GitHub Study Finds AI-Assisted Repos Carry More API Risk

Tsuchida and coauthors studied 622 GitHub users who publicly signal generative-AI coding tool use, per a paper posted to arXiv. From that group, they matched 179 repos carrying visible AI-assistance config against 179 traditional repos with none.

The stakes show up in the issue trackers, not the code: complaints on AI-assisted repos skew toward external dependencies, per the paper. Of 248 issues logged against the AI-assisted set, the recurring ones are provider API rate limits and reliance on the GenAI provider's own API.

The matched repos also look different on the surface. AI-assisted ones carry longer READMEs with more headers and code blocks; traditional repos lean on more external URLs instead.

Their own conclusion, per the paper: generative AI relocates maintenance cost toward verifying generated content and managing provider dependencies.

Separate incident-telemetry numbers from Faros reached the same conclusion using production data instead of repository metadata.

Budget for GenAI providers like vendors, not dependencies. Rate limits and endpoint changes are the outage risk here, and code review won't catch either one.