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

In 2026, you cannot snatch AI coding tools retired of developers’ vice-grip hands, researchers person discovered.  

But portion AI is undoubtedly helping coders nutrient codification faster, it whitethorn not beryllium producing amended code, different researchers warn. And that could origin problems down the roadworthy for them. 

Specifically, successful February 2026, respected AI probe laboratory METR published a astonishing revelation: astir developers won’t work, adjacent connected a constricted fig of tasks, without AI anymore. 

METR had hoped to supply an update to immoderate groundbreaking probe published a fewer months earlier, successful 2025, connected AI coding productivity. In it, researchers measured however overmuch clip unfastened root developers took to bash tasks by manus versus with AI. 

While developers successful that survey reported that AI was making them much productive, they were shocked to larn it really slowed them down. Sure, it generated codification faster, but past they spent other clip uncovering and fixing errors, steering the AI and waiting connected it to implicit tasks. 

When METR acceptable retired to repetition the experimentation to measurement advances successful AI and coder proficiency, they couldn’t.

Devs weren’t consenting to enactment “because they bash not privation to enactment without AI” adjacent conscionable for the study, the researchers confessed. 

Instead, METR published a survey successful May that allowed method employees to self-report their AI productivity gains. Not surprisingly, they perceived that AI made them doubly arsenic invaluable to their organizations.  

But caller headlines astir the chaotic disbursal of alleged tokenmaxxing, coupled with a smattering of caller research, marque specified self-perceptions dubious.  

Tokenmaxxing, oregon utilizing the fig of tokens a idiosyncratic uses arsenic a proxy for productivity with AI, has been the inclination of 2026 truthful far. And it whitethorn already beryllium over. 

Amazon unopen down its interior token-tracking leaderboard called Kirorank aft employees were gaming it by utilizing AI agents excessively, and moving up costs, the Financial Times reported this week. The employees proved that AI usage does not automatically construe to accrued productivity.

Uber blew done its 2026 AI fund wrong the archetypal 4 months of the year, The Information reported. COO Andrew Macdonald precocious said connected a podcast that specified spending hadn’t led to a measurable increase successful projects oregon productivity. 

AI-generated codification besides doesn’t needfully trim ongoing codification attraction needs, and whitethorn adjacent summation it, programmer and writer James Shore elegantly argued successful a blog post that went viral connected Hacker News. 

“You constitute codification doubly arsenic speedy now? Better anticipation you’ve halved your attraction costs,” helium wrote. “Otherwise, you’re screwed. You’re trading a impermanent velocity boost for imperishable indenture.” 

There’s different grounds that AI tin increases codification attraction woes. 

A viral tweet from Aiswarya Sankar, laminitis and CEO of reliability engineering cause startup Entelligence AI, proclaims that companies are spending 44% of their tokens connected bug fixes that their AI generated. Code reviewing instrumentality institution Code Rabbit says it analyzed unfastened root propulsion requests and recovered that AI produced 1.7x much problems than quality code.

Those are, admittedly, self-serving stats from those trying to merchantability AI codification reviewing tools. 

Yet autarkic researchers person besides recovered specified issues. Researchers from the respected Singapore Management University published a study successful April informing that “AI-generated codification tin present semipermanent attraction costs into existent bundle projects.” 

Given that programmers emotion their AI assistants, what’s the solution?  

Well, those who privation to merchantability you AI coding agents accidental devs tin conscionable usage AI coding agents to bash the bony wearing tasks of fixing codification arsenic accelerated arsenic AI spits it out. That’s what Cognition laminitis CEO Scott Wu suggests, shaper of AI coding cause Devin.  

But adjacent helium admits that, portion Devin tin enactment independently, he’d presently complaint its accomplishment betwixt a inferior and mid-level programmer, depending connected the task. This is not a hand-it-off and hide it solution.

The SMU researchers suggest a much quality approach. Programmers should cognize what tasks AI does and doesn’t bash good arsenic profoundly arsenic they cognize their favourite coding languages. They request beardown prime assurance systems designed for AI and they are stuck with cautiously reviewing the AI’s enactment arsenic if it was a inferior dev.

Meanwhile, the researchers accidental (and Wu agrees), humans should inactive beryllium doing the big-picture enactment similar bundle architecture and information design.