Max Ghenis MaxGhenis

Can LLMs Forecast Human Value Evolution? (And Why We Can't Tell Yet)

Draft for EA Forum — feedback welcome

The Hypothesis

Human values change. Sometimes predictably — generational replacement, exposure effects, information cascades. If LLMs have learned these patterns from their training data, they might predict value trajectories better than simple extrapolation.

This matters for alignment. If we could forecast where human values are heading, we'd have a tool for:

Anticipating moral circle expansion

Using PolicyEngine from Stata

Three approaches for integrating PolicyEngine tax-benefit calculations with Stata:

1. Direct Python integration (Stata 16+)

Stata 16+ has native Python integration. Install policyengine-us and call it directly:

python:

	<!DOCTYPE html>
	<html>
	<head>
	<meta charset="UTF-8">
	<title>Complete .rac Format with Versioning</title>
	<style>
	* { box-sizing: border-box; margin: 0; padding: 0; }
	body {
	font-family: 'SF Mono', 'Fira Code', monospace;
	background: #0a0a0f;

	"""
	85th Percentile Household Income in Los Angeles City
	Using 1-year ACS PUMS data for 2023 and 2024

	LA City PUMAs identified from Census ACS API by name matching.
	These 23 PUMAs are explicitly labeled as LA City areas.
	Note: Some LA City areas may be in shared PUMAs (e.g., 03707, 03748) which
	include portions of other cities - these are excluded to avoid overestimation.
	"""