One quarter of PRs are "AI junk", the Linux Foundation released a new report: AI writes code 52% faster, yet open source maintainers are more overburdened.

AI programming tools are drastically cutting down the time developers spend writing code, but the open source community may be paying another price for this efficiency improvement.

On October 20, the Linux Foundation released its latest research report Navigating Open Source Software Development in the Generative AI Era, investigating the impact of generative AI on open source software development, code contributions and community maintenance.

The report presents a set of strikingly contrasting data: respondents estimated that the average time spent on code writing tasks was reduced by 52% after adopting generative AI; at the same time, open source project maintainers estimated that about 25% of the Pull Requests (PRs) they received are low-quality contributions, referred to as "AI slop". In large-scale open source projects, this proportion further rises to 31%. 59% of maintainers stated that the number of code contributions their projects received has grown since the rise of generative AI; this figure reaches 69% among large projects.

AI is making submitting code easier and easier, but whether maintainers can keep up with the speed of code growth has become a new challenge.

1 AI writes code faster, but complex tasks do not necessarily save time

This Linux Foundation research was published by LF Research, sponsored by Red Hat, and supported by Futurewei, OpenSSF, and LF AI & Data.

The Linux Foundation stated in its official press release that the efficiency improvement brought by generative AI to open source development covers multiple links, far more than just code generation. Among the six types of open source development activities covered in the study, respondents reported time savings to varying degrees. The average time reduction for code writing was the largest, reaching 52%; even for communication tasks such as proposal discussions, a time saving of about 19% was achieved.

However, this efficiency improvement does not apply to all scenarios. For tasks with high complexity, 21% of respondents said that using generative AI actually takes more time. This means that the efficiency advantage demonstrated by AI in generating code and handling routine development tasks cannot necessarily be directly translated into an overall efficiency improvement for complex software engineering tasks.

The time saved has not all been converted into more code output. 66% of respondents said they have invested the efficiency gains in technical debt governance, 63% in developing new features, and 58% in learning and skill improvement.

These data show that AI tools are changing the way developers allocate their time. However, the improvement of individual developer efficiency does not mean that the entire open source collaboration process can obtain the same magnitude of benefits.

2 A quarter of PRs are considered "AI slop", increasing pressure on maintainers

Compared with code writing efficiency, what is more noteworthy in the report is the pressure that the open source maintenance link is under.

59% of maintainers reported an increase in the number of code contributions, but these new contributions cannot all be directly converted into project value. According to maintainers' estimates, about a quarter of the PRs received are low-quality content. In large projects, this proportion reaches 31%.

The "AI slop" here usually refers to AI-assisted contributions with poor quality, insufficient verification, or that require maintainers to invest a lot of energy to judge their value. It needs to be emphasized that this proportion comes from the estimates of survey respondents, not objective statistics obtained by researchers after testing each PR on platforms such as GitHub one by one, nor does it mean that all low-quality PRs can be confirmed to be generated by AI.

However, the maintenance pressure reflected by this set of data deserves attention.

In the past, developers needed to spend a considerable amount of time understanding the project structure, writing code, and running tests before they could submit a PR. With the popularization of AI Coding tools, the threshold for generating code and submitting modifications has been significantly lowered. For contributors, proposing a code modification may only take a very short time; for maintainers, judging whether the code is correct, whether it conforms to the project architecture, and whether it introduces security issues still requires a lot of effort.

Especially in large open source projects, maintainers also need to deal with issues such as code compatibility, long-term maintenance costs, contributor communication, and project governance.

The Linux Foundation points out in the report that maintainer burnout has become one of the most prominent concerns for large open source projects.

This change also brings a new measurement dimension to the productivity issue of AI Coding: if the cost of code generation drops significantly, but the costs of code verification, review and maintenance do not drop synchronously, the efficiency gains may be redistributed among different roles.

