Troubleshooting: Claude hallucinating references in long outputs
Claude is inserting plausible-sounding but non-existent citations in multi-chapter outputs; need strategies to force verifiable references or integrate citation checks into the workflow.
Best tools for this use case
Based on the workflow in this discussion, these tools are useful starting points to review.
Claude
Excellent for careful reasoning, long-form thinking and structured analysis.
ChatGPT
Best all-round AI assistant for broad knowledge work and workflow acceleration.
Gemini
Strong AI assistant for users already working inside Google's ecosystem.
Answers
Approved replies, operator insight, and tactical follow-up from the community.
Concrete steps:
1) Use RAG—provide the canonical source set and instruct “only cite from this set; include exact URL/DOI.”
2) Chunk long outputs by chapter; after each chunk, output a citation table (title, URL, confidence).
3) Automate verification (HTTP HEAD/DOI resolver + fallback web search) to flag non-existent refs and replace unverifiable ones with [UNVERIFIED].
4) Add a hard system rule: refuse to fabricate citations.