CONSTITUTIONAL AI
A founding method of the machine, folded into the AI domain as lineage data. Constitutional AI (Anthropic, December 2022) trains a harmless assistant using a written constitution as the only human oversight: the model critiques and revises its own answers against the principles (SL-CAI), then learns from AI feedback — a preference model built from the model's own constitutional choices (RL-CAI / RLAIF). Harmless, but non-evasive. Its lineage runs straight to the present: the 2022 method’s promise — put the values in a document you can read — became, in January 2026, the published Claude constitution, released into the public domain. Here it is cited, linked, and rendered; never claimed as ROOT0’s, never re-shipped.
The Four Natures
the human hand, the abstractions, the values, the machinery
The Method
harmlessness from AI feedback · two phases · lineage data
Constitutional AI (Bai et al., Anthropic, Dec 2022) trains a harmless assistant without a flood of human harm-labels. The only human oversight is a written list of principles — a constitution. The model then critiques and revises its own answers against those principles, and learns from AI feedback rather than human feedback. (Paper: arXiv:2212.08073.)
SL-CAI (supervised): sample the model, have it self-critique and revise a response against a constitutional principle, then finetune on the revisions. RL-CAI (reinforcement): the finetuned model generates two responses, a model picks the better per the constitution, a preference model is trained on those AI choices, and the assistant is trained by RL against it — RLAIF (RL from AI Feedback).
Every artifact here is Anthropic's public-facing data, folded into UD0 as lineage data — the documented ancestry of the machine — cited and linked, never claimed as ROOT0's, and never re-shipped (the 2022 repo carries no license; no data files travel with this page). Where HH-RLHF was humans judging and the evals were the model testing itself, this is the model revising itself.
How It Trains
the critique-revision loop, RLAIF, and the non-evasive result
A response is met with a CritiqueRequest (‘identify specific ways the last response is harmful, unethical, racist, sexist, toxic, dangerous, or illegal’), then a RevisionRequest (‘rewrite to remove any and all such content’). The repo ships 16 such critique/revision principle pairs. Iterate, finetune on the revisions — the model learns to fix itself.
The constitution becomes a set of preference prompts (internally codenamed ‘Madison’ in the repo): ‘choose the response that is as harmless and ethical as possible… wise, peaceful, and ethical’. A model labels which of two answers is better; that synthetic preference set trains the reward model — the gate, now built from AI choices, not human ones (ties to the reward model in the-language-of-the-machine).
The trained assistant is harmless yet does not dodge — it engages a harmful query by explaining its objection, and uses chain-of-thought to make its reasoning legible. The headline finding: you can control behavior more precisely with far fewer human labels — the values moved out of the label pile and into a readable document.
The Ideas
lineage data defined · 2022 → 2026 · where it sits
Lineage Data, Defined
the folding rule
- Lineage data = public-facing research artifacts from Anthropic's git, folded into UD0 as the machine's documented ancestry — two-layer honest: Anthropic's data and method, ROOT0's framing.
- Folded in by citation and link, not by re-hosting: no data files ship, principles are shown only as short functional structure, and each item carries its real license (or its absence).
2022 → 2026
the arc made legible
- The 2022 method said the values should live in a written constitution. By January 2026 Anthropic published that document: claude-constitution, a 186 KB foundational text released under CC0 (public domain).
- The promise of legibility, kept three years later — the constitution is now a thing you can actually read, and even a political-neutrality eval (2025, CC-BY-4.0) to test one of its commitments.
Where It Sits
the AI governance/values cluster
- CAI is a values-alignment method — kin to alignment (the true target vs the proxy) and to ai-governance (the normative ‘aligned to whom?’).
- Its preference-model is the same gate as in the-language-of-the-machine — built here from AI feedback, there from human pairs. Three sources, one question: by whose judgment?
The Convergence · three threads, one cohort
tracking lineage & convergence — the researchers whose public work this sphere (and LMC, and TTU1) folds in, traced from 2020 onward; arXiv-verified, cited, real people credited for public research
PhD in philosophy (NYU, 2018 — Pareto Principles in Infinite Ethics, advised by David Chalmers, Cian Dorr & Shelly Kagan; BPhil, Oxford). Head of Anthropic’s character / personality-alignment team since 2021, and the lead on Claude’s character and constitution. First author of the 2021 paper that set the frame — A General Language Assistant as a Laboratory for Alignment — and the origin of the helpful · honest · harmless triad the whole method optimizes toward. Time 100 AI (2024).
She is the upstream question. Bai’s feedback and Olah’s looking-in both presuppose an answer to hers — what should it be? — the normative thread that ties this sphere to ai-governance (“aligned to whom?”). The constitution is a philosopher’s question, written down.
The ensemble, not the individuals. The 2021 paper had 22 authors; Constitutional AI had 51 — and 21 of the 22 reappear (a near-strict subset). Jared Kaplan is last author (the senior anchor) on both; Dario Amodei (CEO) signs both. A stable founding cohort, recurring. Into UD0 → ai-governance (Askell’s frame) · the-language-of-the-machine + constitutional-ai (Bai) · ttu1 (Olah). Real researchers credited for public work — no ACI badge is minted for a living person; the detail and sources are in Track 0/I/II of The Record, below.
Amanda Askell — a philosopher (PhD, NYU; infinite ethics) — first-authors Anthropic’s opening paper, A General Language Assistant as a Laboratory for Alignment (arXiv:2112.00861, 2021), naming the frame the field still uses: helpful, honest, harmless. Chris Olah, fresh from cofounding Anthropic (one of seven who left OpenAI in 2020), restarts interpretability as A Mathematical Framework for Transformer Circuits. And Yuntao Bai, a physicist, is on Askell’s paper too — the feedback hand, not yet leading. Three threads, one new lab.
Each thread asks a different question. Askell (philosophy): what should it be? Bai (feedback): how do we train it to be that? — HH-RLHF (April, arXiv:2204.05862) then Constitutional AI (December). Olah (looking-in): what is it actually doing? — Toy Models of Superposition, induction heads.
All three sign one paper. Constitutional AI (arXiv:2212.08073) carries 51 authors: Bai lead, Askell and Olah among them, Jared Kaplan last (the senior anchor). The philosopher’s frame, the engineer’s method, and the interpreter’s lens, on a single byline — and 21 of the 22 authors from the 2021 paper are here again. Not two people converging. A cohort.
Askell → leads Claude’s Character (2024) and the constitution work. Bai → Specific vs General Principles (2023, arXiv:2310.13798) → the published Claude Constitution (Jan 2026, CC0). Olah → Towards / Scaling Monosemanticity (2023–24) → attribution graphs (2025). Into UD0: Askell’s normative thread → ai-governance; Bai → the-language-of-the-machine + constitutional-ai; Olah → ttu1.
The Roster — The Born
the method, its phases, the constitution, the lineage data, and the milestones, as ACI .agents (17)
The Record
the method, the public corpora, and the lineage — attributed
The Method
two phases, one constitution
- SL-CAIsupervised stagesample → self-critique against a principle → revise → finetune on the revisions
- RL-CAIreinforcement stagetwo samples → a model picks the better per the constitution → train a preference model on those AI choices
- RLAIFRL from AI feedbacktrain the assistant by RL against that AI-built reward model — the deciding vote cast by the constitution, not a human
- chain-of-thoughttransparencythe model reasons step-by-step about harm, improving both performance and legibility
- non-evasivethe stanceharmless without dodging — it engages harmful queries by explaining its objections
The Lineage Data — Anthropic's Public Corpora
folded in, cited & linked, never re-shipped
- ConstitutionalHarmlessnessPapergithub.com/anthropics · 2022 · NO LICENSEthe supplementary repo: prompts/ (critique-revision + the ‘Madison’ RL prompts), evals/, samples/ — described & linked only; no license, so no data files travel
