# 25% of PRs in open source software are AI slop

> An LF Research survey with 234 responses puts low-quality pull requests made with AI at 25% and time savings when programming at 52%.
> By José Maldonado · Artificial Intelligence · 2026-10-21

Artificial intelligence already writes a good part of open source code and, at the same time, fills the inbox of those who review it. The figure? 25% of pull requests (PRs, or requests to add code to the project) reaching open projects are AI slop, the label that [LF Research](https://www.linuxfoundation.org/research) uses for low-quality contributions generated with models that do not offset the review time they demand from maintainers. The result is a new bottleneck, located at the front door rather than at the keyboard.

The figure came out on October 20, 2026, in a report presented at [All Things Open 2026](https://2026.allthingsopen.org/), in Raleigh, sponsored by Red Hat and backed by Futurewei, OpenSSF and LF AI & Data. The same survey puts a number on the savings: writing code with the help of a model cuts the time for that task by 52%, and across all the activities measured the range runs from 19% to 52%.

The paradox lies in who pays the bill. The tool that speeds up the writer multiplies the work of the reviewer, and 64% of large projects already report slower reviews. With 234 valid responses and a margin of error of ±6.41 percentage points, the study turns a diffuse complaint into an uncomfortable figure for the entire software industry.

## A sample of 234 people, not an anecdote

The report is titled [*Navigating Open Source Software Development in the Generative AI Era*](https://www.linuxfoundation.org/press/new-lf-research-survey-shows-generative-ai-cuts-open-source-coding-time-in-half-but-25-of-incoming-pull-requests-are-considered-ai-slop) and it is signed by Sam Boysel, data scientist at the Linux Foundation, and Anni Lai, of Generative AI Commons.

The methodology section is explicit: a 45-question web survey, fieldwork between April and June 2026, 234 qualified responses and a margin of error of ±6.41 points at 95% confidence. Distribution was handled through the channels of the Linux Foundation and its partners.

The sample is spread across roles that often coexist in the same person: 206 of the 234 respondents identify as contributors and 134 as maintainers. 81% take part out of personal interest, 47% are paid by their company to do so, and 52% of contributors have eleven years or more of experience.

Two things should be qualified before citing the figure in a meeting. The slop percentage is a subjective estimate by each maintainer about their own inbox, not an automatic count of commits. And the questionnaire circulated among those already inside the ecosystem, so the sample skews toward infrastructure projects.

## The real savings, task by task

Respondents estimate how long tasks took before and how long they take now with model assistance across six activities in the development cycle. The result is not a single average but a scale: the 52% saving in code writing drops to 19% in discussions of proposals, where the work consists of reading, understanding and deciding rather than producing text.

Activity measured

Average time saving

Writing code

52%

Discussion of proposals

19%

Technical discussion and review of proposals, maintainers only

25% to 40%

That breakdown has a clear reading for maintainers: AI relieves them of load in coordination tasks, not just at the keyboard. Even so, the report acknowledges an important limit, namely that 21% of respondents admit taking longer with AI on the most complex tasks.

The time gained does not disappear. 66% reinvest it in technical debt, 63% in new features and 58% in training, according to the responses. In other words, AI does not reduce the ecosystem's total load: it shifts it toward work that had been postponed for years, which in turn generates more code that someone will have to review later.

## Who pays the bill for filtering?

Volume is the first half of the problem. 59% of maintainers report more contributions since AI arrived, a percentage that rises to 69% in projects with more than fifty collaborators. The second half is quality: that 25% average becomes 31% in large projects and 23% and 20% in medium and small ones.

Burnout follows the same slope. In small and medium projects the top concern is code quality; in large ones, first place goes to maintainer exhaustion.

Co-maintainers estimate 29% slop, almost six points above lead maintainers, and 64% of large projects see their reviews lengthen versus 46% of small ones. Each junk contribution costs, in practice, the time of several people who could be reviewing useful patches.

Many teams have begun using the same technology to defend themselves. Three out of four maintainers use AI tools to filter contributions, above all to summarize complex logic or look for edge-case errors, while 23% use none. That is the line of work behind [Sashiko, the AI with which the kernel hunts its next big bug](https://observatorioblockchain.com/en/cybersecurity/sashiko-ia-revisa-kernel-no-time-to-pwn/).

## Labels, filters and traceability that nobody complies with

Only 35% of projects have a formal policy allowing the use of AI, 32% have none and do not plan to have one, 23% plan to approve one and 10% explicitly prohibit it.

Among those who have already written rules, 76% require declaring the use of AI in the contribution, a formula that shifts the burden of proof to the author instead of banning the tool.

Practice goes another way. 61% of contributors review line by line and test the generated code before submitting it, but only 52% label AI-assisted work most of the time or always, and 19% never label it. 80% consider it important to know which model and which data generated the patch.

The Linux kernel has been in that debate for months. Its documentation still calls for an \`Assisted-by\` tag on patches with AI assistance, and on October 13, 2026, a patch was published on the mailing list to remove that requirement, following the consensus reached at the Linux Maintainers Summit in Prague.

The dispute remains open, with disagreements within the mailing list itself. At the opposite extreme, many GNOME projects prohibit any model-generated content, a rule that Michael Catanzaro criticized in June 2026 for also banning useful bug reports.

Tools such as [Boro, the Rust CLI with AI to audit the kernel](https://observatorioblockchain.com/en/technology/nvidia-boro-cli-rust-kernel-linux-ia-local/) show the other path: automating the filter instead of trusting the author's declaration. Even so, the underlying problem remains how long it takes a person to validate the result.

## The curl case, where slop can be switched off

The curl story serves as a counterweight to the pessimism. Daniel Stenberg, its lead maintainer, said in July 2025 that about 20% of the security reports received that year were slop, that only 5% turned out to be a real vulnerability, and that each report mobilizes three or four team members for half an hour and up to three hours.

Since 2019, the project's bounty program had validated 81 real bugs with more than $90,000 in rewards. That track record was proof that the model worked, and slop was breaking it on the cheapest side: submitting without checking.

Nine months later, the same project reversed the trend. According to the survey itself, in April 2026 Stenberg reported that *curl* had stopped receiving junk reports and was getting more and more good findings, almost all of them made with AI assistance. Slop is not a fixed property of models: it depends on the incentive, the filter and how explicit the project's rules are.

## A bottleneck measured in other people's hours

The report proposes written policies, standardized attribution and requiring authors to test their own patch before submitting it. All three recipes have the same accounting effect, which is to return to the author a cost that the reviewer absorbs today.

Nobody bills those hours, and that is why the public debate remains focused on writing speed while the real funnel is in review. As we explained when analyzing [how AI is changing open source software development forever](https://observatorioblockchain.com/en/ai-agents/a-ia-cambia-para-siempre-el-desarrollo-del-software-abierto/), the technical side is moving faster than governance.

With a quarter of incoming contributions in quarantine and only a third of projects with clear rules, the profitable question is no longer how much code a model saves, but how many maintainer hours each point of savings costs.
