原文 · 未翻译
In 2026, you cannot pry AI coding tools out of developers’ vise grip, researchers have discovered.
But while AI is undoubtedly helping coders produce code faster, it may not be producing better code, other researchers warn. And that could cause problems down the road for them.
Specifically, in February 2026, respected AI research lab METR published a surprising revelation: Most developers won’t work, even on a limited number of tasks, without AI anymore.
METR had hoped to provide an update to some groundbreaking research published a few months earlier, in 2025, on AI coding productivity. In it, researchers measured how much time open source developers took to do tasks by hand versus with AI.
While developers in that study reported that AI was making them more productive, they were shocked to learn it actually slowed them down. Sure, it generated code faster, but then they spent extra time finding and fixing errors, steering the AI and waiting on it to complete tasks.
When METR set out to repeat the experiment to measure advances in AI and coder proficiency, they couldn’t.
Devs weren’t willing to participate “because they do not wish to work without AI” even just for the study, the researchers confessed.
Instead, METR published a survey in May that allowed technical employees to self-report their AI productivity gains. Not surprisingly, they perceived that AI made them twice as valuable to their organizations.
But recent headlines about the wild expense of so-called tokenmaxxing, coupled with a smattering of recent research, make such self-perceptions dubious.
Tokenmaxxing, or using the number of tokens a person uses as a proxy for productivity with AI, has been the trend of 2026 so far. And it may already be over.
Amazon shut down its internal token-tracking leaderboard called Kirorank after employees were gaming it by using AI agents excessively, and running up costs, the Financial Times reported this week. The employees proved that AI use does not automatically translate to increased productivity.
Uber blew 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 such spending hadn’t 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 it, programmer and author James Shore elegantly argued in a blog post that went viral on Hacker News.
“You write code twice as quick now? Better hope you’ve halved your maintenance costs,” he wrote. “Otherwise, you’re screwed. You’re trading a temporary speed boost for permanent indenture.”
There’s other evidence that AI can increase code maintenance woes.
A viral tweet from Aiswarya Sankar, founder and CEO of reliability engineering agent startup Entelligence AI, proclaims that companies are spending 44% of their tokens on bug fixes that their AI generated. Meanwhile, code-reviewing tool company CodeRabbit says it analyzed open source pull requests and found that AI produced 1.7x more problems than human code.
Those are, admittedly, self-serving stats from those trying to sell AI code reviewing tools.
Yet independent researchers have also found such issues. 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.”
Given that programmers love their AI assistants, what’s the solution?
Well, those who want to sell you AI coding agents say devs can just use AI coding agents to do the bone-wearying tasks of fixing code as fast as AI spits it out. That’s what Cognition founder and CEO Scott Wu —the maker of AI coding agent Devin — suggests.
But even he admits that, while Devin can work independently, he’d currently rate its skill between a junior and 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 should know what tasks AI does and doesn’t do well as deeply as they know their favorite coding languages. They need strong quality assurance systems designed for AI and they are stuck with carefully reviewing the AI’s work as if it were a junior dev.
Meanwhile, the researchers say (and Wu agrees), humans should still be doing the big-picture work like software architecture and security design.
Topics