In February 2026, AI research lab METR published a surprising finding: many developers are now unwilling to work on even a limited number of tasks without AI assistance.

METR had hoped to update its groundbreaking 2025 research into AI coding productivity. In that earlier study, researchers compared how long open-source developers took to complete tasks manually versus with AI support, TechCrunch report says.

The results were unexpected. Although developers believed AI was making them more productive, the data showed the opposite. While AI generated code more quickly, developers spent additional time reviewing outputs, correcting errors, steering the models, and waiting for responses. Overall, AI slowed them down rather than speeding them up.

When METR attempted to repeat the experiment to assess advances in AI capabilities and developer proficiency, it ran into an unforeseen obstacle.

Developers simply did not want to participate.

According to the researchers, participants declined because they “do not wish to work without AI” — even temporarily and solely for the purpose of the study.

Unable to conduct the original experiment, METR instead published a survey in May 2026 asking technical employees to self-report AI productivity gains. Unsurprisingly, respondents believed AI had made them roughly twice as valuable to their organisations.

Yet growing evidence suggests such perceptions may not align with reality.

Recent headlines surrounding the costly trend of “tokenmaxxing” — using AI token consumption as a proxy for productivity — have raised fresh doubts. The trend has become one of the defining workplace phenomena of 2026, but it may already be showing signs of collapse.

According to the Financial Times, Amazon recently shut down its internal token-tracking leaderboard, Kirorank, after employees began gaming the system by excessively deploying AI agents, significantly increasing costs without corresponding gains in output. The episode demonstrated that higher AI usage does not automatically translate into greater productivity.

Meanwhile, The Information reported that Uber exhausted its entire 2026 AI budget within the first four months of the year. Chief Operating Officer Andrew Macdonald later acknowledged on a podcast that the spending surge had not resulted in a measurable increase in projects completed or productivity achieved.

There are also growing concerns about the long-term consequences of AI-generated code.

Programmer and author James Shore argued in a blog post that later went viral on Hacker News that AI may accelerate coding while simultaneously increasing the maintenance burden.

“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.”

Other evidence points in the same direction.

A widely shared post by Aiswarya Sankar, founder and CEO of Entelligence AI, claimed that companies are spending 44 per cent of their AI tokens fixing bugs generated by AI itself. Meanwhile, code-review platform CodeRabbit reported that its analysis of open-source pull requests found AI-generated code introduced 1.7 times more problems than code written by humans.

These figures should be treated cautiously, as they come from companies with a commercial interest in selling AI code-review tools.

However, independent researchers have reached similar conclusions. In April 2026, researchers from Singapore Management University published a report warning that “AI-generated code can introduce long-term maintenance costs into real software projects”.

Taken together, the findings suggest that while AI may accelerate certain aspects of software development, the industry has yet to determine whether those short-term gains outweigh the long-term costs. As developers become increasingly reluctant to work without AI, the risk is that organisations could become dependent on tools whose true productivity impact remains far from settled.