AI Writes Code at Supersonic Speeds. Why Isn't Software Output Rising?

If you hand a novelist a typewriter that can type a thousand words a minute, will they publish more books? Probably not, because the bottleneck in writing isn't the physical act of typing—it's the thinking, editing, and refining.

In the world of software development, artificial intelligence has essentially handed programmers a supersonic typewriter. AI coding assistants can instantly generate massive blocks of functional code, leading to widespread speculation about an impending explosion in software output and inevitable mass layoffs in the tech sector. But a sweeping new study suggests that this narrative fundamentally misunderstands how software is actually made.

Harvard University researchers Fiona Chen and James Stratton recently analyzed an astonishingly large dataset from engineering analytics firm Jellyfish. They looked at 300 million individual work events—like code commits and pull requests—across more than 700 software development firms and 700,000 employees between 2021 and March 2026.

Their findings reveal a fascinating paradox: companies using AI coding tools are not producing significantly more software, nor are they reducing their workforce.

The reason lies in a massive shift of the productivity bottleneck. While AI can write code at blistering speeds, seasoned developers know better than to blindly trust machine-generated output. AI code can contain subtle bugs, security vulnerabilities, or logical errors that look perfectly plausible at first glance. Consequently, the time saved during the actual writing phase is entirely absorbed by the downstream verification process.

The researchers found that human code review has become a severe chokepoint. When developers use AI, the review process takes significantly longer. Pull requests are more frequently kicked back for revisions, and human reviewers are forced to leave more comments and corrections than they did before the AI era.

This dynamic offers a crucial lesson for any industry rushing to adopt generative AI. The technology is not currently replacing human workers; it is transforming their job descriptions. Developers are shifting from being primary creators who write every line from scratch to becoming meticulous editors and quality assurance inspectors.

As we integrate AI into our workflows, the ultimate takeaway is clear: generating content has never been cheaper or faster, but verifying its truth, safety, and utility remains a premium, deeply human endeavor.

Key Points

• A Harvard study analyzed 300 million work events across 700+ companies to measure AI's impact on coding.
• Despite AI's ability to generate code rapidly, overall software output at these firms did not significantly increase.
• The productivity bottleneck has shifted from code generation to human code review.
• AI-generated code requires longer review times, more revisions, and more oversight, effectively absorbing the time saved upfront.

Why It Matters

It demonstrates that the true cost of AI integration lies in verification, shifting human roles from creators to editors rather than replacing them outright.

Sources:

• AI coding agents generate more code, but not more software — Ars Technica AI