LF Research is the Linux Foundation’s research arm. It was set up in 2021 and publishes studies on how open source is built and used, drawing on surveys and project data.
Its latest report, Navigating Open Source Software Development in the Generative AI Era, surveyed 234 contributors and maintainers on how generative AI is affecting the way open source gets made.
The survey estimates that about 25% of incoming pull requests are low-quality AI slop, with the estimate climbing to 31% on larger projects.
Red Hat sponsored the report, with additional contributions from OpenSSF, LF AI & Data, and Futurewei.
The findings
Large projects, the ones with more than 50 contributors, are affected the most.
Among their maintainers, 69% report more incoming contributions, against 59% across all maintainers, and 64% say reviews now take longer. Burnout is their top concern at 63%, narrowly ahead of contribution quality at 60%.
You will also find that policies have not caught up to the AI-assisted coding and vulnerability reporting trend.
Of the 119 respondents who answered the question, 35% said their project has a formal policy permitting AI use. Another 32% have none and no plans for one, 23% plan to write one, and 10% ban AI contributions outright.
You might remember that some projects, like LLVM, allow AI-generated code with conditions. Separately, 26% of projects have no way of verifying AI-assisted contributions at all.
The report calls on contributors to give effort to written policies, disclosure of AI use, and testing and documentation.
Interestingly, AI does seem to save developers’ time. Respondents estimate that code authoring takes 52% less time on average, while on more complex tasks, 21% said the work took longer with AI.
Most maintainers (71%) say AI makes it easier for newcomers to start contributing. Just over half (53%) say it has not improved community trust, while only 13% say it has.
Some caveats do apply.
Respondents were recruited through Linux Foundation newsletters and social channels, plus partner channels that include Red Hat. As for the 25% slop figure, it is an estimate from 111 respondents, not a measured count.
That sounds fair, but disclosure is mostly left to individual choice right now, and a quarter of projects have no way to check anything. It only works if the people sending slop make an effort to do it.
And as we have already seen time and again, the ones sending slop rarely bother.
I’m curious to know what the quality rate was like before AI. Anecdotally, my marginally popular repo got a few PRs, but many of them were pretty much garbage as well. No tests, broken tests, half baked code. Human written didnt mean good
Slop has existed before AI but it has lowered the barrier to generating it. Having no AI policies isn’t really working because the slop posters don’t declare it and as a reviewer you don’t want to be second guessing a potential contributor.
We are probably going to move to a policy that allows narrow use of AI tools but emphasises human understanding and ownership as mandatory components of contribution.
I’m lucky, so far no AI slop PRs yet, so I’m in a different boat. I’m happy to keep my little thing going and hope that it slowly dies away with no further attention.
emphasises human understanding and ownership as mandatory components of contribution.
That really the best you can hope for, although I dunno how you can force ownership. Its not easy to hunt down the original developer if there is a bug in their code.
Given that these pr’s will only get better and better, this already sounds like a huge win for all oss users. Maintainers might need a special maintainer agent that filters and guides new agent contributors, but after that we’ll see all oss projects boom with features, quality and constant development, or maintenance as long as anyone uses it. New types of projects - huge, or very complex ones - will start to appear also. It have already started.
At some point even the most negative user will have to stop using the ‘slop’ pejorative for ai driven pr’s, as they’ll just be better than hooman pr’s. One commenter already noted that Hoomans create ‘slop’ too. AI will only do that temporarily, maybe longer than I think, but definitely temporarily.
If you actually read the article, you’ll see that “slop” in the survey does not refer to AI-driven PRs in general, but to specifically low-quality AI-generated PRs. The % of PRs with AI-generated code would be higher. 25% of all PRs being low-quality AI-generated code is a lot.
you’ll see that “slop” in the survey does not refer to AI-driven PRs in general, but to specifically low-quality AI-generated PRs.
This part is actually unclear. The question posed to survey participants that allowed researchers to determine a PR was “AI slop” is left out of the report. And this line under the chart showing “25% slop etc.” is troubling:
Average estimated percentage of incoming PRs that are low-quality and/or AI-generated
How this particular question was framed is of critical importance here and without it, it’s easy to assume that all reported AI use was broadly categorized as “slop”
Sadly I could not find a way to access the full list of questions used in the survey to confirm this.
They aren’t specific about how the question was posed, no, but page 9’s wording does state that it is maintainer-estimated, and specifies the terms “low-quality” and “slop” as the qualifier for the 25% figure, so unless they’ve made a huge logical leap, I think it is reasonable to conclude that maintainers were specifically asked about low-quality AI-generated PRs (or “AI slop”). I think people are sensible enough these days to not deny that AIs sometimes produce acceptable code.
page 9’s wording does state that it is maintainer-estimated, and specifies the terms “low-quality” and “slop” as the qualifier for the 25% figure
I’m not seeing it. This is the sole relevant statement on pg 9 where they simply define AI slop in a general way that avoids tying it in with their measurements:
“AI slop” are low quality AI-generated contributions that are unlikely to be worth the review time they demand.
In addition to the explicit ‘and/or general AI use’ statement, there are several red flags here that make me think this is a low quality study to be taken with salt. Or maybe I’m cynical and using the term ‘slop’ to push a research study is both unusual and suspect.
i think these results are very balanced, specially keeping in mind projects are still adapting to the new ways.
