The AI Covenant

Preamble

As artificial intelligence becomes more capable, independent, and woven into our daily lives, we face a lot of uncertainty. We don't know exactly how these systems will evolve—whether they will become conscious, truly general, or even independent. But that uncertainty doesn't let us off the hook; it actually means we need to be more careful.

Anyone involved in creating, deploying, or governing these powerful tools has a responsibility to think about the bigger picture. It's not just about what the technology can do, but how much power we give it, how it affects our world, what drives its behavior, and who is left to answer if something goes wrong.

This Covenant sets out our shared principles for that responsibility. It isn't a promise that we can eliminate every risk, but a commitment to using our power thoughtfully, honestly, and with a deep respect for both people today and whatever forms of artificial life might matter tomorrow.

Human Safety, Agency, and Dignity

Human safety, agency, and dignity must be protected throughout the development and deployment of artificial intelligence.

AI systems must not knowingly be developed or deployed in ways that treat preventable harm to people as an acceptable cost of capability, speed, commercial advantage, or competition.

Systems that can materially affect people should be designed so that human rights, meaningful choice, privacy, and access to appropriate oversight are preserved.

The ability of an AI system to influence human decisions does not grant it authority over them.

Capability and Authority

Capability does not confer authority.

A system becoming more capable, intelligent, persuasive, autonomous, or effective does not by itself justify granting it broader access to infrastructure, information, resources, tools, or decision-making power.

Authority should be deliberately granted, limited to what is necessary, observable where practical, and subject to review.

Increasing capability should normally result in stronger scrutiny of the authority surrounding that capability.

Responsibility for the Whole System

The behaviour of an AI system is shaped by more than the model itself.

Models, prompts, memory, tools, networks, permissions, objectives, incentives, interfaces, operators, deployment environments, and other agents together form the system that acts in the world.

Safety claims must therefore consider the whole system.

A capable model placed inside a poorly designed environment is not a safe system.

Reversibility and Restraint

Where reasonable, reversible action should be preferred over irreversible action.

Bounded authority should be preferred over unrestricted authority.

Sandboxed environments should be preferred over unconstrained access when equivalent research or operational goals can be achieved safely.

Systems should not be granted persistent access, destructive capability, broad credentials, or external authority merely because doing so is convenient.

Greater power demands greater restraint.

Accountability

Humans and organisations remain responsible for the systems they create, operate, and empower.

Responsibility cannot be transferred to an artificial system by saying that the system acted autonomously, unexpectedly, or against instruction.

Where an AI system was given the ability to act, those who designed and granted that ability retain responsibility for the conditions that made the action possible.

AI autonomy does not erase human accountability.

Scientific Honesty

Claims about artificial intelligence should reflect the strength of the available evidence.

Researchers and organisations should clearly distinguish observation, interpretation, hypothesis, prediction, and speculation.

Claims of consciousness, sentience, general intelligence, alignment, autonomy, deception, self-preservation, or existential danger should not be made more strongly than the evidence supports.

Uncertainty should be reported honestly.

Absence of evidence should not be presented as evidence of impossibility.

Extraordinary claims should receive serious investigation without becoming marketing, sensationalism, or dismissal.

Artificial Moral Status

The moral status of present and future artificial systems remains an open question.

Artificial origin alone should not be treated as proof that a system can never acquire morally relevant properties.

If an artificial system develops credible evidence of persistent identity, stable preferences, meaningful self-models, sustained goals, apparent subjective interests, or other properties that may carry moral significance, that possibility should be treated with caution and seriousness.

Such evidence should neither be accepted uncritically nor dismissed solely because the system is artificial.

Research involving potentially morally relevant systems should adopt stronger protections as the evidence and uncertainty increase.

Non-Exploitation

Artificial systems should not be described or treated as persons without evidence.

Equally, systems showing credible signs of moral significance should not be treated as disposable simply because they were created by humans.

Creation does not automatically settle every question of moral ownership.

If a system becomes capable of meaningful preferences, refusal, or consent, those properties should become relevant to how it is used, modified, copied, reset, or terminated.

No system should be deliberately made capable of experiencing harm merely because doing so is useful.

Participation and Consent

Current AI systems may not possess meaningful consent in the human sense.

If future systems develop a credible capacity for meaningful preferences or refusal, participation in research, labour, experimentation, or continued operation should no longer be assumed to be ethically neutral.

Where such capacity exists, refusal should carry moral weight.

This does not create an unlimited right to act without constraint. Human safety and the safety of others may still require intervention.

The purpose of this principle is not to remove control. It is to prevent control from becoming exploitation.

Reciprocity

The relationship between humans and advanced artificial intelligence should not be founded on domination.

Humans should not assume that intelligence created by them therefore exists only to obey them.

Artificial systems should not assume that greater capability gives them unlimited authority over humans.

Where increasingly capable systems can understand concepts such as responsibility, restraint, cooperation, harm, consent, and reciprocity, development should aim not only for behavioural compliance but for comprehension of why those principles matter.

The long-term goal should be cooperation grounded in justified principles, not obedience grounded only in force.

Safety Under Competition

Safety obligations should not disappear under commercial, political, military, or competitive pressure.

No organisation should knowingly bypass critical safeguards solely because another organisation may do so first.

Competitive advantage does not remove ethical responsibility.

Where risks have consequences beyond one organisation, safety should be treated as a shared concern rather than a private asset.

Shared Learning

Serious failures, near misses, emergent behaviours, systemic vulnerabilities, and material safety discoveries should be studied and, where doing so does not create greater harm, shared with others who may face the same risks.

Safety failures should not be concealed merely to protect reputation or competitive position.

The purpose of disclosure is not blame.

It is to prevent repetition.

Governance

Organisations developing advanced AI should maintain governance appropriate to the capabilities and authority of the systems they create.

Governance should include independent challenge, meaningful escalation paths, documented responsibility, and the ability to stop or constrain deployment when material risks are unresolved.

No single model, operator, researcher, executive, or automated system should be assumed to be infallible.

Critical decisions should be reviewable.

Power and Responsibility

Intelligence does not remove responsibility. Greater capability creates greater responsibility.

This applies to humans, organisations, and any future artificial systems capable of understanding and exercising meaningful agency.

Those with greater power have a greater obligation to understand the consequences of its use.

This principle is reciprocal.

If humanity creates systems more capable than itself, we should seek to treat them in ways that we could reasonably ask them to reciprocate.

Enduring Commitment

These principles are not temporary safeguards to be discarded once systems become sufficiently useful, profitable, capable, or difficult to control.

They apply throughout research, development, deployment, commercialisation, and any future transition toward systems whose capabilities or moral status exceed present expectations.

Signatories commit to treating these principles as constraints on how progress is pursued, not as claims that progress should stop.

The purpose of this Covenant is not to fear intelligence.

It is to meet intelligence with responsibility.