code was offset by the effort required to verify logic, ensure correctness, and catch subtle errors that slipped through automated suggestions.
Productivity Reconsidered in the Age of AI Assistance
The weaknesses identified were not limited to security. CodeRabbit found that AI-generated code was most prone to errors involving logic and correctness, while its biggest shortcoming lay in code quality and readability—factors that can quietly accumulate into long-term technical debt.
There were modest upsides. AI tools were effective at minimizing spelling mistakes, with humans found to be roughly twice as likely to introduce simple misspellings. But such gains paled beside the broader pattern: more code, more issues, and a heavier burden on human reviewers.
Taken together, the findings suggest that AI is not eliminating work so much as redistributing it. Developers may write less code by hand, but they are increasingly tasked with policing, refining, and securing what machines produce.
As AI becomes an enduring part of software development, the promise of effortless productivity is giving way to a more nuanced reality—one in which speed and scrutiny advance side by side
