Reply To: How to prevent AI ‘hallucinations’ when writing research: simple, practical steps

Quick confirmation: Good call on verifying two details (title + DOI). That alone eliminates a large share of hallucinations.

Core problem: Large language models mix facts with plausible-sounding errors. For research writing, that damages credibility and slows acceptance.

Why this matters — short version: One bad citation or invented statistic can cost you reviewer trust, require rework, or sink a grant or publication. Fixing this is a process, not a guess.

What I’ve learned: Treat AI outputs as structured hypotheses — fast drafts that must be validated. That changes the workflow from “trust then edit” to “test then publish.”

What you’ll need

• AI chat access (GPT-style or similar).
• Two verification sources (PubMed/Google Scholar/Scopus or institutional library).
• Reference manager or single verification doc (Zotero, EndNote, or a verification spreadsheet).
• 15–30 minutes per major claim for verification.

Step-by-step workflow (do this each time)

1. Run the AI with the strict prompt below asking for citations, quotes, and a confidence level.
2. Extract top 2–3 claims you plan to use — list them separately in your verification doc.
3. Search trusted databases for title, DOI, or author. Verify at least two data points (title + DOI or title + authors + year).
4. If verification fails, label claim as “unverified” in draft and either remove or mark as speculative.
5. When writing, include only verified citations. Use AI text for drafting language only; keep original quotes and methods from sources.

Metrics to track (KPIs)

• % of AI-cited claims verified (target: 95%+).
• Average time to verify a claim (target: <15 minutes for key claims).
• Number of flagged/unverified claims per paper (goal: 0–1).
• Reviewer objections related to references (goal: zero on first submission).

Common mistakes & fixes

• Mistake: Taking AI citations at face value. Fix: Verify two details immediately.
• Mistake: Using paraphrases as factual claims. Fix: Pull direct quotes or methods sections from the primary source.
• Mistake: Vague AI prompts. Fix: Ask for exact citations, quotes, DOI, and confidence level.

1-week action plan (practical)

1. Day 1: Pick one paper or paragraph to revise. Run the strict prompt below.
2. Day 2: Verify top 3 cited claims in databases; record verification results.
3. Day 3: Rewrite paragraph using only verified citations; preserve quotes and methods excerpts.
4. Days 4–5: Repeat for next paragraph or section; monitor verification time.
5. Days 6–7: Consolidate verified references into your reference manager and prepare for submission.

Copy-paste AI prompt — primary (use as-is)

“You are an expert research assistant. Topic: [insert topic]. Provide a 3–5 sentence factual summary. For each key claim, list up to 3 peer-reviewed studies that directly support it. For each study include: full citation (authors, year, journal), article title, DOI (if available), one direct one-sentence quote from the paper that supports the claim, and a confidence level (high/medium/low) with a one-line reason. If you cannot find supporting studies, say ‘I don’t know’ and list how to verify.”

Strict variant (forces transparency)

“Do not invent citations. If unsure, reply ‘I don’t know’. For topic: [insert topic], return only verified peer-reviewed citations with DOI and a verbatim supporting quote. For each citation include a confidence score and the exact search terms you’d use to verify this in PubMed or Google Scholar.”

What to expect: Faster drafting, slightly more time upfront for verification, near-zero citation errors, and improved reviewer confidence.

Your move.

— Aaron Agius