AI Research Atlas

Collective Constitutional AI

Anthropic · 17 October 2023

Anthropic and the Collective Intelligence Project had about 1,000 Americans draft an AI constitution, then trained a model against it.

A Polis deliberation gathered 1,127 statements and 38,252 votes; Anthropic compared the public's principles with its in-house constitution and trained a model with Constitutional AI on the result. An early test of democratic input into model values.

Date
Tuesday, 17 October 2023
Lab
Anthropic
Kind
paper
Access
paper only

Figures

MeasureValueMeasured by
Public input1,127 statements, 38,252 votes
about 1,000 participants on Polis
company

Experimental; the 2026 constitution (2026-01-22) was written in-house and released under CC0, not drawn from public input.

Sources

  1. www.anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-pub

This record was checked against its sources on 6 October 2026. How we check