Who first used the word constitution for an AI?
From a December 2022 research paper by Anthropic, "Constitutional AI: Harmlessness from AI Feedback". Instead of having people rate every answer during training, the model critiques and revises its own answers against a short written list of principles, which the paper called a constitution. Anthropic published the list it used for its Claude models in May 2023, drawing on sources that included the UN Universal Declaration of Human Rights. In January 2026 it replaced that list with a much longer document that explains its reasoning and ranks its priorities, putting safety and ethics ahead of helpfulness.
Other companies now use the word more loosely, for a public statement of how their AI is meant to behave, whether or not it was used in training.
How is a constitution different from terms of service?
| Document | What it covers | Who can see it |
|---|---|---|
| Terms of service | What you and the company agree to, including who is liable for what | You, as a contract you accept |
| Privacy policy | What happens to your data | Anyone; Australian businesses covered by the Privacy Act must have one |
| System prompt | The hidden instructions an app sends its model | Nobody outside the company |
| Constitution | How the AI is meant to behave, and what it puts first | Anyone, which is the point |
What should you read for in one?
The priorities, more than the virtues. Every constitution says the AI should be honest and safe. The useful part is what happens when values collide. When your request conflicts with someone else's interest, which wins? When the AI does not know, does it say so? What will it refuse, and what can it never do? A document that answers those in plain sentences gives you something to hold the product to, and one that only lists good intentions gives you nothing to check.
How would a pilates studio weigh two AI tools by their rules? (example figures)
A pilates studio on the Sunshine Coast spends $1,200 a month on Meta ads and is choosing an AI tool to help with its marketing. Before connecting the ad account, the owner reads what each tool publishes.
| Question | Tool A: terms of service only | Tool B: published principles |
|---|---|---|
| Can it change my ad account? | Terms say it "may take actions on your behalf" | A published limit: it reads and never changes |
| What if it does not know? | Not addressed | Says it will tell you so |
| Whose interest comes first? | Not addressed | Puts the customer's interest ahead of selling more |
| What does it keep? | A 14-page privacy policy | A plain summary and the full policy |
For Tool A, the risk sits in one phrase. A tool that "may take actions on your behalf" could, say, double the daily budget, and two unnoticed weeks of that is about $560 the studio never meant to spend. Tool B's published limit closes that door, and the owner can check the product against it.
Where can you read Cora's constitution?
CoraVision publishes Cora's constitution at coravision.ai/constitution, so the principles Cora works by can be read before you apply for the beta. The limits are set out on CoraVision's AI safety page. Whatever any document says, the working limit is easy to check for yourself: CoraVision reads your Google Ads, Meta and Google Analytics 4 accounts and never changes anything in them, and you make every change yourself.
Questions
Does a constitution guarantee how the AI behaves?
No. It states intent and, where it is used in training, shapes behaviour, but models still make mistakes. What it gives you is a public standard to hold the company to, while limits built into the software, such as read-only access, do the hard stopping.
Is an AI constitution a legal document?
Usually not. The terms of service are the contract. A constitution is a public statement of how the AI is meant to behave, which makes it something customers, journalists and critics can test the product against.