Cognition's DeepWiki, launched in 2025, will generate a navigable wiki for almost any public GitHub repository in a few minutes. Engineering teams started pointing it, and similar internal tools built on the same idea, at their own decade-old codebases, hoping for a shortcut past years of undocumented decisions. What came back read well. What came back was not always true, and the gap between the two took longer to notice than anyone expected.
A model summarizing a function can describe what it does, not why it exists, what it assumes about the systems around it, or which of its edge cases were intentional. Stripe's API documentation is held up across the industry as close to a gold standard specifically because Stripe engineers wrote it with that intent baked in from the start, not because a model summarized the surface area after the fact. When the underlying code carries that intent, through naming and structure, generated docs read as genuinely useful. When it does not, the model produces confident, plausible prose about behavior nobody actually decided on purpose.
Teams inheriting older systems feel this most acutely. A generated doc that describes an API's quirky behavior without noting that the behavior is a decade-old bug nobody has fixed is worse than no documentation at all, because it carries the authority of something that looks reviewed.
Some engineering organizations have responded by building a review pipeline distinct from ordinary code review, specifically to catch the moment a model has documented an accident as if it were a design choice. That review step costs exactly the time that instant documentation was supposed to eliminate.
What is opening up is demand for tools that ship with explicit confidence markers, flagging which claims come directly from code structure versus which are inferred from naming conventions or comments that may be stale. Mintlify and several smaller documentation startups have started building this distinction into their AI review flows, treating it as a selling point rather than an afterthought.
Stripe's documentation took roughly a decade of engineers writing sentences on purpose. DeepWiki can produce the same shape in an afternoon. What it cannot do, unprompted, is tell you that the shape is hollow in places the original engineers never bothered to explain even to each other.