How Outdated API Docs Silently Trained an AI Model to Write Broken Code

The team behind dailymeteo.com, a European meteorological archive, discovered that their fine-tuned GPT model had been trained on examples built from inaccurate API documentation. Key errors included a listed variable, sea-level pressure, that the API had never actually served, a hardcoded start year of 1960 when the archive begins in 1961, and incorrect timestamp format descriptions for climate mean data. Because training examples were written against the documentation rather than the live API, the model confidently generated well-formed but non-functional code for every affected query type. The incident highlights a systemic risk in fine-tuning workflows: documentation describes what a system is supposed to do, not necessarily what it does, and that gap can go undetected until someone tests the real endpoints.