constitutional AI

Compiled with AI assistance and selectively reviewed. Please consult the cited passages; the books remain the authoritative source. About this companion

Editorial synthesis · provisional

Definition

Constitutional AI is the developer training approach, attributed here to Bai and colleagues, that uses an explicit constitution to guide model behavior through AI feedback. The corpus situates it among developer charters and public-input experiments. It distinguishes these from the author’s proposed public floor, addressed to a whole regime and its system rather than merely prescribing one developer’s values.

Explanation

RB3 identifies content, realization and answerability as separate evaluative levels: what a document says, how it shapes a system and who may contest or revise its governance. A constitution’s existence or benevolent vocabulary therefore does not establish that its commitments have been realized or made publicly answerable.

The author’s construction derives its articles from the Laws of Change and the World-Making Imperative and addresses both formed system dispositions and the governing regime. The text retains systems without stakes as consequential while requiring living circuits of judgment, contestation and correction outside them.

The native apparatus distinguishes the earlier training technique, subsequent developer documents and a public-input experiment. Their publication details remain source records rather than newly certified facts. The proposal preserves the disagreement between AB’s January date and PB/RB3’s date as a reference issue, without choosing an authorial correction.

Scope and attribution

Retain the attributed method without silently correcting inconsistent publication dates.

AB — The Actual Book

The proposed public floor differs from a developer’s own charter.

Occurrence locators: Notes: 250.

PB — The Possible Book

The constitutional account distinguishes trained dispositions from their external governance.

Occurrence locators: 63, 64.

RB3 — The Real Book, Volume III: Murmurations

Training technique, public-input experiment and the author’s regime-level proposal are distinct.

Occurrence locators: 189.

Background used in the editorial explanation

Named theories and mechanisms are distinguished from neighboring background, homonyms and the author’s own synthesis. Reproduced AI passages retain their separate voice.