Developers refuse to write code without AI, despite signs of declining productivity

Software developers have become so accustomed to AI-powered coding tools that many now refuse to work without them—even for a limited number of tasks in controlled experiments. At the same time, more data shows that these systems often create more problems than they solve. Thus, the mass adoption of AI in programming is beginning to look more like a matter of perception and dependency than of real efficiency.

The productivity paradox

In February 2026, the AI research lab "METR" attempted to replicate a popular study comparing how much time developers spend on tasks with and without AI assistance. The attempt failed for an unexpected reason: participants simply refused to work without their AI tools, even for a limited number of tasks and in a strictly controlled environment.

The original METR study from 2025 had already painted a worrying picture. Programmers claimed that AI had "sped them up" by about 24%, but actual measurements showed the opposite—on average, they spent 19% more time completing tasks. The delay was due to the additional effort required for debugging, managing the AI assistant, and waiting for the generated code.

Even after encountering this slowdown in practice, participants continued to rate AI as providing a roughly 20% boost in speed—a testament to the strong divergence between subjective feeling and objective results. After the lab failed to replicate the experiment in 2026, it published a survey in which developers themselves stated that AI makes them "twice as valuable" to their organizations.

Lower quality code and more hidden costs

Independent analyses suggest that this acceleration comes at a serious cost to quality. The code review platform "CodeRabbit" examined 470 open pull requests and found that AI-generated code contained an average of about 1.7 times more issues compared to manually written code. Critical and serious defects in changes created with the help of AI occur up to 1.7 times more frequently, logical and correctness errors jump by about 75%, and security vulnerabilities increase between 1.5 and 2 times.

The startup "Entelligence AI", which specializes in reliability and monitoring, reports that companies spend about 44% of their AI tokens on fixing problems created by the AI itself—an estimate based on data from over 2,400 organizations. Researchers from the Singapore Management University reached similar conclusions in an April report, warning that "AI-generated code could lock real software projects into long-term maintenance costs."

How to manage AI dependency

Experts' recommendations boil down to a clear principle: treat code created by AI the way you would treat the code of a junior programmer. The SMU team and "Cognition" founder "Scott Wu"—creator of the AI programming agent "Devin"—are unanimous that it is critically important to maintain reliable quality control systems and carefully review every change suggested by AI.

Experts further insist that the most important tasks—such as architectural design, technology selection, security, and managing complex dependencies—should remain in the hands of experienced engineers. AI can be useful for generating boilerplate code, helper functions, and initial drafts, but not as an autonomous source of truth for the overall system design.

"You write code faster, but maintenance has no autopilot"

Programmer and author "James Shore" formulated the dilemma particularly sharply in a post that quickly became popular on "Hacker News": "Writing code twice as fast now? Better hope maintenance costs have halved too. Otherwise, you're trapped. You're trading a temporary rush of speed for eternal servitude."

This quote captures the essence of the industry's new reality: AI can create a sense of explosive productivity, but if organizations do not account for higher review, testing, and maintenance costs, they risk accumulating invisible "technical debt" that will, in the long run, erode any short-term gains in speed. The question is no longer whether to use AI, but how—and with what protective mechanisms—so the tool does not become a source of constant vulnerability.