The conversation about "citing sources" misses a deeper tension: LLMs excel at tasks where we can verify outputs (code, translations, summarization of provided text) but struggle where we can't (novel research, edge cases in specialized domains).

The practical heuristic I've settled on: use LLMs as a first-pass generator when you have the expertise to audit the output, not as a replacement for building that expertise. The danger isn't hallucination per se—it's the confidence gap where LLM fluency exceeds your domain knowledge. That's where subtle errors become invisible.

The tool calling approach (RAG, web search) helps but introduces a new problem: now you're trusting the LLM's retrieval strategy and synthesis judgment, not just its factual recall. It's trading one set of failure modes for another.