ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›AI MEANINGS›WHAT IS CONSTITUTIONAL AI?
AI MEANINGS · MEANINGINTERMEDIATE

What Is Constitutional AI? Meaning in Plain English

Constitutional AI trains a chatbot with a written list of principles that the AI uses to check and fix its own answers. Anthropic developed it for Claude.

READ6 min
WORDS1,141
SECTIONS7
TYPEMEANING
CHECKED12 OCT 26

Constitutional AI is a way of training an AI model to behave well by giving it a written list of principles, a “constitution”, and having the AI check and correct its own answers against those principles, instead of relying mainly on people to label harmful answers. Anthropic developed it for its Claude models.

◈ IN PLAIN TERMS

Picture a trainee who is handed the company handbook and told: write your reply, reread it against the handbook, then rewrite anything that breaks the rules. A manager does not need to stand over every reply. Constitutional AI does the same for a chatbot: the handbook is the constitution, and the AI does much of the checking itself.

In a sentence: “The chatbot explained why it would not help with that request instead of just refusing, which the article put down to constitutional AI training.”

What does constitutional AI mean?

The plain point: the main human input is a list of written rules; the AI uses them to judge and improve its own answers.

Most chatbots learn manners partly from people rating thousands of answers. Constitutional AI keeps people in charge of the rules but lets an AI do much of the rating. Anthropic says the method makes it possible “to control AI behavior more precisely and with far fewer human labels.”1 That is Anthropic’s own description of its own method. (Anthropic’s tools are used behind parts of this site.)

Where the word comes from

The name comes from a paper by Anthropic researchers (first author Yuntao Bai), posted on 15 December 2022. Because “The only human oversight is provided through a list of rules or principles”, they called the method constitutional AI.2

How does it work?

The plain point: two stages, both using AI feedback guided by the written principles.

  1. Critique and rewrite. The model answers a prompt, writes a critique of its own answer against the principles, then revises it. The model is then trained on the improved answers. Anthropic describes this stage as: “we sample from an initial model, then generate self-critiques and revisions”.1
  2. AI picks the better answer. An AI model compares pairs of answers and says which fits the principles better. Those choices train a scoring model, which then guides further training. The paper calls this “RL from AI Feedback” (RLAIF).1 “RL” is reinforcement learning: learning from rewards.

The researchers report that the result was “a harmless but non-evasive AI assistant”: one that explains its objection to a harmful request rather than just shutting the conversation down.1

TAKEAWAY

Constitutional AI swaps much of the human rating work for an AI that grades itself against written rules. People still choose the rules, so people still decide the values.

What is in Claude’s constitution?

The plain point: a long public document setting out four priorities, in order.

Anthropic published a new constitution for Claude on 22 January 2026, saying its previous one “was composed of a list of standalone principles” and that the new one explains its reasoning instead.3 The document asks Claude to be, in priority order when they conflict: broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. Anthropic says it “plays a crucial role in our training process”.4

Anthropic also says plainly that Claude’s “outputs might not always adhere to the constitution’s ideals.”3 A constitution describes what the company is aiming for; it is not a guarantee.

Often confused with

  • RLHF: people rate answers and the model learns from their choices. Constitutional AI replaces much of that rating with AI feedback guided by written rules. See RLHF meaning.
  • A system prompt: instructions given to a model each time it runs. A constitution is used during training, so it shapes the model itself. See system prompt meaning.
  • A country’s constitution or a law: Claude’s constitution is a company document, not a legal one. For laws, see AI governance meaning.
  • Alignment: the wider goal of making AI do what people intend. Constitutional AI is one method for it. See AI alignment meaning.

Questions people ask

Who writes the constitution?

Anthropic does. In a 2023 experiment it also tried a “public input process involving ~1,000 Americans” to draft a constitution, noting that Claude “currently relies on a constitution curated by Anthropic employees.”5

Can anyone read or reuse Claude’s constitution?

Yes. Anthropic publishes it on its website under a CC0 public-domain licence, saying it “can be freely used by anyone for any purpose without asking for permission.”4

Does constitutional AI make a chatbot safe?

It makes certain behaviour more likely, not certain. Anthropic itself says outputs may not always follow the constitution, and people still find ways around trained rules, as jailbreak meaning explains.

Learn more on this site

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What is constitutional AI?

  • Training an AI to check and correct its own answers against a written list of principles
  • A law that governments pass to control AI
  • Instructions typed to a chatbot at the start of each chat
Show the answer

Answer: Training an AI to check and correct its own answers against a written list of principles. Constitutional AI gives the model a written list of principles and has it critique and improve its own answers against them during training.

2. In constitutional AI, who still decides the values?

  • The AI decides its own values
  • Nobody: the values come from the internet
  • People, because they choose the rules
Show the answer

Answer: People, because they choose the rules. The AI does much of the grading, but people still write the rules, so people still decide the values.

3. Does constitutional AI make a chatbot safe?

  • Yes, it guarantees the chatbot always follows its rules
  • It makes good behaviour more likely, not certain
  • No, it has no effect on behaviour
Show the answer

Answer: It makes good behaviour more likely, not certain. The page says it makes certain behaviour more likely, not certain, and Anthropic itself says outputs may not always follow the constitution.

Related terms: Checkpoint · Clustering · Computer vision · Distillation · Emergent abilities · Fine-tuning

SOURCES
  1. 1Anthropic, Constitutional AI: Harmlessness from AI Feedback, 15 Dec 2022, read at source 12 Oct 2026“These methods make it possible to control AI behavior more precisely and with far fewer human labels.” ↑a“In the supervised phase we sample from an initial model, then generate self-critiques and revisions” ↑b“We then train with RL using the preference model as the reward signal, i.e. we use ‘RL from AI Feedback’ (RLAIF).” ↑c“we are able to train a harmless but non-evasive AI assistant” ↑d
  2. 2Bai et al. (Anthropic), Constitutional AI: Harmlessness from AI Feedback, arXiv:2212.08073, submitted 15 Dec 2022, read at source 12 Oct 2026“The only human oversight is provided through a list of rules or principles”; the abstract continues that the authors therefore refer to the method as ‘Constitutional AI’. ↑
  3. 3Anthropic, Claude’s new constitution, 22 Jan 2026, read at source 12 Oct 2026“Our previous Constitution was composed of a list of standalone principles.” … “we need to explain this to them rather than merely specify what we want them to do.” ↑a“Claude’s outputs might not always adhere to the constitution’s ideals.” ↑b
  4. 4Anthropic, Claude’s Constitution, undated page, read at source 12 Oct 2026“It plays a crucial role in our training process, and its content directly shapes Claude’s behavior.” … “In cases of apparent conflict, Claude should generally prioritize these properties in the order in which they’re listed”, listed as broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. ↑acites a “Creative Commons CC0 1.0 Deed” licence; the document “can be freely used by anyone for any purpose without asking for permission.” ↑b
  5. 5Anthropic, Collective Constitutional AI: Aligning a Language Model with Public Input, 17 Oct 2023, read at source 12 Oct 2026“public input process involving ~1,000 Americans to draft a constitution for an AI system.” … “Anthropic’s language model Claude currently relies on a constitution curated by Anthropic employees.” ↑

Bai et al. (Anthropic), Constitutional AI: Harmlessness from AI Feedback, arXiv:2212.08073, 15 Dec 2022, read at source 12 Oct 2026.

The comparison on this page (a trainee and the company handbook) and the example sentence are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE