Bias and Hallucination Checks for AI-Aided Drafts

Reading time ~6 minutes · Published October 22, 2025

Pre-flight sanity

• Model provenance: Note the model and date used. Re-run key prompts with a second model for convergence.
• Prompt discipline: Keep the brief in front of the model. Add a “cite-or-omit” instruction when asking for references.
• Scope boundary: Confirm the intended audience, target journal, and allowed claims before editing further.

Fact checks (numbers, names, claims)

• Named entities: Verify spellings, affiliations, trial names, datasets, and grant numbers against primary sources.
• Numbers: Recompute simple rates, totals, and percentages. For statistics, confirm test choice, effect sizes, and CIs match the described design.
• Quotes/paraphrases: If a direct quote is included, confirm the exact wording and page/figure location in the source.

Citation cross-checks

• Existence: For every reference, confirm it exists and resolves to the correct work (title, authors, year).
• DOI resolution: Look up or validate the DOI via Crossref; paste the citation or title and confirm a match. Link to the DOI in the reference. (Crossref search)
• Biomedical: Cross-check PubMed entry and PMID; confirm article type and publication status. (PubMed)
• Open discovery: Use OpenAlex to sanity-check venue, year, and related works. (OpenAlex)
• Retractions: Screen references for retractions, expressions of concern, or corrections. (Retraction Watch Database)
• Mapping: Ensure one-to-one mapping between in-text citations and the reference list; no orphaned entries.

Bias controls

• Balance of evidence: Add at least one high-quality counterpoint source if the topic is contested.
• Sampling bias: Avoid over-reliance on a single geography, lab, or time period when summarising literature.
• Sensitive domains: For clinical, legal, or safety-critical content, add a plain-language limitation statement and route high-risk claims to expert review.

Reasoning & scope checks

• Claim-evidence alignment: Every claim should point to a verify-able source or a clearly marked hypothesis.
• Specification creep: Remove features or results the study did not measure. Separate exploratory from confirmatory analyses.
• Terminology consistency: Standardise acronyms and units; keep them consistent across text, figures, and tables.

Language & framing checks

• Hedging: Use calibrated language when evidence is preliminary; avoid definitive verbs for unconfirmed findings.
• Ambiguity: Replace ambiguous pronouns and long nested clauses with clear, specific statements.
• Disclosure: Add data/code availability, funding, and competing interest statements as required by the target venue.

Packaging for submission

• Tracked changes: Provide a clean file and a tracked-changes file with line numbers.
• Response letter: Use point-by-point structure; quote the reviewer in bold, respond in plain text, and cite exact line numbers for edits.
• Figure audit: Correct DPI, font embedding, colour-blind-safe palettes, and self-contained legends with units.