3 71% believe AI lowers the open source participation threshold, but community trust has not improved synchronously

There is another set of contradictions in the impact of AI on the open source community.

The survey shows that 71% of maintainers agree that generative AI lowers the threshold for participating in open source projects. This means that more developers who lack experience in specific programming languages, frameworks or projects have the opportunity to participate in open source collaboration with the help of AI tools.

But at the same time, 53% of maintainers disagree with the statement that "generative AI has improved community trust".

Collaboration in the open source community depends on many factors beyond code. Maintainers not only need to know whether a piece of code can run, but also need to judge whether the contributor understands the modified content, can explain the design choices, and has the ability to continue maintenance when problems arise later.

AI can help developers quickly generate modification solutions, but it cannot automatically build trust between contributors and maintainers.

The report also found that 46% of maintainers believe that generative AI has increased the demand for mentoring new contributors; 32% believe that this demand has decreased to some extent.

This also shows that although AI has lowered the technical threshold for new developers to enter the open source community, newcomers still need to understand project specifications, collaboration processes and software engineering practices. Part of the work that was originally used to guide code writing may be shifting to code verification, problem explanation and contribution quality management.

4 Only 35% of projects have established formal AI contribution policies

Faced with the rapid growth of AI-assisted code contributions, the governance mechanisms of the open source community have not yet fully kept up.

The survey results from the Linux Foundation show that only 35% of projects have formulated formal policies that explicitly allow the use of generative AI in code contributions. Another 32% of projects have no relevant policies and no plans to formulate such policies. On the contributor side, 61% of respondents said they will conduct detailed reviews and complete functional tests before submitting AI-generated code. However, there is still no consistent standard for the disclosure of AI usage.

This brings a practical problem: when maintainers receive a code contribution, they may not be clear about which parts of it are generated by AI, what verifications the contributor has performed, and whether the code modification has been fully tested. For open source projects that rely on volunteer maintainers, this information asymmetry may further increase review costs.

The Linux Foundation recommends that open source projects establish clear policies for AI-assisted contributions, put forward requirements for code testing, documentation and AI usage disclosure, and at the same time improve the contribution attribution and verification mechanism. The relevant person in charge of OpenSSF also emphasized in the official statement that AI can speed up code writing, but contributors still need to understand and test the content they submit, and maintainers should not bear all the security review responsibilities alone.

5 The efficiency dividend of AI Coding: who will bear the verification cost?

In the past two years, discussions on AI Coding have largely focused on code generation speed, task completion rate and developer productivity. From code completion to Coding Agents that can autonomously modify repositories, execute tests and submit PRs, automation capabilities are constantly being enhanced.

This report from the Linux Foundation provides another perspective for observation: after the code production speed increases, the bottleneck in the software engineering process may gradually shift to the review, verification and maintenance links.

This change is particularly obvious for open source projects. Code contributions to open source projects can come from any developer around the world, but maintenance permissions are usually concentrated in a relatively limited number of core maintainers. AI tools can allow more people to generate code quickly, but cannot ensure that the time and energy of maintainers grow at the same speed.

This does not mean that AI-assisted contributions will necessarily reduce the quality of open source software. The report also shows that AI is lowering the participation threshold, helping developers pay down technical debt, develop new features, and improve learning efficiency.

The problem is that as AI-generated code becomes the norm, the open source community needs to establish matching verification and governance mechanisms. In the future, to measure the value of AI Coding, in addition to how much coding time developers have saved and how many tasks the Agent has completed, we also need to pay attention to the final acceptance rate of code, the time required for review, the cost of defect repair, and the long-term maintenance burden.

AI makes code easier and easier to produce, but making this code truly enter a reliable software system still requires costs.

Linux Foundation Report Link:

https://www.linuxfoundation.org/research/oss-development-gen-ai-era-2026

This article is from the WeChat Official Account "InfoQ" (ID: infoqchina), written by Cai Fangfang, and published with authorization from 36Kr.