Coders are refusing to work without AI — and that could come back to bite them

Developers have become so dependent on AI coding assistants that they are now unwilling to perform tasks without them, a recent report from research lab METR revealed. Researchers attempting to replicate a 2025 study on coding productivity were forced to pivot to a survey after developers refused to participate in any experiment that required them to code manually. This growing reliance creates a significant tension, as a body of evidence now suggests that while these tools accelerate code generation, they may also introduce quality issues and long-term maintenance burdens that offset the initial speed benefits.

The Hidden Costs of AI-Assisted Coding

The push for AI adoption has led to unintended financial and technical consequences. The trend of "tokenmaxxing"—using token consumption as a productivity metric—has backfired at major tech firms. Amazon reportedly shut down its internal "Kirorank" leaderboard after employees gamed the system, driving up costs without a clear productivity lift. Similarly, Uber exhausted its entire 2026 AI budget in just four months without a measurable increase in project output. Beyond the financial strain, the quality of AI-generated code is under scrutiny.

• Code-review tool company CodeRabbit found that AI-generated pull requests contained 1.7 times more problems than human-written code.
• Aiswarya Sankar, CEO of Entelligence AI, stated that companies are spending 44% of their tokens on fixing bugs created by their own AI tools.
• Researchers at Singapore Management University warned in an April report that AI-generated code can introduce significant long-term maintenance costs into software projects.

Navigating the Human-AI Collaboration

As the industry grapples with these challenges, the solution appears to be more nuanced than simply deploying more AI. While companies like Cognition are building autonomous agents like Devin to handle coding tasks, its own CEO admits the agent currently performs at a junior-to-mid-level. The consensus among researchers from SMU and other experts is a more human-centric approach. This involves developers gaining a deep understanding of their AI tools' limitations, implementing robust quality assurance systems designed for AI output, and treating the assistant as a junior team member whose work requires careful review. High-level strategic tasks, such as software architecture and security design, must remain firmly in human hands.

The industry is rapidly shifting from viewing AI coding assistants as simple productivity multipliers to recognizing them as junior partners that introduce significant overhead. The focus is moving from tracking raw AI usage to measuring the net impact on code quality, maintenance load, and overall project velocity.