New research June 2026 · 623M changes analyzed

Write-only mode: refactoring is down 70% while code duplication is up 81%.

Four years of commit data show AI-era throughput climbing while the structural habits that keep codebases maintainable — reuse, refactoring, legacy upkeep — erode across every signal we track.

−74%
Long-term legacy maintenance
+47%
Error-masking constructs
+41%
Within-commit copy/paste
Read the paper
AI Code Quality Index / GitClear Maintainability Gap Research

How AI is reshaping the code we ship

Code quality signal metrics, tracked weekly across the largest dataset of LLM-attributed commits in the industry. We compare every major code-generating model against the all-developer baseline.

Updated monthly · Q3 2026 · methodology

Code lines evaluated
821.2 M
since Dec 2024
AI-attributed lines
6.2 M
all providers
Developers observed
263,286
in commit window
Models fingerprinted
209
Claude, GPT, Gemini, +206
Metric 01 free · public

Obfuscations per 1,000 lines changed

Defensive idioms like rescue, safe-navigation operators, null checks, and mock guards. Higher generally signals more cautious code.

Metric 02 free · public

Duplicate lines per 1,000 changed

Block-level duplication detected within the same commit or across the recent codebase. A leading indicator of maintenance debt.

Metric 03 free · public

AI-assisted commit percent

Percent of commits with evidence of AI authorship (commit author, AI API stats, etc).

Metric 04 gated · subscribers

Copy/paste percent

Fingerprint-detected reuse — text either repositioned without modification or duplicated from elsewhere in the repo.

Metric 05 gated · subscribers

Moved lines percent

Share of authored lines that were moved. A proxy for refactoring activity.

Metric 06 gated · subscribers

Method calls per 1,000 lines changed

Function invocations within newly authored code. Higher indicates greater connectivity between new and existing code.

Competitive landscape

Could GitClear improve your AI ROI measurement?

As AI reshapes what teams ship each quarter, measuring code quality and velocity has never mattered more — and the analytics market is crowded with capable tools. Each cell below names what a competitor does well, then how GitClear answers in the same lane. Click through for the full side-by-side.

See it on your own repos No credit card. Connect a repo and watch the metrics populate in minutes.
Metric 08 gated · subscribers

Mature-code update percent

Share of changed lines that updated code older than one year. Higher signals more changes to load-bearing, long-lived code.

Metric 09 gated · subscribers

Churn percent

Share of authored lines that were subsequently rewritten or removed within two weeks. A proxy for defect or rework rate.

Metric 10 gated · subscribers

Throwaway percent

Share of authored lines that were rewritten or removed within one week

Methodology

Since 2020, GitClear has been iterating on a code change interpretation engine (Commit Cruncher) that weaves together the tapestry of changes that a line undergoes over time. Commit Cruncher keeps lines connected from initial inception, through possible file moves & find/replaces, up to its current-day content. By labeling all the human-recognized code change idioms, we can deduce the share of consequential code changes (which constitute around 3% of the "Lines change count" reported by conventional git stats) relative to the count of other emergent quantities of interest.

Each signal above is computed weekly across hundreds of millions of changed lines, segmented by the LLM fingerprint we attribute to the commit (Claude, GPT, Gemini, Copilot, Cursor, Augment, Grok, and a human-or-unattributed baseline). Rates are normalized per 1,000 changed lines or expressed as a percent of authored lines so that providers with different commit volumes remain comparable.

For the full methodology — fingerprinting heuristics, attribution confidence thresholds, normalization formulas, the dataset's repo composition, and how each signal is defined and validated — see The Maintainability Gap, our research paper detailing the four-year longitudinal study behind these charts.

GitClear's past AI research has been analyzed and written about by industry-leading journalists and analysts, including MIT Technology Review.