According to TheRegister.com, Google has previewed Code Wiki, an AI project that automatically documents code in repositories and regenerates content after every code change. The preview includes documentation for hundreds of open source projects, with plans to extend to Gemini CLI for internal repositories soon. Code Wiki appears to be a rebuild of Auto Wiki, a project developed by Mutable.ai before Google acquired the company. The service includes a chat interface where developers can ask questions about codebases, though early testing shows mixed accuracy. Google claims reading existing code is among the “biggest, most expensive bottlenecks in software development,” positioning Code Wiki as a solution to this challenge.
The AI documentation dilemma
Here’s the thing about auto-generated documentation: it sounds amazing in theory but gets messy in practice. Developers who’ve tested Code Wiki report it produces “educated guess” documentation full of LLM-isms that oscillates between being too vague and too specific. One person noted the documentation for Vue.js was one-fifth the length of the actual code itself – which is honestly kind of ridiculous when you think about it. And there’s the fundamental question: if you’re going to spend time verifying AI-generated docs, wouldn’t you be better off just reading the code directly?
Accuracy concerns and context gaps
The ASP.NET Core example really highlights the limitations. When asked about using Postgres instead of SQL Server for distributed caching, Gemini gave a partially incorrect answer that missed Microsoft’s official documentation about PostgreSQL support. This isn’t just a small error – it’s the kind of thing that could send developers down completely wrong paths. And Code Wiki didn’t even flag that the vuejs/vue repository it documented has been obsolete for two years. That’s a pretty massive context failure for something claiming to understand codebases.
The regeneration problem
Google‘s approach of regenerating documentation after every code change creates another headache. As one commenter put it, “Does anything persist? If I could be in the middle of reading it, and the next day it’s completely different, that’s a huge waste of my time.” That’s a real concern for teams trying to maintain consistent understanding of their codebase. The constant churn could make documentation feel less like a reliable reference and more like a moving target.
Not the first to try
Google isn’t pioneering this space – Devin’s DeepWiki already offers similar functionality for both open source and private repositories. The real question is whether any AI system can truly grasp architectural decisions and nuanced code relationships. As one developer noted on Hacker News, “Repo-wide documentation really feels like the last thing I’d want AI to touch.” Given that Code Wiki itself includes the disclaimer “Gemini can make mistakes, so double-check it,” maybe we’re not quite at the point where we can trust AI with our most critical documentation needs.
