AI coding boosts output — and hidden tech debt
New analyses claim GPT‑5.2 and similar models are noticeably raising developer productivity, shifting focus toward higher‑level tasks. At the same time, a codebase analysis found AI‑generated contributions roughly doubled code duplication and reduced refactoring, creating maintenance debt that can be hard to spot. (startuphub.ai) (dev.to)
AI coding tools are speeding up software work, but the code they add is also showing more duplication and less cleanup. (openai.com) OpenAI said on December 11, 2025 that GPT‑5.2 improved software-engineering benchmark scores to 55.6% on SWE-Bench Pro and 80.0% on SWE-bench Verified, up from 50.8% and 76.3% for GPT‑5.1 Thinking. The company also said partners including JetBrains, Warp, Cognition, and Augment Code saw gains in interactive coding, code review, and bug finding. (openai.com) Broader surveys show how fast these tools have spread. Google’s 2025 DORA report drew on nearly 5,000 technology professionals, while Stack Overflow’s 2025 survey said 84% of respondents were using or planning to use AI tools in development and 51% of professional developers used them daily. (research.google) (survey.stackoverflow.co) The basic tradeoff is simple: code generators make it cheaper to produce new code than to reorganize old code. Refactoring is the maintenance work that removes repetition, splits large functions, and reuses shared components without changing what the software does. (arxiv.org) (sciencedirect.com) GitClear’s 2025 report, based on 211 million changed lines from repositories owned by Google, Microsoft, Meta, and enterprise companies from 2020 through 2024, found a spike in duplicate code blocks, more short-term churn, and a continued decline in moved lines, which it uses as a signal of code reuse. GitClear said “copy/paste” code exceeded “moved” code for the first time in its dataset. (gitclear.com) A widely shared April 2026 write-up of that research said copy-pasted code rose from 8.3% to 12.3% of all changes, refactored code fell from 25% to under 10%, and duplicated blocks of five or more lines increased eightfold in 2024. Those figures come from GitClear’s analysis, not from a peer-reviewed journal paper. (dev.to) (gitclear.com) Other research points in both directions. A November 2025 arXiv study of 15,451 refactoring instances across 12,256 pull requests found AI coding agents do refactor, but mostly through low-level edits like renaming variables and parameters rather than larger design changes. (arxiv.org) And a July 10, 2025 randomized trial from Model Evaluation and Threat Research, or METR, found 16 experienced open-source developers took 19% longer on their own repository tasks when AI tools were allowed. The study used 246 real issues and mostly involved Cursor Pro with Claude 3.5 and 3.7 Sonnet, which the authors said made it a snapshot of early-2025 tools rather than a verdict on newer systems. (metr.org) Developers themselves sound more enthusiastic about access than about accuracy. Stack Overflow’s 2025 survey found 46% of respondents distrusted AI output, compared with 33% who trusted it, and only 3.1% said they “highly trust” the results. (survey.stackoverflow.co) That leaves teams measuring two different things at once: how quickly code gets written this week, and how much duplicated, harder-to-maintain code stays in the repository next year. The output gains are getting easier to see; the maintenance bill still mostly shows up later. (research.google) (gitclear.com)
