One GSoC mentor got applicant proposals that still contained the chatbot’s reply: “here is the proposal you asked for.” Another says applicants “just AI it up, complete it, and say yeah, I understood everything.” Both are quoted in a 28 October 2026 study of Google Summer of Code , the program that pays newcomers to work on open-source projects under a volunteer mentor. Its authors include Google’s own GSoC staff. The tension is simple: AI lets a newcomer hand in polished work, so mentors can no longer tell who understands it, and the checking that follows eats the time the program depends on.
When a good proposal stops being evidence
A proposal used to show effort. Writing one meant reading the codebase and committing to a plan. Now it takes a prompt. The study, by Xing, Feng, Radomile, Taylor, Vargas and Sarma, drew on 15 interviews and 194 replies across 15 threads on the restricted GSoC mentors’ mailing list, from 25 organizations. One mentor put the problem plainly: “If it’s easier to write a submission, then it’s harder to review a submission.”
Mentors responded with more review. One reports the only significant change as “the amount of effort we’re putting into reviewing PRs”, which after disclosure came to “maybe 20, 30% more” time. Another says “It does create fatigue.” Two organizations paid more heavily. One withdrew: “It is just not in our bandwidth.” Another sat out a cycle for lack of mentors.
That cost lands somewhere specific. The paper describes a cycle in which “today’s mentees become the people who mentor tomorrow’s.” Spend mentor hours on forensic review, and you spend the hours that make next year’s mentors.
A rule you can’t check pushes the work onto you
The paper observes that “the boundary an organization drew determined who did the checking.” Take VideoLAN, the group behind VLC. Its 16 March 2026 GSoC policy says “Proposals and qualification tasks that show signs of AI generation will result in the applicant not being selected.” It does not say how those signs are found.
Compare a mentor in the study whose organization runs a strict ban and later left the program. This mentor says “it is next to impossible to confidently attribute something” to AI, so they act only when they “do find definitive proof.” A ban on authorship asks mentors to prove something they mostly cannot.
Another mentor uses a different test: “Why did you choose to do it this way?” and “How did you check it?” The paper notes that “Rules about understanding let mentors check compliance by asking contributors to explain their work.” A newcomer who can answer is learning. One who cannot has been found out without anyone accusing them.
A ban on AI authorship asks mentors to prove something they mostly cannot, while an understanding rule turns the check into the mentoring itself.
Read across from the paper’s findings, that suggests the choice of rule is a choice about workload. That is Pipeline’s inference, not the study’s measurement. The authors argue the mechanism but did not compare outcomes across organizations, and the only time cost anyone quantifies is that one mentor’s own estimate. It also applies a single program’s evidence to newcomer onboarding in open source generally. It is separate from whether AI-written code is any good, a question we took up when projects began closing their doors to public pull requests .
The program is not collapsing, and AI helps too
The honest complication is that GSoC’s 2026 numbers don’t look like a program in trouble. Google’s own count lists 15,245 applicants and 1,141 accepted contributors across 184 organizations. The Rust Project reported 96 proposals, up 50%, and said it “somewhat struggled with some AI-generated proposals” but “it stayed manageable.” Many strong proposals went unaccepted mainly because of “limited mentorship capacity,” not because of AI spam.
The study’s limits cut the same way. It covers only GSoC, its mailing list may amplify difficult views, it has 15 interviewees, and it recruited no mentees. It also reports that AI helped with onboarding, coordination and first-pass review, though not with “the work of assessing contributor understanding and code quality.”
So the evidence supports a narrower claim than decline. Capacity, not applicant volume, is the scarce thing, and rules should be judged by how much of it they spend. GSoC leaves the rules to each organization , so this winter’s decision belongs to whoever is deciding whether to apply for 2027.
If you mentor newcomers, the rule you write sets your own weekly load. Under an authorship ban, you are the one hunting for signs of AI. Under an explain-it-to-me rule, the check is the conversation you were already having. One of those scales with the number of applicants. The other scales with the number of mentors you still have.
