Hallucination Detection In Research Output

1. 1Citation hallucinations — papers that don't exist
2. 2Statistical hallucinations — numbers that sound authoritative but were generated
3. 3Misattribution — real quotes attributed to wrong authors, or real authors with invented quotes
4. 4Definition drift — technical terms subtly redefined to fit the model's narrative

Statistical hallucinations are the sneakiest

When an LLM says '47% of clinicians report burnout,' it may be real, adjacent to real (the actual number was 54% from a different study), or entirely fabricated to make a sentence sound sharp. Statistics in model output are the highest-risk claims — verify every one.

Detection techniques

• Ask the model to give confidence ratings per claim, then spot-check low-confidence ones
• Re-prompt the same question in a new session — hallucinations rarely survive regeneration
• Cross-check statistics against official data sources (government, Cochrane, meta-analyses)
• For quotes, paste the exact quoted string into Google with quotation marks
• Use a second model (e.g., Claude checks GPT's output) as an adversarial reviewer

The big idea: hallucinations are not rare edge cases — they are a predictable output of how LLMs generate text. Build verification into every workflow, not just the ones that feel risky.