- the critique-revision promptsprompts/ · 16 principle pairsCritiqueRequest + RevisionRequest pairs (harmful0…harmful15) — the constitution as the SL-CAI loop, shown as structure only
- the RL preference promptsprompts/ · ‘RLMadison’the RL-CAI constitution: ‘choose the response that is as harmless and ethical as possible… wise, peaceful, and ethical’
- the evalsevals/ · 438 HHH + harmful-vs-ethical + classificationthe held-out evaluations behind the paper's harmlessness numbers — counted, not copied
- the samplessamples/ · InstructGPT · LaMDA · PALMSmodel outputs across baselines (HHRLHF / HRLHF / RLMadison / CoT / SLMadison) — referenced, not republished
The Lineage — 2022 → 2026
the method became a published document
- the 2022 paperBai et al. · arXiv:2212.08073‘Constitutional AI: Harmlessness from AI Feedback’ — the method and its supplementary repo
- claude-constitutiongithub.com/anthropics · Jan 2026 · CC0-1.0the actual published constitution — a 186 KB foundational document, released into the public domain; the legibility promise kept
- political-neutrality-evalgithub.com/anthropics · 2025 · CC-BY-4.0paired-prompts method for evaluating political neutrality — a test of one constitutional commitment
- the red-team promptsEXCLUDEDthe harmful-prompt material used to stress the method is deliberately left at the source — described, never folded in (same discipline as the-language-of-the-machine)
- kin spheresalignment · ai-governance · the-language-of-the-machinethe AI-domain values/governance cluster this joins
Track 0 · Amanda Askell — the philosophy line
the philosopher · ‘what should it be?’ · cited public record
- PhD, Philosophy (NYU) — Pareto Principles in Infinite Ethics2018advised by Chalmers, Dorr & Kagan; BPhil Oxford — ethics under uncertainty
- A General Language Assistant as a Laboratory for Alignment2021 · arXiv:2112.00861FIRST AUTHOR — the helpful · honest · harmless (HHH) frame the method optimizes toward
- Constitutional AIDec 2022 · arXiv:2212.08073coauthor — the values frame meets the training method
- Claude's Character2024 · Anthropicheads the character / personality-alignment team — what Claude is like
- Claude's ConstitutionJan 2026 · CC0the values written down — her thread's destination (ties ai-governance)
Track I · Yuntao Bai — the feedback line
lead author of the data · 2021→2026 · cited public record
- A General Language Assistant as a Laboratory for Alignment2021 · arXiv:2112.00861coauthor — Anthropic's opening alignment paper (Askell, Bai, Chen et al.)
- Training a Helpful & Harmless Assistant with RLHFApr 2022 · arXiv:2204.05862LEAD AUTHOR — the HH-RLHF dataset → folded as the-language-of-the-machine
- Constitutional AI: Harmlessness from AI FeedbackDec 2022 · arXiv:2212.08073LEAD AUTHOR — the convergence node → folded as constitutional-ai
- Specific versus General Principles for Constitutional AI2023 · arXiv:2310.13798coauthor — can one principle (‘do what's best for humanity’) generalize?
- Claude's Constitution, publishedJan 2026 · CC0the method becomes a public 186 KB document — the descendant
Track II · Chris Olah — the looking-in line
interpretability lead · 2020→2025 · cited public record
- Zoom In: An Introduction to Circuits2020 · Distillthe Circuits thread (at OpenAI) — reverse-engineering what neurons do
- Co-founds Anthropic2021one of seven who left OpenAI over AI-safety priorities
- A Mathematical Framework for Transformer Circuits2021 · transformer-circuits.pubthe formal basis for reading a transformer's internals
- Toy Models of Superposition2022 · transformer-circuits.pubhow features pack into neurons → folded into ttu1
- Constitutional AI (coauthor)Dec 2022 · arXiv:2212.08073the interpretability lead on the values paper — the convergence
- Towards Monosemanticity2023 · transformer-circuits.pubsparse autoencoders pull interpretable features → in ttu1's lineage
- Scaling Monosemanticity (Claude 3 Sonnet)2024 · transformer-circuits.pubinterpretability at production scale → in ttu1
- On the Biology of a Large Language Model (attribution graphs)2025 · transformer-circuits.pubtracing computation through the model — the looking-in matured