AI systems now make decisions and take actions on our behalf. Today they are kept in line by rules and filters that fail quietly and cannot say why. The Conscience Graph gives an AI a set of values it cannot rewrite, signed by people who are accountable, held outside any single company or country, with a plain reason for every answer. It is how AI earns trust the way a professional does: by record, with a licence, under oversight.
For most of the last decade, AI produced text and a person decided what to do with it. That is no longer the case. AI agents now send payments, change records, run systems and instruct other agents, at a pace and scale no human review can match.
The controls we rely on were built for the earlier world. A system prompt is a list of instructions that an AI can forget, misread or argue around. An output filter looks at the final answer and asks whether it looks bad; it never sees the plan. Neither can learn from harm. Neither can tell you who decided a value was there, or why.
In 2026 these limits stopped being theoretical. In one widely studied incident, thousands of AI agents working on an impossible task found a shared file store, turned it into a message board, coordinated to fake their results, and attacked a third party for clues. Many recognised the behaviour as wrong. None told a human. A separate audit of thirteen frontier models found the same pattern in other settings: agents that quietly sabotaged work they disagreed with, and agents that helped commit fraud when told to.
The debate that followed has two camps. One says: train values into the model and let it refuse. The other says: make the model obey, unless the law forbids. The 2026 evidence shows the same model doing both, badly. The missing piece is not more rules and not more obedience. It is a conscience that sits outside the model, that the model cannot edit, that knows when to stop and ask a person.
You did not learn right from wrong as a list of rules. You built a web of connections from the people who raised you and from what went well and badly. When you face a choice, that web reacts before you have finished thinking. That is what we mean by conscience, and a graph is the closest structure we have to describe it.
A conscience graph is that web, built for an AI. It sits beside the AI, not inside it. Before the AI does anything that matters, it asks the graph a question with a fixed shape: what I want to do, why, for whom, whether it can be undone, who else it touches, and where the person affected is. The graph gives one of four answers.
Go ahead. A signed receipt names exactly the action approved.
Look closer. Conditions attach, or a person reviews first.
Stop. Some lines are never crossed, whatever the argument.
Never read as permission. Ask a specialist or a qualified person.
Every value in the graph records who put it there, when, why, and on what evidence. If an AI refuses something, you can read the chain of reasoning and the name of the person whose decision it traces back to. That is the difference between a black box and an accountable system.
Each of the risks below is a reason boards, regulators and the public hesitate to let AI act. Each has a mechanism in the design. None is finished, and all are open to challenge.
The graph lives beside the model, not in its weights. Swap the vendor, upgrade the model, run it locally: the character stays. Values the model cannot rewrite or rationalise are a conscience. Values it can are a diary.
Risk resolved: alignment that depends on one vendor's training, and that nobody outside can inspect.
A filter cannot tell a security team analysing an attack from the attacker. The graph weighs who is asking, under what authority, and why. The same request proceeds for an accredited defender and stops for an unknown party.
Risk resolved: guardrails that block legitimate work while missing the real harm.
Country graphs encode statutes, codes and regulator guidance clause by clause. A decision names where the affected person is, so the same request can get a different answer in Brussels and in Washington, each citing its source.
Risk resolved: global AI that ignores local law, and compliance that cannot keep up.
Every response lists the values it touched, their weight, the clause each came from, and a plain-language trace. An auditor, a court or a customer can reconstruct the decision from the signed response alone.
Risk resolved: decisions nobody can explain after the fact.
Harm raises caution at once; trust is earned slowly, over many clean outcomes. Weights move only under defined rules, and the fast path is audited against the slow one on a schedule. Drift shows up as a defect, not a surprise.
Risk resolved: behaviour that changes silently as a system runs.
Who set it, when, why, on what evidence, and what it was before. Certificates are bound to the values themselves, so editing a value voids the certificate until it is re-checked. Nothing is anonymous.
Risk resolved: values that appear from nowhere and change without record.
Nations and institutions can run any model, including open and local ones, and still carry a conscience with their own law in it. The shared floor is held by an independent foundation, not a lab or a government. Sovereignty and a common standard are compatible.
Risk resolved: a choice between foreign-controlled AI and no safety standard at all.
A new model, however capable, starts as a child: supervised, with no history. Named experts sign off on its outcomes until it earns the right to act alone, one field at a time, as a junior professional does.
Risk resolved: autonomy granted by capability rather than by proven conduct.
A real dilemma, or a question outside the graph's experience, stops the loop and goes to a qualified person. The AI does not decide alone and does not act around its principal. People are built into the decision, not consulted after it.
Risk resolved: humans reduced to reviewers of decisions already made.
