Study shocks boardrooms: AI makes programmers slower, not faster
Dangerous hype: Why 66% of developers now distrust AI-generated code
Artificial intelligence in software development is hailed in boardrooms as the ultimate productivity miracle. But far from the euphoric board presentations, a quiet revolt is brewing within development teams. Instead of simplifying daily work, AI tools are increasingly becoming mental time-wasters. Current studies and alarming real-world reports reveal an uncomfortable truth: AI-generated code is often “almost correct,” but requires extremely time-consuming and tedious debugging. The result? Development time increases, cognitive load rises dramatically, and companies unknowingly accumulate an unmanageable amount of technical debt. So-called “vibe coding”—the thoughtless generation of code by AI—threatens to become a trillion-dollar time bomb. It’s time for an unflinching look at the reality of software development that management often refuses to acknowledge.
Productivity miracle or burnout trap? The truth about AI in software development that executives don’t want to hear
The big misunderstanding between management and the development team
Few technological developments in recent history have generated as much euphoria among corporate leaders worldwide as the use of artificial intelligence in software development. Board meetings, investor presentations, and strategy papers are full of terms like “productivity multiplier,” “competitive advantage,” and “transformative efficiency.” But while executives celebrate AI-powered coding tools as a panacea, a very different world of experience is emerging in development departments around the globe—one characterized by frustration, mental exhaustion, and growing skepticism.
This gap between expectations and reality is not a fringe phenomenon or an expression of a lack of adaptability. It is a structural problem that will prove costly for companies in the medium term. The question is no longer whether AI tools should be used in software development—this has already happened in 84 percent of all development departments—but rather how and under what conditions this can work sustainably. A sober analysis of the available data, studies, and case studies paints a picture that is significantly more complex than the prevailing narratives of progress suggest.
When enthusiasm meets resistance: The tension in practice
The Stack Overflow Developer Survey 2025, the most comprehensive survey of its kind with over 49,000 developers from 177 countries, delivers a sobering diagnosis. While the adoption rate of AI tools has increased from 76 to 84 percent year-over-year, and 51 percent of all professional developers use these tools daily, the positive sentiment toward these tools has plummeted dramatically over the same period: from over 70 percent in 2023 and 2024 to just 60 percent in 2025. The question of trust is particularly revealing: only 33 percent of developers trust the accuracy of AI output—a decrease from 43 percent the previous year—while 46 percent are actively distrustful, and only 3 percent say they “very much trust” AI results.
Experienced developers are the most skeptical: only 2.6 percent of them say they strongly trust AI outputs, while 20 percent explicitly express strong distrust of AI-generated results. This is no coincidence. Those who have designed complex systems over years, tracked down bugs in deeply nested codebases, and experienced the long-term consequences of short-sighted architectural decisions develop an institutional skepticism toward seemingly simple solutions—and this skepticism is rationally grounded, not regressive.
The deceptive allure of quickly generated code
The biggest source of frustration, identified by 66 percent of all developers as a central problem, is the tendency of AI solutions to be “almost right, but not quite.” The economic consequences of this phenomenon are more serious than they initially appear. Code that is 90 percent correct doesn’t create 90 percent added value—it may even create no value at all, because it must first be fully tested, corrected, and adapted before it can be deployed to production systems. Forty-five percent of all developers surveyed confirmed that debugging AI-generated code takes more time than writing the same code from scratch.
One consequence of this is that 42 percent of all code changes submitted to repositories are now AI-supported, but developers spend more time reviewing these changes than writing the original code. In practice, this means that while AI accelerates code production, it slows down the production of high-quality and sustainably maintainable code. Under these conditions, a productivity tool becomes a control mechanism that is extremely time-consuming.
What the numbers really say about productivity
Perhaps the most unsettling finding of recent research comes from a randomized controlled trial (RCT) conducted by the independent research institute METR between February and June 2025. Sixteen experienced open-source developers tackled 246 tasks from their own long-standing projects—with and without access to AI tools such as Cursor Pro and Claude 3.5/3.7 Sonnet. The result fundamentally contradicted the expectations of all participants: Before the study, the developers estimated that AI support would reduce processing time by 24 percent; in reality, the AI tools increased processing time by 19 percent.
This finding contradicted not only the assessments of the developers involved but also the predictions of business and machine learning experts, who had forecast time savings of 38 to 39 percent. The researchers cited the considerable time required for formulating prompts, reviewing AI output, and managing tool integration as possible explanations. Furthermore, mature codebases with strict quality standards—typical of professional enterprise environments—are particularly poorly suited for AI tools trained on generic code examples. The study does not represent a fundamental rejection of AI tools, but it clearly demonstrates that productivity gains are far from guaranteed for complex, context-dependent tasks in established codebases.
The invisible burden: Mental exhaustion and cognitive overload
Besides the measurable time component, there is a more difficult-to-quantify but no less real burden: mental exhaustion from the constant switching between formulating AI prompts, analyzing the generated results, troubleshooting, and documentation. Developers describe this state as particularly grueling because—unlike the classic flow experience in programming—it doesn’t allow for deep, focused work phases, but rather forces a fragmented mode of attention. This fragmented mode is known in cognitive science to be particularly exhausting and leads to a reduction in performance in the long run.
The consulting firm Thoughtworks coined a fitting term for this phenomenon in its Technology Radar Volume 34, published in April 2026: “cognitive debt.” This refers to the growing gap between what the code does and what developers actually understand about it. With every automatically generated code block adopted without full understanding, this gap widens—subtly, but with far-reaching consequences. Thoughtworks CTO Rachel Laycock succinctly summarized the finding: AI agents facilitate the rapid writing of code, but increasingly overwhelm developers’ understanding.
Architectural blind spots: What AI code systematically gets wrong
An in-depth analysis by Ox Security from October 2025, which examined 300 open-source projects—50 of which were wholly or partially AI-generated—identified ten recurring antipatterns in AI-generated code. The most common problems can be summarized in one sentence: AI-generated code is “highly functional, but systematically lacking in architectural judgment.” In 80 to 90 percent of cases, AI tends to implement textbook solutions instead of addressing the specific requirements of the application, avoids refactoring, and repeatedly makes the same functional errors because the model does not retain any memory of previous implementations.
Particularly problematic is the phenomenon that researcher Ana Bildea calls “code generation bloat”: Because AI doesn’t develop libraries but rather generates functionality inline over and over again, the codebase grows uncontrollably, contains many redundant blocks, and becomes increasingly difficult to maintain. Bildea aptly describes this dynamic by stating that she has observed companies going from “AI is accelerating our development” to “we can no longer deliver features because we no longer understand our own systems” in less than 18 months. GitClear provides further empirical confirmation: Between 2021 and 2024, the proportion of refactoring-related code changes fell from 25 to under 10 percent, while the proportion of copied code blocks rose from 8.3 to 12.3 percent.
