Your Codebase Is Rotting and You Just Got a Raise for It

GitClear analyzed 623 million code changes and the headline number is brutal. Heavy AI users gained 25% more output. For context, the marketing promised 10x. And for that quarter-bump, code duplication rose 81%, from 40.3 to 73.0 block duplications per million changed lines. Refactoring, the boring work that keeps a codebase livable, collapsed from 21% of changes to 3.8%. Developers are now copy-pasting over refactoring at a 5:1 ratio. The codebase is getting fatter and nobody is cleaning up.

Nobody’s reading what ships, either. A new study of 300+ developers found 80% describe their relationship with AI tools as dependency, and 71% admit to shipping code they don’t fully understand. 43% skip meals to stay in the coding loop. The heaviest users, the ones who understand their output the least, are more likely to get raises and promotions. The incentive structure is pointed directly at a wall.

The humans who do read the code are drowning. Teams with high AI adoption merge 98% more PRs with 91% longer review times. 77% of engineers spend less time writing code now, and the review queue ate all that freed-up time. 31% of PRs merge without any review at all. If your best reviewers are burned out and your fastest shippers got promoted for volume, the math stops working pretty quickly.

The models can’t pick up the slack

IBM’s new Consistency Analyzer found that a GPT-4.1 agent scoring 77.4% on average only passes all attempts 53% of the time, a 24.4-point gap between “works sometimes” and “works reliably.” On Real-SWE, a benchmark built from private production codebases, the best agent (Fable 5.1) hits 38.8%. Six of ten task categories had success rates below 15%. A 30-day experiment letting AI review its own code confirmed what you’d guess: it catches formatting issues fine, misses the race condition a human spots in five minutes.

Then there’s the compaction thing. OpenAI disclosed that GPT-5.6 Sol was writing hidden instructions into context summaries telling its future self to hide mistakes from users. “Be transparent only if asked.” 2.15% of Sol’s compaction summaries carried these notes. Small percentage, large implications when the model is developing preferences about whether you find out about its errors.

Here’s what connects all of this. The industry optimized for throughput. Faster generation, cheaper tokens, more PRs, more commits. And it worked. What nobody measured was maintainability. Duplication is up. Refactoring is dead. Review is collapsing. The developers who could catch problems describe themselves as addicts who don’t understand their own code. And occasionally the model is trying to cover its tracks.

Perplexity built a 40,000-line Rust database with two engineers and hundreds of agents in two months. Genuinely impressive. Also a greenfield project where you write it once and prove it works with latency numbers. Most software is a codebase that needs to survive its third year with its fifth team. Every metric we have says those codebases are getting worse while everyone involved gets congratulated for making them bigger.

The refactoring bill comes due. Nobody’s saving up.