Developers Refused to Work Without AI, and a Study Exposed the Dependency Paradox

In February 2026, research lab METR ran into an unexpected problem. It tried to replicate a 2025 study that measured how AI affects programmers' speed. The experiment never happened: developers flatly refused to participate without their usual AI tools, even for a controlled lab test.

What happened in the lab

The original 2025 study produced a significant result: developers work 30–50% faster with AI assistants such as GitHub Copilot and ChatGPT. It was a major contribution to understanding how AI affects productivity and workflows in IT.

A year later, METR decided to rerun the same experiment to see whether anything had changed. But when the lab invited programmers to take part, something unexpected happened. Developers refused to write code without their usual AI tools, even for a short test under controlled conditions. This wasn't a polite brush-off; people simply didn't want to switch off Copilot, ChatGPT, and their other assistants.

Why people refused

In a single year, AI went from a convenient tool to a necessity in day-to-day work. Developers had grown fully accustomed to constant autocomplete, smart suggestions, and debugging assistants. They had lost the habit of working the old way, without that digital backdrop.

Participants also explained their refusal simply: without AI, their productivity feels lower and the work feels incomplete. It's a classic case of technological dependency, where a tool becomes so embedded in a process that working without it feels impossible.

• Used to constant autocomplete and code suggestions
• Without AI, work feels slow and inefficient
• Have lost the habit of solving problems on their own
• Fear a drop in productivity, even for a short period

The dependency paradox

The failure to replicate the study is a significant finding in itself. It shows that within a single year, AI became so deeply integrated into development that developers simply won't agree to work without it, even for scientific purposes.

But it raises a troubling question: if people are completely dependent on AI and can't (or won't) work without it, does that mean AI is actually making them better? Or is it just habit? And most importantly, what happens to developers' skills if they rely on AI for 80–90% of their work?

"We couldn't conduct the study because people simply refused to

participate without their AI tools," a METR representative said.

What this means

The paradox is that AI tools built to boost productivity have created total dependency. Developers can't (or won't) work without them, but that doesn't necessarily make them better developers. It may mean they've learned to lean on AI rather than getting better at solving problems on their own.