Developers have become so reliant on AI coding tools that some now refuse to work without them — even temporarily, even for research studies trying to measure the tools’ effectiveness. That refusal is itself a data point, and not a reassuring one.
The core tension:
- Speed is up. AI tools help developers produce code faster. No one disputes this.
- Quality isn’t. Researchers warn the code coming out of AI-assisted workflows may not actually be better — just more of it.
- Skill atrophy is real. Developers who won’t code without AI risk losing the ability to code without AI. The dependency becomes self-reinforcing.
- Research is harder to run. When participants refuse to put down the tool, studying its actual impact becomes nearly impossible — a methodological problem that should concern anyone citing “productivity gains.”
The Reddit thread on this story put it bluntly: “The problem with coders who won’t work without AI is that they might become incapable of coding without AI.” That’s not a hypothetical. It’s the same pattern seen with GPS navigation and spatial reasoning, or spell-check and spelling ability. Tools that remove friction also remove practice.
This matters most for junior developers. A senior engineer who leaned on manual debugging for a decade before picking up Cursor or Copilot has deep mental models to fall back on. A junior who started with AI autocomplete filling every function body may never build those models in the first place. The gap between “can prompt an AI to write code” and “can reason about why code breaks” is wide, and it’s getting wider.
NOTE
The research difficulty here is worth flagging on its own: if you can’t get developers to code without AI long enough to run a controlled study, the productivity claims from tool vendors are essentially unfalsifiable.
More code, shipped faster, with less understanding of what it does. That’s the trajectory researchers are flagging. Teams that treat AI coding tools as accelerators rather than replacements for comprehension will handle this fine. Teams that don’t will find out during their next production incident.