New Research Reveals Developers Refuse to Code Without AI, Raising Big Concerns

Here is a surprising twist in the world of coding this year. Software developers have become so attached to their AI coding assistants that they are now refusing to do certain tasks without them. This strong reliance has come to light even as new questions emerge about whether AI tools truly make coding better or just faster.

This reliance became clear when a respected AI research lab, METR, tried to update a study from last year. That earlier study in 2025 found something unexpected. While developers felt AI made them more productive, it actually slowed them down. They generated code quicker, but then spent extra time finding and fixing errors, guiding the AI, and waiting for it to finish.

When METR attempted to repeat the experiment in February to see how AI and coders had improved, they hit a wall. Developers were unwilling to participate if it meant temporarily working without AI, even for the study. Instead, METR conducted a survey in May, where coders reported they felt twice as valuable to their companies thanks to AI.

However, recent reports and other research suggest that this feeling of boosted productivity might not be entirely accurate. One trend this year, called "tokenmaxxing," involved people trying to show their productivity by using as many AI tokens as possible. This approach proved problematic and potentially expensive.

Amazon, for instance, shut down an internal leaderboard called Kirorank that tracked AI usage after employees were found to be using AI agents excessively, driving up costs without real gains. Uber also reportedly burned through its entire 2026 AI budget in just the first four months of the year. Despite the massive spending, a top Uber executive stated there has been no clear increase in projects or overall productivity.

The story began with the rapid rise of AI coding assistants, promising to revolutionize how software is built. These tools quickly moved from experimental novelties to essential parts of many developers' daily routines. The idea was simple: let AI handle the repetitive parts of coding, freeing up human developers for more complex, creative work.

Companies rushed to integrate AI into their development pipelines, driven by the promise of faster development cycles and reduced costs. Developers, eager for efficiency, embraced these new digital helpers. This widespread adoption created the current landscape where AI tools are deeply embedded in the coding process, making it difficult for many to imagine working without them.

This development matters because it challenges the core assumption that more AI use directly equals better or cheaper software. If developers are becoming dependent on tools that might introduce new problems or hidden costs, it could have significant long-term implications for the entire tech industry. It raises questions about how we measure true productivity and the future of coding skills.

For everyday people, this could mean a few things. Software might become more expensive to develop, or projects could take longer than expected if companies struggle with hidden AI-related costs and bugs. If developers become less skilled at writing code from scratch, it might also make future software less robust or harder to maintain in the long run.

Zooming out, this situation highlights a crucial challenge in integrating AI into complex human tasks. While AI can undoubtedly speed up certain processes, it does not automatically guarantee quality or efficiency. The rush to adopt AI might be creating a new set of problems, shifting the burden from initial coding speed to long-term maintenance and debugging. This could lead to an "AI debt" where companies pay later for the shortcuts taken now.

The concerns about AI-generated code are growing. Programmer and author James Shore wisely pointed out that if AI helps you write code twice as fast, you better hope it halves your maintenance costs too. Otherwise, he says, you are merely trading a temporary speed boost for a permanent problem. Other reports back this up. For example, the founder of an AI reliability startup claimed that companies are spending 44 percent of their AI tokens just fixing bugs that AI itself created. Another company analyzing code found AI-generated code had 1.7 times more problems than human-written code. While some of these statistics come from companies selling solutions, independent researchers from Singapore Management University also warned in April that AI-generated code can indeed introduce long-term maintenance costs into real software projects.

So, what happens next in this evolving relationship between coders and AI? Companies and developers will need to find a smarter way forward. Some AI tool makers suggest using AI agents to fix the very bugs that AI generates, but even they admit these tools are often only as skilled as a junior or mid-level developer. This means it is not a "hand it off and forget it" solution.

The Singapore Management University researchers propose a more human-centered approach. They believe programmers must deeply understand what AI does well and what it does not, just as they know their favorite coding languages. Strong quality assurance systems designed specifically for AI are crucial, and developers must carefully review AI's work as if it came from a junior colleague. Most importantly, humans should continue to handle the big-picture tasks, like software architecture and security design, where their judgment remains irreplaceable.

Do you think companies should implement AI-free coding challenges or periods to ensure developers maintain their fundamental skills, or would that hurt innovation?

How can we develop better ways to measure the true productivity and long-term value of AI in software development, beyond just counting lines of code or tokens used?

---

Filed under: AICoding, DeveloperProductivity, SoftwareDevelopment, TechTrends, AIinDev