AI's Impact on Developer Productivity: Challenges and Solutions

In 2026, taking AI tools out of developers' hands is almost impossible. Research by the METR lab revealed that most programmers refuse to work without AI, even for limited tasks. The issue is that although AI helps speed up code production, it does not necessarily improve the quality of that code. And this could bring problems down the road.

METR, in February 2026, attempted to update a 2025 study on AI productivity. At the time, it measured how long open-source developers took to complete tasks manually versus with AI. The surprise? Instead of speeding things up, AI was slowing down the process. Despite generating code quickly, programmers spent time fixing errors and waiting for AI to complete tasks. When METR tried to repeat the experiment, developers refused to participate, claiming they did not want to work without AI, not even for a study.

Instead, METR published a survey in May, where technical employees reported their perceptions of productivity gains with AI. They believed AI made them twice as valuable to their organizations. But is this perception real? Recent news about the cost of so-called "tokenmaxxing" (using token count as a productivity metric) suggests it isn't. Amazon, for example, shut down its internal token leaderboard, Kirorank, after realizing employees were gaming the system, generating high costs without any real productivity increase.

Uber also blew past its AI budget in just four months without a measurable increase in projects or productivity. James Shore, a programmer and author, highlighted in a post that went viral: "Are you writing code twice as fast now? Better hope your maintenance costs have been cut in half. Otherwise, you're in trouble."

Additionally, AI can increase code maintenance issues. Aiswarya Sankar, CEO of Entelligence AI, tweeted that companies spend 44% of their tokens fixing bugs generated by the AI itself. Code Rabbit, for its part, analyzed open-source pull requests and found that AI generated 1.7 times more issues than human code. Of course, these are statistics from companies selling AI code review tools, but independent researchers have found similar problems.

So what is the solution? Some, like Cognition's Scott Wu, suggest using AI coding agents to fix code quickly. But even he admits that, currently, these agents perform like junior-to-mid-level programmers. Meanwhile, researchers from Singapore Management University recommend a more human-centric approach: programmers must understand well what AI does or does not do well, just as they know their favorite programming languages. Additionally, strong quality assurance systems and careful reviews of AI work are essential.

In the end, high-level work such as software architecture and security design should still be done by humans. After all, as much as we love our AI assistants, they still do not replace human intuition and judgment.