AI agents aren't squeezing out new open-source contributors, study finds

A widespread concern in the open-source community holds that AI-powered coding agents could harm project sustainability by automating the beginner-level tasks that traditionally serve as on-ramps for new contributors. Without fresh faces tackling these gateway issues, the theory goes, projects would eventually struggle to find experienced maintainers. But recent research suggests this doomsday scenario may not be unfolding in practice.

What the study measured#

A July 2 submission to arxiv.org from Peking University researchers examined how 1,888 GitHub projects changed after adopting AI coding assistants like Cursor and Claude Code. The team identified adoption by pinpointing when each repository first committed a configuration file (such as .cursorrules or CLAUDE.md), then compared those projects against similar ones that never adopted such tools.

Using difference-in-differences methodology—a rigorous statistical approach designed to isolate causation from background trends—the researchers found minimal disruption. Newcomer participation remained flat or edged slightly upward, with the most pessimistic statistical model showing only a 1.5% decline that fell short of statistical significance.

Complexity rises, but newcomers stay put#

Code complexity did increase after AI adoption. Cyclomatic complexity (which counts independent execution paths in a function) climbed 3–4% across languages. Cognitive complexity in Python projects jumped roughly 11%—substantial, but markedly lower than a 2023 Carnegie Mellon study that found a 41% increase under Cursor adoption.

Critically, the Peking researchers then focused on the 128 Python projects where complexity actually spiked. Even there, newcomer retention held steady and the active contributor base grew. The data suggests that rising code complexity and flat newcomer participation are two independent patterns, not causally linked. Higher complexity hasn't deterred fresh contributors—at least not yet at observed levels.

Significant limitations#

The analysis has important bounds. It concentrated on established projects with at least six months of pre-AI history, excluding nearly two-thirds of repositories that adopted AI tools almost immediately upon creation. For those younger projects, no meaningful before-and-after comparison was possible.

Additionally, the study identifies adoption through configuration files rather than actual usage frequency. It therefore reveals what happened after adoption, but cannot distinguish between heavy and light users of AI tools.