AI saves coding time—but who absorbs the review work?
A Linux Foundation survey of 234 open-source contributors and maintainers reports a 52% average self-estimated reduction in code-authoring time. Maintainers also estimated that one in four incoming pull requests was low quality.
The short version
- What happened
- Among 234 qualified respondents, contributors self-estimated an average 52% reduction in code-authoring time when using generative AI.
- Who it affects
- Volunteer maintainers may absorb more triage and review work when AI makes it easier to submit plausible but weak patches.
- What remains unproven
- The study uses a self-selected web survey rather than repository telemetry or a controlled comparison.
Keep in mind
The study uses a self-selected web survey rather than repository telemetry or a controlled comparison.
Key questions have smaller denominators than the 234-person full sample, increasing uncertainty for subgroup estimates.
Reported time savings and pull-request quality are participant estimates and do not measure downstream defects or maintenance cost.
The work was sponsored by Red Hat and supported by organisations active in open source and AI governance.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Faster authoring does not remove the verification work
Open-source contributors using generative AI estimated that it cut their code-authoring time by 52% on average, according to a Linux Foundation Research survey published on 20 October. Yet maintainers answering a separate question estimated that 25% of incoming pull requests were low quality. In projects with more than 50 contributors, that estimate rose to 31%. The pairing matters: time saved while drafting code can reappear as review, testing and explanation work for someone else.
The report also says 59% of maintainers saw contribution volume growing. More participation can strengthen a project, but maintainers must decide whether a change is correct, secure, licensed appropriately and consistent with project conventions. Generative AI may lower the effort needed to propose a patch without lowering the expertise needed to accept it or document why it was rejected for another contributor. The survey therefore points to a redistribution of work, not simply a universal productivity gain.[1]
The survey reached 234 people, with smaller denominators for individual questions
Researchers fielded a 45-question web survey from April to June 2026 and retained 234 qualified responses. The sample included 206 contributors and 134 maintainers, with roles overlapping. Participants were recruited through Linux Foundation newsletters and social channels, plus communities connected to Red Hat, OpenSSF, Kaiyuanshe, GOSIM and LF AI & Data. The report gives the full sample a margin of error of plus or minus 6.41 percentage points at 95% confidence.
Individual results often use smaller bases. The low-quality pull-request estimate came from 111 maintainers, so its uncertainty is wider than the full-sample figure. The 52% authoring-time result is a respondent estimate rather than observed elapsed time. Respondents were self-selected and already engaged with open source, which limits generalisation to all developers, companies or projects. The study is cross-sectional and cannot show that AI caused the reported changes.[1]
The distinction that matters
Where an apparent coding-time gain can move
The survey separates authoring from the downstream work needed before code becomes dependable.
Draft
AI assists code authoring
Respondents report time saved
Review
Maintainers inspect and test
Quality burden may shift
Operate
Projects maintain accepted code
Long-term cost remains unmeasured
Teams need review capacity, disclosure rules and evidence from actual repositories
Governance is uneven. Thirty-five per cent of respondents said their organisation had a formal policy permitting generative AI, while 32% reported no policy and no plans for one. Sixty-one per cent of contributors said they performed detailed review and full testing of AI-assisted contributions. Those answers suggest that conscientious practice exists, but they do not reveal whether the checks consistently catch security, licensing or maintenance problems.
For maintainers, the practical response is to make contribution expectations explicit, automate routine checks where appropriate and budget human attention for complex review. Employers adopting AI coding tools should measure the whole workflow: authoring time, review time, rework, defects, security findings and long-term maintenance. They should also distinguish a patch that is rejected quickly from one whose hidden defects emerge after release. The report was sponsored by Red Hat, with support from Futurewei, OpenSSF and LF AI & Data. Stronger evidence would link disclosed AI use to repository telemetry and compare matched projects over time.[1]