Researchers estimated the share of AI code in F-Droid catalog apps

An independent researcher conducted an audit of F-Droid to determine the scope of software created using generative models. The sample included 102 applications from the popular free software catalog, which were analyzed for signs of using LLM when writing the source code. The results showed that some projects, such as Amber, are classified as "mostly AI-generated". During the check, characteristic patterns were identified: a specific commit structure, the presence of infrastructure for AI agents, and atypical code fragments for manual writing. The author of the study notes that accurate detection remains a difficult task due to the lack of uniform standards for marking AI content in repositories. In the analysis process, methods of statistical evaluation of the frequency of occurrence of certain constructions typical for models like GPT-4 or Claude were used. Despite the subjectivity of some criteria, the study confirms the trend of an increase in the share of automated code in open-source projects. The full report and methodology are available in the discussion on Hacker News. Such studies highlight the problem of transparency in the development of free software. With the increasing availability of code autocompletion tools, the line between author's contribution and machine generation becomes less noticeable, which requires the development of new verification methods to ensure security and support code quality in the long term.