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

By 2026, you won’t be able to wrest AI coding tools from developers’ grasp, researchers have found.

But while AI undoubtedly helps programmers produce code faster, it may not produce better code, other researchers warn. And that can cause problems for them in the long term.

Specifically, in February 2026, respected AI research lab METR published a surprising revelation: Most developers will no longer be able to work without AI, even on a limited number of tasks.

METR had hoped to provide an update for some groundbreaking research published a few months earlier, in 2025, on the productivity of AI coding. In it, researchers measured how much time open source developers needed to complete tasks by hand versus with AI.

Although developers in that study reported that AI made them more productive, they were shocked to discover that it actually slowed them down. Sure, it generated code faster, but then they spent extra time debugging and fixing bugs, directing the AI, and waiting for it to complete tasks.

When METR wanted to repeat the experiment to measure progress in AI and coding skills, they couldn’t.

Developers were unwilling to participate “because they don’t want to work without AI,” even just for the study, the researchers admitted.

Instead, METR published a survey in May, tech workers were able to self-report their AI productivity gains. It’s no surprise that they realized AI made them twice as valuable to their organizations.

But recent headlines about the wild costs of so-called token maxxing, coupled with a bit of recent research, make such self-perceptions questionable.

Tokenmaxxing, or using the number of tokens a person uses as a measure of productivity with AI, is the trend of 2026 so far. And it may already be over.

Amazon has shut down its internal token-tracking leaderboard, called Kirorank, after employees were gaming it by overusing AI agents and driving up costs. The Financial Times reports this this week. The employees proved that AI use does not automatically translate into higher productivity.

Uber burned through its 2026 AI budget within the first four months of the year. The information reported. COO Andrew Macdonald recently said on a podcast that something like that expenditure had not led to a measurable increase in projects or productivity.

AI-generated code also doesn’t necessarily reduce ongoing code maintenance needs and may even increase them, programmer and author James Shore elegantly argued in a blog post that went viral on Hacker News.

“You now write code twice as fast? I just hope you’ve halved your maintenance costs,” he wrote. “Otherwise you’re screwed. You’re trading a temporary speed boost for a permanent contract.”

There is even more evidence that AI may increase code maintenance problems.

A viral tweet by Aiswarya Sankar, founder and CEO of startup Entelligence AI, claims that companies spend 44% of their tokens on bug fixes generated by their AI. Meanwhile, code review tools company CodeRabbit says it analyzed open source pull requests and found that AI caused 1.7x more problems than human code.

Those are, admittedly, self-serving statistics from those trying to sell AI code review tools.

Yet independent researchers have also discovered such problems. Researchers from the respected Singapore Management University published a report in April warning that “AI-generated code can introduce long-term maintenance costs into real software projects.”

Considering that programmers love their AI assistants, what’s the solution?

Well, those who want to sell you AI coding agents say that developers can simply use AI coding agents to perform the tiresome tasks of fixing code as fast as AI spits it out. That’s what Cognition founder and CEO Scott Wu – the creator of AI coding agent Devin – suggests.

But even he admits that while Devin can work independently, he currently rates his skills between a junior and a mid-level programmer, depending on the task. This is not a ‘hand-it-off and forget-it solution’.

The SMU researchers suggest a more human approach. Programmers need to know what tasks AI does and doesn’t do as much as they know their favorite coding languages. They need strong quality assurance systems designed for AI, and they are stuck carefully reviewing the AI’s work as if it were a junior developer.

Meanwhile, the researchers say (and Wu agrees) that humans still need to do the big work, like software architecture and security design.