Agent-Generated Code Maintenance Asymmetry¶
AI-generated files get about half the commit frequency of human-authored ones, and their changes skew toward features over bug fixes — a distinct maintenance footprint.
The Evidence¶
An empirical study of 508 AI-generated files and 1,543 modifying commits across 100 GitHub repositories in the AIDev dataset measured how AI-generated and human-authored files diverge after merge (arXiv:2605.06464). Three asymmetries appear in a file's first six months.
Frequency¶
AI-generated files receive roughly half the commit frequency of human-authored files in the first month, with activity declining further over the subsequent two months. The magnitude of each modification is also smaller — fewer lines touched per commit (arXiv:2605.06464).
A line-level replication on a separate AIDev sample finds the same direction with statistical significance: AI-authored lines have a 16% lower hazard of modification (HR = 0.842, p < 0.001) and a 15.8 percentage-point lower modification rate (arXiv:2601.16809).
Modification mix¶
The composition of changes inverts between origins.
| Modification type | AI-generated | Human-authored |
|---|---|---|
| Feature addition | 21.78% | — |
| Refactoring | 14.19% | — |
| Bug fix | 11.73% | 16.76% |
| Documentation | — | 16.22% |
For AI-generated files, feature additions lead. For human-authored files, bug fixes lead (arXiv:2605.06464).
Who maintains¶
Humans perform 83.21% of maintenance commits on AI-generated files. On human-authored files the share is 92.98%. The agent that wrote the file rarely returns to maintain it (arXiv:2605.06464).
Two Readings of the Same Data¶
The authors flag the interpretive ambiguity directly: lower modification rates "might suggest superior code quality, yet developers may avoid modifying AI-generated code due to difficulty in comprehending it" (arXiv:2605.06464).
The second reading has separate support. A study of 302,600 AI-authored commits across 6,299 repositories found 22.7% of AI-introduced issues survive to the repository's latest version, with code smells accounting for 89.3% (arXiv:2603.28592). Lower commit volume is not the same as fewer defects.
Why the Mix Shifts to Features¶
The Sawada et al. paper offers one interpretation: "generated files may lack sufficient coverage of requirements" — agents under-deliver on scope, leaving humans to add the missing capability after merge (arXiv:2605.06464). This fits the broader pattern that agents skew simpler — Codex-assisted PRs change cyclomatic complexity 9.1% of the time versus 23.3% for human PRs (arXiv:2507.15003).
Practical Implications¶
Treat AI-authored regions as a distinct maintenance category. The 83/17 split means the team that ships an AI-authored file maintains it, rarely with help from the agent that wrote it.
Do not read low commit frequency as a quality signal. Two mechanisms produce the same observation; pair the metric with defect-survival or comprehension-test scores before treating stability as cleanliness.
Audit for missing scope, not just bugs. Feature additions dominate post-merge maintenance, so ask "what did the agent leave out?" rather than "what did it get wrong?"
Modification timing tracks organizational factors. Predicting when AI-authored code gets touched scores Macro F1 = 0.285 on textual features alone — review depth, ownership, and comprehension drive timing more than code shape (arXiv:2601.16809).
Key Takeaways¶
- AI-generated files see roughly half the commit frequency of human-authored files in the first month, with smaller per-commit change magnitude
- The modification mix inverts: AI files are dominated by feature additions (21.78%) and refactoring (14.19%); human files are dominated by bug fixes (16.76%) and documentation (16.22%)
- Humans perform 83.21% of maintenance on AI-generated files; agents return for only 16.79%
- Lower modification rates have two competing explanations — better quality or comprehension-driven avoidance — and the data does not distinguish them
- Feature-addition dominance suggests AI-generated files under-cover requirements at merge time
When This Pattern Doesn't Apply¶
The 6-month observation window and the 100-repository AIDev sample bound generalizability. The asymmetries are unlikely to hold in three contexts:
- Throwaway scaffolding: prototypes, spike branches, and generated boilerplate that no one was going to maintain regardless of authorship
- Highly specified ticket workflows: narrow, fully specified tasks (e.g. "implement this DTO") where low post-merge modification reflects spec stability, not avoidance or quality
- Closed-loop AI authoring and maintenance: pipelines where an agent both writes and maintains code (e.g. agentic refactor jobs on cron) collapse the human-vs-agent maintenance split and invalidate the 83/17 ratio
One caveat on that last context: routing maintenance back to agents does not make the lower-volume reading safe. Agents introduce fewer breaking changes when generating new code (3.45% vs 7.40% for humans) but more during maintenance — 6.72% on refactoring, 9.35% on chore changes — a "Confidence Trap" where highly confident agentic PRs still break callers (arXiv:2603.27524).
Related¶
- Shadow Tech Debt — the architectural-incoherence mechanism that produces issues agents do not return to fix
- Comprehension Debt — the human-side mechanism for why developers avoid modifying code they did not author
- Agent PR Volume vs. Value — pre-merge counterpart: agents author at high volume but lower acceptance, and skew structurally simpler
- CRA-Only Review and the Merge Rate Gap — review-side mechanism by which AI-authored work makes it through merge despite missing scope
- Predicting Reviewable Code — predictive signals for AI-generated functions likely to be deleted, complementary to maintenance-survival data
Sources¶
- arXiv:2605.06464 — Sawada, Shirai, Kashiwa, Yamaguchi, Iwata, Iida: "To What Extent Does Agent-generated Code Require Maintenance? An Empirical Study"
- arXiv:2601.16809 — "Will It Survive? Deciphering the Fate of AI-Generated Code in Open Source" — line-level survival analysis on a separate AIDev sample
- arXiv:2603.28592 — "Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild"
- arXiv:2507.15003 — AIDev dataset paper
- arXiv:2603.27524 — "Safer Builders, Risky Maintainers: A Comparative Study of Breaking Changes in Human vs Agentic PRs" — agents are safer at generation but riskier during maintenance (the "Confidence Trap")