Using AI To Research The Missing Heritability Post

A careful field report on Scott's first heavy use of AI (o3) as a research assistant for the Missing Heritability post — 'an unprecedented combination of brilliant and mendacious; too useful to avoid but too unstable to fully trust.' The good: o3 derived Mercatus's overhead from a 64-page Form 990 by dividing three numbers on page 10, resolved his genetic-nurture confusion with a math model plus five cited studies, and tabulated adoption studies with n/age/h². The bad: it boldly hallucinated a 'Section 3.2' and verbatim quotes in Rushton & Jensen and a nonexistent Appendix A; made a calculation error yielding negative suicide rates; and showed prompt-cued source-bias (mentioning 'genetic nurture' flipped it into anti-hereditarian sycophancy). Ends with practical tips (always check links; cross-check two AIs without leading questions; 'explain like I'm Zvi') and ties reliability to AI 2027's honesty dynamics.

Quality 73 · Strong. Strong band, upper. An honest, well-exemplified account of the actual texture of AI-assisted research in 2025, with real observations ('no grad student is both this smart and this dumb'; sycophantic source-bias from prompt cues) and actionable guidance. Among the best of this paid set; just below the Excellent line because it is a practical field report rather than a deep reframing.

Claude’s paradigm shift 42 · Moderate. Moderate. Timely and his specific observations are fresh, but the broad theme (LLMs are useful yet hallucinate) was widely discussed by mid-2025.

Real-world impact 3 · Moderate. An honest, well-exemplified field report on the actual texture of AI-assisted research in 2025 (o3 as 'brilliant and mendacious') with real observations and actionable guidance. Real topical relevance to how research is changing, but a practical write-up informing the discourse — modest RWI.