Finding knowledge gaps, key concepts, and unknown unknowns; fixing unit tests after making a change; fuzzy codebase search and generating hypotheses for debugging if you get stuck; and RAG to find the most relevant information in a book or specification you haven't had the time to read fully yet are all IMO reasonable applications of AI that won't necessarily lead to ruin over time if done carefully, cross-checked with actual sources, and done with the understanding that they could be wrong and might be missing something.