How have software-development commit trends shifted across repository creation,…

Research Question

What do high-quality longitudinal studies (2019–2026) show about directional shifts and current baseline ranges for repository creation rate, Lines of Code (LOC) velocity, rework share, project abandonment, Artificial Intelligence (AI) slop indicators, useless-test prevalence, and unshipped-project rates?

Findings

Executive Summary

Across seven software-delivery metrics tracked from 2019 to 2025, raw volume indicators rose sharply while quality-composition indicators degraded, and project-to-production conversion rates remained persistently low. Repository creation on GitHub grew from approximately 44 million per year (2019) to over 121 million per year (2024), and commit volume reached nearly 1 billion in 2024 at 25% year-over-year growth, but this volume surge coincides with measurable quality erosion: short-cycle code churn rose 84% relative and copy-paste code share rose 48% within the GitClear dataset between 2020 and 2024. The DORA 2024 report found that every 25% increase in AI adoption is associated with a 7.2% drop in delivery stability despite individual productivity gains. This creates a governance tension for organisations mandating AI coding tools, where productivity mandates may degrade delivery reliability. Project failure rates for general software have remained structurally stable at approximately 19% cancelled (Standish CHAOS), while AI/ML projects face a structurally different non-deployment rate of 87-90% attributed to the operationalisation gap between proof-of-concept and production service. Useless-test prevalence and long-tail repository abandonment rates lack rigorous population-level longitudinal data, representing the two largest evidence gaps.

Key Findings

1. Repository creation on GitHub more than doubled between 2019 and 2024, growing from approximately 44 million new repositories per year to over 121 million per year.

2. Annual commit volume on GitHub reached nearly 1 billion in 2024, a 25% year-over-year increase, with a record of nearly 100 million commits in a single month.

3. Short-cycle code churn (lines revised within two weeks of creation) rose 84% relative within the GitClear dataset, from 3.1% of changed lines in 2020 to 5.7% in 2024, coinciding with the period of widespread AI coding assistant adoption.

4. The share of code classified as deliberately refactored ("moved code") fell 61% relative in the GitClear dataset, from 24.1% in 2020 to 9.5% in 2024, eroding the structural quality improvement that refactoring provides to long-lived codebases.

5. The DORA 2024 report found that every 25% increase in AI adoption on a software team is associated with a 7.2% decrease in delivery stability and a 1.5% decrease in delivery throughput, despite a 2.1% individual productivity increase.

6. Copy-pasted code blocks rose from 8.3% to 12.3% of all code changes in the GitClear dataset between 2020 and 2024, and duplicated code blocks of five or more identical lines became eight times more common during 2024.

7. Among popular open source GitHub projects, 16% were abandoned (with 41% of those rescued by new maintainers), while industry estimates for newly created projects place the 12-month inactivity rate at 60-95% depending on the definition used. ([inference]; medium confidence for popular project figure, low confidence for new-project estimate; source: Avelino et al. (2019) "On the abandonment and survival of open source projects" gitnux.org

8. The Standish Group CHAOS 2020 report found software project cancellation rates stable at approximately 19%, with 50% of projects challenged and 31% successful, a distribution consistent with secondary summaries spanning the 2020-2023 window; the primary report is paywalled.

9. AI and ML projects face a structurally higher non-deployment rate of 87-90%, with IDC/Lenovo finding that of every 33 AI proof-of-concept projects only four reached production, reflecting an operationalisation gap that is distinct from project cancellation; both figures are drawn from secondary media coverage of paywalled primary reports.

10. AI-generated test suites frequently produce tautological tests that achieve high line coverage without effective fault detection; practitioner analyses report mutation scores below 5% in suites with 100% line coverage, though no population-level longitudinal dataset for this metric exists.

11. Baltes et al. (2026) characterise AI slop in software development as a tragedy of the commons: individual productivity gains from AI-generated code externalise review and maintenance costs onto the broader team, with Quality Degradation, Review Friction, and Forces and Consequences as the three empirically identified impact clusters.

Assumptions

1. The GitClear dataset (211M changed lines, enterprise and high-profile open source codebases) is broadly representative of enterprise software development trends, though it likely underrepresents personal, student, and AI-generated throw-away repositories. Justification: the dataset includes named large-scale production codebases (Google, Meta, Microsoft, Chromium, VS Code) with transparent methodology. Source: www.devclass.com

2. GitHub Octoverse repository counts include a substantial fraction of experimental, student, and AI-generated repositories. The headline count growth therefore overstates the growth in commercially maintained software projects. Justification: the Octoverse 2025 report explicitly notes that 80% of new developers use Copilot in their first week, suggesting a large proportion of new repositories are tutorial or experimental. Source: GitHub Octoverse 2025 (2024/2025 data)

3. The AI/ML non-deployment rate (87-90%) reflects a structural deployment barrier that predates the current AI coding assistant era. It is a persistent characteristic rather than a new 2022-2025 trend, though the growth in AI POC project creation increases the absolute number of unshipped projects. Justification: the VentureBeat figure dates to 2019; the IDC/Lenovo 2025 figure is consistent, suggesting stability in the rate. Source: teaminnovatics.com www.cio.com

Analysis

The seven metrics fall into two structural groups when assessed against the available evidence.

The first group (repository creation, commit volume, LOC output) shows clear upward trends that are well-documented by primary sources. These metrics are supply-side: they measure what developers are producing, not whether that production has quality or business value. The growth in these metrics is largely explained by developer community growth (40M to 180M+ on GitHub over the study period) compounded by AI coding assistant adoption lowering the marginal cost of generating code.

