Code wikis are documentation theater as a service (passo.uno)

Google’s new Code Wiki (and similar tools like DeepWiki) promises to “Stop documenting. No more stale docs. Ever” by ingesting a repository, using an LLM to generate complete developer docs and diagrams, and exposing an agent-driven Q&A — a process that can take roughly ten minutes even for large codebases. In practice these wikis can be useful for quick project summarization and exploration, but the output is often superficial: impressive at first glance, then riddled with hallucinations, subtly wrong facts, brittle information architecture and misleading diagrams. The result is “documentation theater as a service” — a polished façade that looks like docs but fails the tests of accuracy, accountability and long-term maintenance. The broader significance for the AI/ML community is cultural and practical. These tools expose a demand for faster, more accessible knowledge, but they also risk lowering the bar for what counts as documentation and displacing the human work of technical writing. Best practice: treat LLM-generated wikis as a starting point—use them to surface gaps, draft outlines, or power exploratory Q&A—but never ship AI-only docs. Human ownership, content integrity and sustained curation remain essential; AI should augment documentation workflows, not be given the final word.