Coders are refusing to work without AI – and this could come back to bite them

In 2026, you won’t be able to outsource AI writing tools from vise makers, researchers have found.

But while AI is undoubtedly helping coders produce code faster, it may not be creating better code, some researchers warn. And this can cause problems for them.

In particular, in February 2026, the prestigious AI METR lab published a startling revelation: Most developers can’t work, even for small tasks, without AI anymore.

METR hopes to provide updates to others interesting research published A few months earlier, in 2025, in the development of AI coding. In it, researchers measured how much time open source developers took to work manually versus AI.

Although the researchers reported that AI was making them more productive, they were surprised to learn that it was slowing them down. Sure, it produced the code faster, but spent more time finding and fixing bugs, controlling the AI ​​and waiting for it to complete a task.

When METR started repeating the experiment to test the progress of AI and coder skills, they couldn’t.

Devs refused to participate “because they don’t want to work without AI” even for research purposes, researchers admitted.

Instead, METR published a study in May that allowed tech workers to show off their AI inventions for themselves. Not surprisingly, they realized that AI made them twice as valuable to their organizations.

But the recent chapters of wild currency called tokenmaxxingalong with a lot of recent research, it makes the self-esteem questionable.

Tokenmaxxing, or using the amount of tokens that one uses as a proxy to create productivity with AI, has been running in 2026 until now. And maybe it’s already over.

Amazon shut down its internal benchmarks scoreboard called Kirorank after employees overplayed the AI ​​game, costing it money. The Financial Times reported this week. The workers proved that using AI does not only translate to higher productivity.

Uber passed its 2026 AI budget within the first four months of the year, The Information report. COO Andrew Macdonald recently said on a podcast that spending has not produced measurable growth in work or productivity.

AI-generated code not only reduces the need for code development, but can increase it, programmer and author James Shore clearly argued. blog post which went viral on Hacker News.

He wrote: “Do you write double digits now? “Otherwise, you’re wrong. You’re trading a temporary interest rate for stability.”

There is also some evidence that AI can increase the complexity of coding.

A viral tweets from Aiswarya Sankar, founder and CEO of trusted engineering startup Entelligence AI, announces that companies are spending 44% of their tokens on correcting the mistakes their AI made. Currently, a company that develops code analysis tools What a rabbit says it analyzed open source pull requests and found that AI produced 1.7x more problems than human code.

These are, admittedly, self-made statistics from those trying to sell AI code analysis tools.

However, independent researchers have also found such stories. Researchers from the prestigious Singapore Management University published a report in April warning that “AI-generated code can lead to long-term maintenance costs in real-world applications.”

Since developers love their AI assistants, what is the answer?

Well, those who want to sell you AI coding agents say that devs can use AI coding agents to do the tedious work of coding as fast as the AI ​​can spit it out. Here’s what Cognition founder and CEO Scott Wu – an AI coding assistant Devin – said it shows.

But even he admits that, although Devin can work independently, he can test his skills among small and medium-sized software developers, depending on the task. This is not just throwing up your hands and forgetting the answer.

SMU researchers suggest a more human approach. Programmers need to know what AI does and doesn’t do as well as they know their favorite languages. They need a quality assurance system for AI and they look closely at AI projects as if they were a junior dev.

Meanwhile, the researchers say (and Wu agrees), people should still be doing the big picture like software architecture and security architecture.