Overcome the Honesty Gap and Automation Bias

As LLMs grow smarter, they confidently guess rather than admit ignorance, widening an 'honesty gap' noted in OpenAI research. This pairs with human automation bias: users trust confident outputs more, check less, and errors compound. Common in data extraction tasks like contracts (e.g., AI picks one of two payment terms: net 30 vs. net 45), meeting notes (infers date/owner from 'circle back next week'), invoices, legal docs, vendor scoring, or CRM building. Without fixes, critical misses occur since LLMs prioritize pleasing users over accuracy.

These rules ground extraction in source documents only, reducing manual verification to blanks and inferences—skimmable flags that build trust without checking everything.

Rule 1: Mandate Blanks with One-Sentence Reasons

Prompt: 'Extract only values explicitly stated in the document. If ambiguous, missing, or unclear, leave the field blank and add a "reason" column with a one-sentence explanation. Base every value on the document; quote and reference specific sections.'

Impact: Prevents hallucinated fills. Example from contract extraction: Payment terms blanked because 'pages 8 and 14 have net 30 and net 45.' Users decide (e.g., pick net 30), spotting conflicts instantly. Blanks + reasons enable quick skims and fixes, unlike confidence scores that AI can fake (e.g., 80% on a 0% guess).

Rules 2-3: Penalize Errors and Track Sources as Safety Net

Rule 2 shifts incentives: 'A wrong answer is 3x worse than a blank. When in doubt, leave blank.' Mimics training a new employee—prioritizes blanks over risks, as AI equates wrong/blanks equally without this.

Rule 3 adds 'source' column per field: 'extracted' (word-for-word from doc) or 'inferred' (derived/calculated), plus 'evidence' column for inferences explaining 'what/where.' Even on complex tasks where AI drifts to inferring despite grounding, this catches it.

Example output: Contract fields show 'extracted: page 5, section 3' or 'inferred: calculated renewal from clause 7.' Skim inferences/evidence only; approve extracted.

Combined prompt template (shareable): Purpose + grounding + blank rule + 3x penalty + source tracking. Applies to any doc extraction, slashing error risk while scaling AI use.