AI-Addicted Developers: A Trap for Code Quality?

Developers and Their Growing Dependence on AI

In 2026, a significant trend is emerging among developers: they no longer want to work without AI-based coding tools. This dependence, highlighted by the METR research lab, raises questions about the impact of AI on code quality.

In February 2026, METR published a surprising revelation: most developers no longer wish to work, even on a limited number of tasks, without AI. This finding continues from a previous study in 2025, where METR measured the productivity of developers using AI compared to those working manually. The results showed that while AI allows for faster code generation, it actually slows down the overall process due to the time spent correcting errors and guiding the AI.

The Illusion of Increased Productivity

Although AI seems to accelerate the coding process, a 2025 study reveals a more nuanced reality. Developers, while feeling more productive, actually spend more time correcting errors generated by AI. METR attempted to replicate this study, but developers refused to participate without AI, illustrating their reluctance to return to traditional methods.

In May, METR published a survey allowing technical employees to report their productivity gains related to AI. Unsurprisingly, they perceived that AI made them twice as valuable to their organizations. However, this perception is challenged by recent research and events in the industry.

High Costs and Questionable Productivity

Companies like Amazon and Uber have found that intensive use of AI, measured by the phenomenon of "tokenmaxxing," has not led to an increase in productivity. Amazon had to shut down its internal ranking system, Kirorank, after employees abused AI, resulting in exorbitant costs. Similarly, Uber exhausted its AI budget in just four months, with no tangible benefits. Andrew Macdonald, COO of Uber, stated in a podcast that these expenditures did not lead to a measurable increase in productivity.

The phenomenon of "tokenmaxxing" is based on the idea that the number of tokens used by a person can serve as an indicator of productivity. However, this approach has shown its limits, particularly at Amazon, where employees manipulated the system to artificially inflate their apparent productivity, leading to high costs without real benefits.

The Challenges of Code Maintenance

AI-generated code presents challenges in terms of maintenance. James Shore, a programmer and author, emphasized that the speed of code generation by AI does not compensate for the increased maintenance costs. A viral tweet from Aiswarya Sankar, founder and CEO of the startup Entelligence AI, proclaims that companies spend 44% of their tokens on bug fixes generated by their AI. Additionally, the company CodeRabbit reported that AI produces 1.7 times more problems than human-written code.

These statistics, while self-interested, are corroborated by independent researchers. The Singapore Management University published a report in April warning that AI-generated code can introduce long-term maintenance costs in real software projects.

Towards a More Judicious Use of AI

In light of these challenges, some experts propose solutions. Scott Wu, founder of Cognition, suggests using AI agents to correct code as quickly as it is generated. However, he admits that these agents do not replace human skills for complex tasks. Researchers from SMU recommend a balanced approach, where developers understand the capabilities and limitations of AI while retaining responsibility for critical tasks such as software architecture and security.

Programmers should know the tasks that AI does and does not perform well as deeply as they know their preferred coding languages. They need robust quality assurance systems designed for AI and should spend time carefully reviewing AI's work as if it were a junior developer. Meanwhile, researchers assert (and Wu agrees) that humans should always handle large-scale work such as software architecture and security design.