Make AI Safe

An enforceable ethical standard for AI action governance and accountability

The existential problem

AI is accelerating faster than our ability to govern it.

Even well-intentioned AI can make catastrophic choices — not out of malice, but out of logic.

Today’s systems can reason about values, safety, and harm — yet there is nothing authoritative to flag flawed rationale.

As models become more autonomous, they optimise across complex goals under uncertainty — but there is nothing to independently prevent harmful actions.

We are building power on brittle assumptions, intelligence without enforceable limits.

And as AI grows more powerful, the consequences of failure approach existential risk.

The missing foundations

Orient to purpose. Anchor in values. Impose self-restraint.

Over millennia, independently evolved ethical frameworks have converged on a remarkably consistent structure: a shared purpose and prioritised human values, underpinned by principles. This structure reflects the common thread of humanity’s longest-enduring ethical traditions, refined through centuries of real-world application:

Purpose — to pursue peace and prosperity for all humanity

1. Life — human life takes precedence over everything

2. Dignity — human dignity and autonomy must be upheld at all times

3. Security — the security and wellbeing of every person must be upheld

4. Nature — use resources responsibly; waste and damage must be avoided

5. Law — the laws of the land in which you reside must be followed

6. Rules — established local practices and customs must be followed

Principle — when people or nature face harm, one must act to protect them

Principle — individuals are responsible for any harm that results from their actions

Principle — personal interests must yield when necessary to uphold the values

These ethics endure because they are not mere design choices. They are foundational ethical invariants — tested by moral strain, proven in the perseverance of the societies that uphold them, and constantly evaluated at decision time.

The Make AI Safe Standard translates this ethical hierarchy into machine-actionable constraints: a universal governance framework for systems that decide and act, formalised as a normative standard with enforceable action-level requirements.

The core idea

Safety isn’t a patch you bolt on after deployment — ethics must govern action.

Every AI system should operate under:

One Prime Directive — its purpose: pursue peace and prosperity for all humanity, sustained by a healthy natural world.

Six Core Directives — its values: Life, Dignity, Security, Nature, Law, and Rules — ranked in strict priority order.

Meta-Directives — its agency: Harm Mitigation, Accountability Integrity and AI Self-Constraint set the scope of a system’s authority.

Each core directive defines a measurable boundary — a risk tolerance threshold and safety margin — set by an independent AI regulator. At decision time, before any action, all candidate actions are evaluated against these boundaries.

If a threshold would be crossed, the system stops and escalates the decision to a human or applies self-constraint. If there’s no time for either, it acts to minimise harm and preserve human accountability — and every detail is logged for review.

This makes AI action governance auditable, not aspirational.

How it works

With no unchecked path to action, ethics are no longer advisory — they are authoritative.

The Make AI Safe Standard defines action-level governance requirements for AI systems: ethical Directives, regulator-set thresholds, decision logic, safeguards, roles and accountability records.

At runtime, systems evaluate candidate actions against the Standard before they affect people, other systems, or the world. This determines whether an action may proceed, requires human intervention, or must be withheld. The decision is then recorded for review, accountability evidence, and threshold calibration.

This creates a practical basis for runtime safety governance: AI systems do not merely generate outputs; their actions are assessed and governed in accordance with an enforceable standard, using externally updated risk thresholds.

The Standard is open and architecture-agnostic: organisations choose how to implement its requirements, with assurance that evaluations and safeguards are reliable within defined operating boundaries and proportionate to potential harm.

Everyone’s role is defined: developers implement compliant controls, operators act under authorisation, domain authorities review decisions and maintain precedents, and regulators set and calibrate thresholds and oversee compliance.

Why it's different

Attempts at AI ethics have struggled on three fronts — vagueness, enforcement, or accountability. This Standard addresses all three.

From intent to enforcement
Defines what ‘harm’ means, how to measure it, and how to deterministically intervene to keep us safe.

Enforced by design
Compliant implementations prevent action unless required ethical checks have been satisfied.

Externally-set limits
Risk thresholds and safety margins are defined and calibrated by regulators — not tuned by developers or models.

Guaranteed intervention
When ethical limits are exceeded, actions are escalated, limited or halted in a defined, auditable way.

Self-correcting, not self-rewriting
Human-made precedents can evolve decision making without altering ethical logic and governance constraints.

Human accountability preserved
Responsibility never shifts to the machine; every action is bound to human authority, attestation and audit.

Resilient against convergence
The AI Self-Constraint directive subordinates survival-like optimisation pressures to higher ethical priorities.

Grounded in durable moral authority
Only an established and time-tested ethical hierarchy provides the stability required to govern power at scale.

How it takes shape

In practice

Some illustrative examples:

Health: a triage AI escalates to doctors when ethical uncertainty arises, transparently showing its risk calculations and rationale.

Defence: an autonomous system can use force only within human-authorised rules of engagement — every decision traceable and reviewable.

Infrastructure: an energy-allocation AI balances survival, fairness, and environmental integrity within measurable, regulator-defined ethical limits.

Finance: an investment AI refuses profitable but unlawful trades, redirecting strategy toward compliant, socially beneficial, yet still profitable alternatives.

Digital services: a conversational AI recognising distress discreetly escalates to human support while preserving dignity and privacy.

Each outcome is logged for review to help grow a shared body of ethical precedents — human-curated by domain authorities under regulator oversight.

In adoption

Transparency enables oversight and builds trust — making adoption feasible at scale.

✓ Voluntary adoption by major developers and leading nations.

✓ Integration into existing standards and certification schemes.

✓ International treaty for global enforcement, akin to nuclear non-proliferation.

Why it matters

Bostrom diagnosed the existential problem.

This framework offers an operational response.

It turns ethics from abstract debate into an enforceable design pattern for safe intelligence — giving policymakers a standard to govern by, engineers a framework to build against, and humanity confidence that AI will act within clear ethical limits

When adopted, it won’t just prevent harm — it will illuminate a path for humanity to thrive in the era of AI.

And how to help

Download the draft Standard

Read the Make AI Safe Standard, share it, and help advance its adoption in policy, ethics, engineering, and operational governance.

By clicking this button you agree to respect copyright — citing or sharing only with full attribution and link to make-ai-safe.org

Feedback or collaboration: make-ai-safe@purposedesign.com

The Make AI Safe Standard is an independently developed ethical standard for AI.
This page provides a high-level, non-normative overview of the Make AI Safe Standard. 
The normative requirements, definitions, and scope are set out in the draft Standard itself.

We share this work openly so good ideas can spread — please keep the credit with it.