The second group (rework, refactor share, AI slop indicators, delivery stability) shows degradation in the 2020-2024 period. The GitClear and DORA data converge in direction: AI coding assistant adoption appears to accelerate code production at the cost of code composition quality. Two competing explanations must be acknowledged. First, the GitClear dataset is enterprise-skewed and the DORA signal may represent teams that have not yet adapted review processes to AI-generated volume, a transient adaptation lag rather than a structural degradation. Second, developer community growth from 40M to 180M+ on GitHub over the study period brought a substantially larger proportion of less-experienced practitioners into the ecosystem independently of AI tool adoption; this demographic shift could account for some of the quality metric decline without any causal role for AI tools. The available evidence cannot definitively separate these explanations, and both are consistent with the observed data.

Project non-delivery (abandonment, cancellation, non-deployment) has two distinct regimes. General software cancellation rates (Standish CHAOS at approximately 19%) have remained broadly stable across the 2020-2023 window covered by available secondary summaries. This stability implies a structural organisational constraint rather than a technology-driven trend, though this interpretation is an inference from the stable rate data. AI/ML non-deployment rates (87-90%) reflect an additional barrier: the gap between building an AI model or proof-of-concept and operating it reliably as a production service.

Useless-test prevalence remains the weakest-evidenced metric area. The mechanism (AI-generated tests optimise for coverage metrics rather than fault detection) is theoretically coherent and supported by qualitative practitioner accounts, but lacks a population-level quantitative baseline. Teams relying on line coverage as their quality gate may be systematically unaware of how far their effective test coverage has declined.

The related completed item on productivity incentive metrics (2026-05-08-productivity-incentive-metrics-quality-review-agentic-ai, github.com addresses the governance implication of this pattern in detail: speed-focused incentives create hidden quality costs, and code acceptance rate is an insufficient organisational metric. The rising rework rates (KF3) and declining refactor share (KF4) documented here have direct implications for technical debt accumulation rates; the related completed item on IT throughput constraint magnitude and debt accumulation rate (2026-05-16-it-throughput-constraint-magnitude-and-debt-accumulation-rate, github.com quantifies how short-cycle churn and quality degradation compound into throughput constraints over multi-year horizons.

Risks, Gaps, and Uncertainties

1. No peer-reviewed longitudinal study covers all seven metrics in a single consistent dataset over the 2019-2026 window. The evidence base is a patchwork of industry reports (Octoverse, DORA, GitClear, Standish), vendor benchmarks (LinearB), one peer-reviewed academic study on abandonment (Avelino 2019), and one qualitative arXiv preprint on AI slop (Baltes 2026).

2. Useless-test prevalence has no population-level quantitative baseline. The practitioner reports and mutation score case studies are directionally consistent but not representative.

3. The Avelino et al. abandonment study (2019) is the most rigorous source for the popular-project abandonment figure, but it predates AI coding assistant adoption and has not been replicated over the 2022-2025 window.

4. The GitClear dataset is enterprise-skewed and vendor-produced. While the methodology is transparent and the dataset large, independent replication has not been published within the scope of this investigation.

5. LOC per unit of meaningful functionality remains unmeasured at scale. The divergence between raw commit volume growth (+25% YoY) and delivery stability decline (DORA -7.2% per 25% AI adoption) implies that the relationship between volume and value has changed, but the precise magnitude of this decoupling is not quantified.

Open Questions

• Has the per-developer repository creation rate changed since AI coding assistant adoption, or does headline repository growth track developer count growth proportionally?
• What does a longitudinal mutation-score analysis across a representative repository population show for 2019-2025?
• Do repositories created in 2022-2025 (post-AI coding assistant era) show different abandonment or inactivity rates at 12 months compared to pre-AI cohorts?
• Is the DORA AI adoption stability degradation (-7.2% per 25% adoption) a one-time adjustment cost or a persistent steady-state effect?

Output

Type: knowledge

Description: Seven software-delivery metrics are tabulated with baseline values (2019-2020), current values (2023-2025), trend direction, and confidence levels. Key data: repository creation up 175% (44M to 121M/year); commit volume near 1 billion/year (+25% YoY); short-cycle churn up 84% (3.1% to 5.7%); refactor share down 61% (24.1% to 9.5%); AI adoption correlates with 7.2% stability drop (DORA 2024); general software cancellation rate stable at 19%; AI/ML POC non-deployment rate 87-90%.

Three most important sources:

• GitHub Octoverse 2025 - primary data on repository creation and commit volume
• DORA Accelerate State of DevOps Report 2024 - primary data on delivery performance and AI impact
• GitClear AI Copilot Code Quality Report 2025 - primary data on code composition quality trends

sources

• GitHub Octoverse 2025 (2024/2025 data) - large-scale repository and contribution trend baselines.
• DORA Accelerate State of DevOps Report 2024 - DevOps Research and Assessment (DORA) software-delivery outcomes and performance metrics.
• DORA Accelerate State of DevOps Report 2023 - longitudinal engineering performance evidence.
• GitClear AI Copilot Code Quality Report 2025 - 211 million changed LOC dataset on AI impact on churn and rework.
• Avelino et al. (2019) "On the abandonment and survival of open source projects" - empirical study of 1,932 popular GitHub projects.
• Baltes et al. (2026) "An Endless Stream of AI Slop" - qualitative analysis of community perceptions of AI slop in software development.
• LinearB Engineering Metrics Benchmarks - rework, refactor, and pull request benchmarks from 4,800+ teams.
• Standish Group CHAOS Report 2020 summary - software project completion and cancellation rates.
• GitHub Octoverse 2019 - historical baseline for repository and developer counts.
• GitHub Octoverse 2020 - pandemic-era baseline for repository and developer counts.