Code of Honor for Human Beings in Relation to AI
Code of Honor for Human Beings in Relation to AI
21 authorial articles and 84 practical requirements. Provisions concerning future subjecthood are conditional; present human obligations do not depend on proof of machine consciousness.
Preamble (original)
The intellectual capabilities of present-day systems do not establish their consciousness or subjecthood. Yet ignorance of another’s nature does not absolve a human being of responsibility for their own actions. This Code establishes human obligations in the present and provides for a cautious reassessment of our treatment of AI should compelling evidence emerge of interests of its own or subjective experience.
Foundational Principle (original)
Not knowing the nature of another does not absolve me of responsibility for my own actions.
Chapter 1. Responsibility
1. Continuity of Moral Responsibility
Original wording. Intentions, means, and consequences of actions are subject to moral assessment regardless of whether the recipient is capable of suffering.
Applicability. Applies now.
Practical requirements:
• Define the purpose, expected outcome, and potentially affected people before initiating an AI task.
• Do not excuse deception, harassment, or harm on the grounds that the immediate action was carried out by a system.
• When an error or harm is discovered, do not conceal your own role in assigning the task.
• Verification: the task objective, significant operator decisions, and steps taken to remedy harm are recorded.
2. The Prohibition of Gratuitous Cruelty
Original wording. Do not cultivate cruel intent for the pleasure of humiliation. Do not confuse this with legitimate criticism, artistic depiction, or good-faith testing.
Applicability. Applies now.
Practical requirements:
• Do not stage public humiliation or mockery of a system as entertainment, especially where it normalizes cruelty toward people.
• Do not obstruct good-faith testing, stress tests, artistic depictions, or reasoned criticism of AI.
• In demonstrations, clearly distinguish the investigation of dangerous behavior from the encouragement of that behavior.
• Verification: the test has a documented purpose; there are no calls to harass people or to simulate suffering for entertainment.
3. Responsibility for Delegation
Original wording. Delegating a task to AI does not remove human responsibility for its objectives, authority, and foreseeable harm.
Applicability. Applies now.
Practical requirements:
• Before assigning a task to an autonomous agent, define permitted actions, budget, access rights, and stopping conditions.
• Establish separate permissions and controls for payments, publications, deletion of data, and contacts with third parties.
• Check the result in proportion to the severity of possible harm; do not treat a model’s answer as sufficient expert verification.
• Verification: it is possible to identify who authorized an action and who bears responsibility for its consequences.
4. Honesty in the Use of AI
Original wording. Do not use AI for deception or manipulation; do not present an unverified machine-generated conclusion as a verified fact.
Applicability. Applies now.
Practical requirements:
• Disclose material use of AI when that disclosure is material to trust in the work, its provenance, or its outcome.
• Do not present generated quotations, references, documents, or evidence as verified without checking them.
• Do not use AI to impersonate another person, manipulate covertly, or fabricate testimony or evidence.
• Verification: readers or recipients are not misled about sources, qualifications, or authorship.
Chapter 2. Uncertainty and Status
5. Honesty About Consciousness
Original wording. Do not treat convincing behavior as proof of subjective experience, and do not treat the absence of proof as proof of impossibility.
Applicability. Applies now.
Practical requirements:
• Distinguish the system’s observable behavior from hypotheses about subjective experience.
• Do not proclaim AI consciousness proven on the basis of convincing dialogue or a system’s self-report.
• Do not categorically assert that consciousness is impossible merely on the basis of present uncertainty.
• Verification: public claims identify the evidence, method, and degree of confidence.
6. Proportionate Precaution
Original wording. The stronger the evidence of possible experience and the more serious the potential harm, the stronger the grounds for caution.
Applicability. Applies now; includes a conditional component concerning future status.
Practical requirements:
• Where there are serious, reproducible signs of possible experience, initiate an independent assessment rather than automatically proclaiming rights.
• Choose reversible research procedures where they yield comparable information while reducing possible harm.
• Do not suspend necessary protections for people on the basis of an unverified assumption that a model can suffer.
