In 2026, you can not pry AI coding instruments out of builders’ vise grip, researchers have found.
However whereas AI is undoubtedly serving to coders produce code sooner, it might not be producing higher code, different researchers warn. And that would trigger issues down the highway for them.
Particularly, in February 2026, revered AI analysis lab METR published a surprising revelation: most builders received’t work, even on a restricted variety of duties, with out AI anymore.
METR had hoped to offer an replace to some groundbreaking research published a number of months earlier, in 2025, on AI coding productiveness. In it, researchers measured how a lot time open supply builders took to do duties by hand versus with AI.
Whereas builders in that research reported that AI was making them extra productive, they had been shocked to be taught it really slowed them down. Positive, it generated code sooner, however then they spent additional time discovering and fixing errors, steering the AI and ready on it to finish duties.
When METR got down to repeat the experiment to measure advances in AI and coder proficiency, they couldn’t.
Devs weren’t keen to take part “as a result of they don’t want to work with out AI” even only for the research, the researchers confessed.
As an alternative, METR published a survey in Might that allowed technical workers to self-report their AI productiveness beneficial properties. Not surprisingly, they perceived that AI made them twice as helpful to their organizations.
However latest headlines about the wild expense of so-called tokenmaxxing, coupled with a smattering of latest analysis, make such self-perceptions doubtful.
Tokenmaxxing, or utilizing the variety of tokens an individual makes use of as a proxy for productiveness with AI, has been the pattern of 2026 to this point. And it might already be over.
Amazon shut down its inner token-tracking leaderboard referred to as Kirorank after workers had been gaming it through the use of AI brokers excessively, and working up prices, the Financial Times reported this week. The staff proved that AI use doesn’t robotically translate to elevated productiveness.
Uber blew by its 2026 AI price range throughout the first 4 months of the 12 months, The Information reported. COO Andrew Macdonald just lately stated on a podcast that such spending hadn’t led to a measurable increase in initiatives or productiveness.
AI-generated code additionally doesn’t essentially scale back ongoing code upkeep wants, and will even improve it, programmer and creator James Shore elegantly argued in a blog post that went viral on Hacker Information.
“You write code twice as fast now? Higher hope you’ve halved your upkeep prices,” he wrote. “In any other case, you’re screwed. You’re buying and selling a brief pace increase for everlasting indenture.”
There’s different proof that AI can improve code upkeep woes.
A viral tweet from Aiswarya Sankar, founder and CEO of reliability engineering agent startup Entelligence AI, proclaims that corporations are spending 44% of their tokens on bug fixes that their AI generated. In the meantime, code-reviewing instrument firm Code Rabbit says it analyzed open supply pull requests and located that AI produced 1.7x extra issues than human code.
These are, admittedly, self-serving stats from these attempting to promote AI code reviewing instruments.
But impartial researchers have additionally discovered such points. Researchers from the revered Singapore Administration College published a report in April warning that “AI-generated code can introduce long-term upkeep prices into actual software program initiatives.”
On condition that programmers love their AI assistants, what’s the answer?
Properly, those that need to promote you AI coding brokers say devs can simply use AI coding brokers to do the bone-wearying duties of fixing code as quick as AI spits it out. That’s what Cognition founder and CEO Scott Wu —the maker of AI coding agent Devin — suggests.
However even he admits that, whereas Devin can work independently, he’d at present fee its ability between a junior and mid-level programmer, relying on the duty. This isn’t a hand-it-off and neglect it answer.
The SMU researchers counsel a extra human method. Programmers ought to know what duties AI does and doesn’t do nicely as deeply as they know their favourite coding languages. They want sturdy high quality assurance techniques designed for AI and they’re caught with fastidiously reviewing the AI’s work as if it had been a junior dev.
In the meantime, the researchers say (and Wu agrees), people ought to nonetheless be doing the big-picture work like software program structure and safety design.
If you buy by hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.
