Here’s the pitch every developer has heard by now: AI will write your code, slash your workload, and free you up for the interesting stuff. The reality? Developers using AI coding assistants are actually working longer hours, not shorter ones. And the reasons why should concern every engineering leader banking on AI-driven productivity gains.
A new analysis from Scientific American lays out the emerging evidence that AI coding tools — GitHub Copilot, ChatGPT, and their competitors — aren’t delivering the time savings their marketing promises. The data paints a more complicated picture than the one vendors are selling.
The productivity paradox is real, and the numbers are stark
The centerpiece finding comes from a study published in February 2025 by researchers at METR (Model Evaluation & Threat Research), an AI safety nonprofit. They ran a randomized controlled trial with experienced open-source developers working on real repositories they already knew well. The result: developers using AI tools were actually 19 percent slower at completing tasks than those working without them.
Nineteen percent slower. Not faster. Slower.
The developers themselves predicted they’d be 24 percent faster with AI assistance. That’s a massive perception gap. They felt more productive while measurably being less so. This disconnect between perceived and actual productivity is one of the most important findings in the study, and it should give pause to any organization measuring AI’s impact through developer surveys alone.
But the METR study isn’t the only signal. Research from Uplevel, a developer analytics company, tracked hundreds of engineers across multiple companies before and after GitHub Copilot adoption. Their findings, reported in late 2024, showed no statistically significant change in the number of pull requests or the time to merge them. What did change? Bug rates went up by 41 percent. And developers reported more burnout.
More bugs. More burnout. Same output. That’s not the ROI pitch anyone signed up for.
Why AI tools add hours instead of subtracting them
So what’s actually happening when developers use these tools? Several compounding factors explain the paradox.
First, there’s the review tax. AI-generated code still needs to be read, understood, tested, and debugged by a human. This isn’t trivial. Reading someone else’s code — even when that someone is an LLM — takes real cognitive effort. Developers spend time verifying that AI suggestions are correct, don’t introduce security vulnerabilities, and actually fit the architecture of the project. The METR researchers noted that much of the time overhead came from prompt crafting, waiting for responses, and then reviewing output that frequently missed the mark on complex tasks.
Second, AI tools are genuinely good at generating boilerplate and simple functions. This creates a seductive loop. Developers start relying on the tool for easy stuff, then keep reaching for it on harder problems where it’s less reliable. The tool becomes a default rather than a selective aid. Time that would have been spent just writing the code gets spent in a back-and-forth with a chatbot that doesn’t understand the deeper context of the codebase.
Third — and this is the one that should worry managers most — AI tools appear to be expanding the scope of what developers attempt. When generating code feels cheap, you write more of it. More features, more refactors, more experiments. The work expands to fill the perceived capacity. Scientific American points to this as a version of the Jevons paradox: when a resource becomes more efficient to use, people don’t use less of it. They use more.
This is exactly what happened with email, spreadsheets, and every other productivity tool before AI. The work just grows.
There’s also an organizational dynamic at play. When leadership sees AI as a force multiplier, expectations ratchet up. Deadlines tighten. Headcount gets questioned. Developers feel pressure to produce more, and AI becomes both the justification for higher targets and an imperfect tool for meeting them. The result is longer hours, not shorter ones.
What this means for engineering organizations right now
None of this means AI coding tools are useless. They clearly help with specific tasks — autocomplete, boilerplate generation, documentation, translating between languages, and rubber-ducking ideas. Many developers genuinely like using them. But liking a tool and being measurably more productive with it are different things.
The gap between marketing claims and measured outcomes is significant. GitHub has cited internal data suggesting Copilot helps developers complete tasks 55 percent faster. But that figure comes from controlled lab settings with isolated tasks, not from the messy reality of maintaining large codebases with complex dependencies and institutional knowledge baked in. The METR study specifically tested experienced developers on repos they knew — a much harder and more realistic benchmark — and found the opposite result.
For engineering leaders, the takeaway is straightforward: don’t set staffing levels or sprint expectations based on assumed AI productivity gains that haven’t been validated with your own teams on your own codebases. Measure actual output. Track bug rates, cycle times, and developer well-being alongside adoption metrics. If your developers say they feel faster but your data shows otherwise, trust the data.
And be honest about what you’re buying. AI coding assistants at current pricing — Copilot runs $19/month per developer for business plans — may still be worth it for developer satisfaction and specific workflow improvements. But framing them as a replacement for headcount or a justification for compressed timelines is premature at best and counterproductive at worst.
The broader industry conversation is starting to catch up. On X, developers have been sharing their own experiences that mirror the research findings, with many noting that AI tools help most on greenfield projects and least on established, complex codebases. Some senior engineers report spending more time cleaning up junior developers’ AI-generated code than they would have spent mentoring them through writing it themselves.
The pattern here isn’t new. Every wave of developer tooling — from IDEs to Stack Overflow to low-code platforms — has been accompanied by inflated productivity claims that eventually settle into a more nuanced reality. AI coding tools are following the same arc, just with louder hype and bigger price tags.
The tools will get better. Models will improve at understanding context, maintaining coherence across large codebases, and producing reliable output. But right now, in mid-2025, the evidence says they’re adding complexity as often as they’re removing it. And developers are paying for that complexity with their time.
WebProNews is an iEntry Publication