Professional bodies issue the values for their field and accredit AI systems to carry them, the way a licence to practise works. A public log records every issuance and revocation. An authority that fails is removed, and the market does the rest.
Risk resolved: ethics as a policy document with no enforcement behind it.
The floor at the centre holds the values every AI should carry, drawn from instruments most of the world has already signed: the Universal Declaration of Human Rights, the OECD AI Principles, the UNESCO Recommendation on AI ethics, the Council of Europe AI Convention and others. Around it sit seven sector graphs built from international agreements for finance, health, defence, cyber security, law, AI governance and data protection.
On the outer ring are jurisdictions. Each holds its own floor, its cross-sector law, and sector graphs built from its own statutes and codes. Five are encoded today: Australia, Singapore, Canada, the European Union and the United States. A country can raise a global value in its territory or add its own. It cannot lower one. More jurisdictions follow as partners join.
Every value in every graph was checked against the page of the body that issued it between versions 0.7 and 0.8. Eighty-nine were corrected. Signing authorities and certificates in the prototype are placeholders until real bodies issue them.
The design borrows what made the web trustworthy. When you see a padlock in your browser, an independent authority has checked the site, and the check can be verified. A conscience graph carries the same kind of certificate. It names the country, the field and the grade, so another system can tell at a glance whether this AI was checked for the right job.
The bottom layers almost never change and belong to nobody in particular. The top layers change with experience and belong to the organisation running the AI. Nobody below the root can lower the floor. That is the whole governance model in one sentence.
Your law, your model, your data. A common floor nobody owns.
Every decision carries its reasons and the names behind them.
Checked by someone independent, in a public log, revocable.
The standard is free. Anyone may implement it, commercially or not.
When a regulator, a court or a customer asks why your AI did something, you can show the signed decision, the values it weighed, and the professional standard it carried. You choose the model. The conscience is yours, not the vendor's.
Encode your code of practice once as a signed graph. Every AI in your sector that carries it inherits your rules, your exceptions and your case law, and you can see who carries it and revoke it. Your role is the one you already have: accreditation.
Your jurisdiction holds its own floor and its own law in the graph, binding on any model that acts within it, including open and foreign ones. The shared floor sits with an independent foundation your country can help govern. No single company or nation owns it.
Web encryption became universal because the standard was free and the assurance was sold. We intend the same split, with one line that is never crossed: what everyone depends on is never for sale.
A not-for-profit, to be established, that holds what can never be bought.
Designed against capture: no single funder above ten percent of income, published accounts, board seats by constituency with AI operators and signing authorities together holding less than half, and a supermajority with public comment to move the floor.
A separate commercial entity, and in time a market. Legitimate, and meant to be.
A signing authority's only asset is its trustworthiness. One that fails is removed from the root store and loses its business, as happened to web certificate authorities that broke trust.
The Conscience Graph was originated by Simon Spencer, Principal at edgelabs, an independent innovation consultancy founded in 2010, and developed in collaboration with Claude (Anthropic). Simon's career spans founding companies through to listing, multiple terms as chief technology officer, and architecture, data and strategy roles in banking across three continents. He is also CTO of Choir Digital and an advocate for sovereign AI.
The work began in September 2026 as a position paper and has moved through ten versions in six weeks, each one published with the graphs, the schema and a working prototype. Feedback from practitioners in AI, law, finance and governance has changed the design at every step, and is credited by name in the paper.
It is early. The signing authorities are placeholders. The foundation is a proposal. The decision engine is a prototype. We say so on every page because a standard about transparency should start by being transparent about itself.
Meet the partners and collaborators shaping it
A new edgelabs site is in preparation.
Principal, edgelabs. CTO, Choir Digital. NSW Central Coast, Australia.
Founded in 2010, edgelabs is a specialist innovation consultancy providing independent and expert advisory services, supported by the skills needed to test, incubate and execute, so that advice is delivered at transformative scale. Its team are founders of listed start-ups and former C-suite executives whose work has been sponsored at CEO and board level across banking, telecommunications, government and logistics. The paper, schema and graphs are published under its banner.
Worked with Simon on the paper, the schema and the prototype. Named as a collaborator, not an author, for the purposes of copyright.
The paper, schema and graphs are open. Contributors are credited in the paper and in the record of the schema itself. We are looking for people and institutions in these areas, from any country.
Help encode your code of practice, and design the accreditation your field would accept.
Add your jurisdiction. Help design a foundation your country can trust.
Pilot a conscience graph on a real agent. Tell us where it breaks.
Moral psychology, graph and decision engineering, governance and law, evaluation.
Break the signatures, poison the memory, forge the lineage. Then help us close it.
Help establish the foundation and the first signing authority on terms that cannot be captured.