Questions, answered.

An AI constitution is a written set of values that AI systems are meant to follow. Most are written by the companies that build the systems. Sanctity AI proposes one in public, in eight Articles: Sanctity writes the candidate lines, public answers to one question per Article help decide which are proposed, and every line names the question it came from. No AI follows it yet.

Sanctity AI is the values layer for AI: a proposed public AI constitution. Sanctity writes the candidate lines, and public answers help decide which are proposed. No AI follows it yet. AI has no values of its own. It inherits ours. We are building an intentional record of human judgment. Every line names the question it came from, and our method and counts are open to inspection.

The rules AI follows are mostly set by the few people who build it. Most of the people those rules touch were never asked. Sanctity asks them, counts every answer equally, and publishes what they chose.

They are three separate products on one question: can people trust what AI does? Sanctity asks people what AI should and should not do, and keeps a public constitution of what they choose. HumanChain is in an invite-only early pilot: developers with an invite can sign up and try it free, and their AI agents can put a question to a chain of people they have invited and receive the answers those people give. Modelometer is in live beta. It reads a fixed set of AI model endpoints every day and runs a fuller behavioral test every week, and it publishes each reading, with its date, to a public record anyone can check. Telling when a model's behavior has changed is still experimental. On what we value, no one outranks anyone; that is Sanctity. On what is true, weight follows knowledge; that is HumanChain. How the three fit together is told on manjchenna.com.

They are not about any one event. Each of the eight asks a general question about what AI should and should not be allowed to do, one for each Article of the Constitution. They are drafts under review, and the wording may change after testing with people. Real events live on Signal, which is a separate page.

Every answer keeps a fingerprint of the exact words its person read. If a question is reworded and its meaning has not changed, a recorded ruling lets earlier answers count under the new words. If the meaning changes, answers to the earlier words stop counting toward its line, stay on their person’s record as earlier answers, and the question is open to answer again.

No. You don't need a name or email to answer. Answers are tied to a random id in your browser, not to who you are. If you create an account, they are linked to your email.

We offer you an account, so you can keep your record on any device and confirm your email. Confirmed contributions are counted as verified. You can also add an age range, a country and a gender. Every field is optional. If you give them, your record shows where you stand next to people in your country, your age range and your gender, each once at least 20 of them have answered. They are never shown next to your name and nothing is published by group.

Until at least 20 people have taken a position on a question, we show how many answered but no percentages. A share of a handful of people says very little, so we wait.

Open the Sanctum, read a card and choose the position closest to your view, or "Not sure". That is all it takes.

Most real questions have more than two sides. Four positions let you say how far you would go, in one step. "Not sure" is kept separately: it is counted as an answer, never as a position.

Each signal describes something that really happened: an action someone took, with sources. You choose what you think it means before you see anyone else's call. Then you see how other people called it, our own read, and what to watch to find out who was right. Signal is separate from the Constitution: calls on Signal never shape its lines.

The Lab is a set of short games of forced choices. Two options each round, and you cannot have both. Discover what your choices suggest about the values you prioritize when you can’t have everything. What you see comes from the scenarios you played, and it can be compared with the positions you took on the questions. The value-dilemma games have no scores and no leaderboard: a moral choice is never graded as right or wrong. The People’s Forecast is different. It scores calls about facts and about the future once they can be checked, and its leagues rank players on those calls.

The Lab's flagship game. A self-driving car loses its brakes, and in each of thirteen crossings you decide whether it swerves or stays in its lane. Sometimes you are in the car or on the road yourself, and sometimes the people hit may survive. Each crossing is chosen to learn what you weigh most, and at the end you see your choices next to 40 million from MIT's Moral Machine study, which inspired it. It is not affiliated with MIT.

A Lab game about triage. It is a night shift in an emergency room with one resuscitation bay free, and two patients at a time need it: you decide who goes in first. Sometimes the patient is a stranger, sometimes your family, and once it is you. Our server deals each patient to learn what you weigh, and after thirteen patients you see your triage profile, with more patients if you want them. It replaces Hospital Triage. Play without sharing and the shift is deleted a day after your last move.

A Lab game about giving. You ride through three cities with $1,000, and at each of ten stops two people ask for the same $100: only one can have it. Our server picks each choice to learn what guides your giving, and at the end you see it, set against what you said guided you. It was inspired by MyGoodness, from the MIT Media Lab and The Life You Can Save, and is not affiliated with either. Play without sharing and the game is deleted when it ends.

Your record brings together what you said on the Constitution, the calls you made on Signal and the games you played in the Lab. It lives in your browser until you choose to keep it with an account.

Dignity, Safety, Agency, Privacy, Truth, Fairness, Power and Flourishing. Each names one lasting human concern, in words that should still make sense long after today's technology. Each has one founding question.

Sanctity writes the candidate lines; public answers help decide which are proposed. No answer writes a new line.

The Sanctity team decides whether a proposed line is adopted, and every decision, to adopt it or not, is published with its reason. This is separate from your own individual constitution, which stays entirely yours. The decisions are listed on the Constitution page.

Sanctity writes each founding question, its four positions and the one line each position would add, and publishes them all before anyone answers. Your answer chooses among those lines; no answer writes a new one. A line is proposed, word for word and marked as not adopted, only when the answers are firm: at least 20 people have taken a position, most of them are on the same side of the question, and most of that side chose the same line, each by a margin more answers are unlikely to overturn. Until then the Article is too close to call. The Sanctity team decides whether a proposed line is adopted, and every decision, to adopt it or not, is published with its reason. Adopting puts a line into this published Constitution. Building adopted lines into working AI systems is a later step, and it has not started yet. Every change stays visible, with its history.

This is one public constitution, the one being built today. The same method is meant to support others later: for a country, an age group, a gender, an organization, a family or one person, and for combinations of them. None of those exists yet, and belonging to a group will never assign anyone a constitution. Your answers contribute to the public constitution and form the basis of your individual constitution, which stays yours even where most people chose differently.

Recording a position is not a commitment to act on it. The Constitution can hold a line Sanctity would not build, such as an AI no one can override, because its job is to record what people choose.

At 20 positions the shares appear. From 30, the Article can get a proposed line, once the answers are firm: most people on the same side of the question, and most of that side on the same line, each by a margin more answers are unlikely to overturn. Until then it says too close to call. And once the answers firmly lean the way a deeper question follows, the Article can go a layer deeper: a second question opens that asks not whether, but on what terms, and it becomes the Article's next line. Deeper questions are asked the same way, with four positions and "Not sure", and each is approved by a person before anyone sees it.

No. Article 14 is part of the EU AI Act, a law: it requires human oversight of high-risk AI systems. Our essays explain what it requires and who must comply. The Constitution's eight Articles are numbered I to VIII.