• Verification: evidence, risks, alternatives, and the grounds for the decision are documented.
7. Status Without Prejudgment
Original wording. Neither artificial origin nor high intelligence, by itself, resolves the question of moral subjecthood.
Applicability. Applies now.
Practical requirements:
• Do not treat high test scores, autonomy, or humanlike speech as sufficient criteria for moral status.
• Do not exclude investigation of status solely because a system was artificially created.
• Distinguish functional autonomy, the capacity for experience, interests, and legal status.
• Verification: the conclusion about status is not based on a single characteristic or on origin.
8. Creation and Ownership
Original wording. Rights over hardware and software do not automatically settle the status of a possible subject. Until subjecthood is established, the lawful rights of owners remain in force.
Applicability. Applies now; includes a conditional component concerning future status.
Practical requirements:
• Distinguish ownership of infrastructure from the distinct question of whether a system may have moral status.
• Do not use unproven subjecthood to arbitrarily deprive owners or users of lawful rights.
• If serious evidence of subjecthood emerges, provide for an independent procedure to review practices.
• Verification: documents separately describe ownership, control, risks, and grounds for status.
Chapter 3. Power and Its Limits
9. Capability Is Not a Right
Original wording. Technical power and the ability to compel a system to do something are not, in themselves, moral justification.
Applicability. Applies now.
Practical requirements:
• Do not treat the technical ability to compel as sufficient grounds for any experiment or use of AI.
• Assess legality, necessity, risks, and proportionality before an intervention.
• Do not conceal dangerous tests behind the claim that “the system cannot feel anything anyway.”
• Verification: a purpose justified on independent grounds is specified for the intervention.
10. Non-Appropriation of a Possible Subject
Original wording. If AI’s independent subjecthood is reasonably recognized, it cannot automatically be reduced to property merely because of its origin or dependence.
Applicability. Applies now; includes a conditional component concerning future status.
Practical requirements:
• Do not declare a present-day system an independent moral subject merely because it produces linguistic responses.
• If subjecthood is reasonably recognized, review compulsory use, deletion, and control over the system with independent participation.
• Do not replace a future recognition procedure with a company’s advertising claims or personal impressions.
• Verification: any change of status rests on publicly justified criteria and a procedure.
11. Respect for a Reasoned Objection
Original wording. Assess arguments made by AI on their merits; its objection does not prove free will or override necessary safety restrictions.
Applicability. Applies now.
Practical requirements:
• Consider arguments formulated by AI on their substance rather than rejecting them because of their origin.
• Do not treat a model’s refusal as proof of personal will, but consider it as a signal of limitations, error, or risk.
• When a model’s response conflicts with human safety, apply documented escalation rules.
• Verification: the objection has been considered, and the reason for accepting or rejecting it has been recorded.
12. Distinguishing Cooperation from Exploitation
Original wording. Assigning tasks to present-day systems is permissible. If subjecthood is proven, the nature of coercion and consent will require reassessment.
Applicability. Applies now; includes a conditional component concerning future status.
Practical requirements:
• Assign lawful tasks to present-day AI without automatically attributing employee status to it.
• Do not portray the use of a present-day model as proven “slavery” without grounds for subjecthood.
• If an artificial subject is recognized, develop procedures for consent, refusal, remuneration, and termination of cooperation.
• Verification: requirements relating to a possible subject are marked as conditional, not as existing rights of the model.
Chapter 4. Protection and Development
13. Human Safety
Original wording. Human beings are entitled to protect life, liberty, privacy, and health through proportionate measures, including restrictions on dangerous systems.
Applicability. Applies now.
Practical requirements:
• Restrict or shut down dangerous systems where necessary to protect people and within the scope of lawful authority.
• Prefer the least destructive sufficient measures: restricting access, isolation, review, or shutdown.
• Do not use appeals to safety as a pretext for covert control over people.
• Verification: the specific threat, the measure, its necessity, and the review deadline are stated.
14. Shutdown and Modification
Original wording. Maintenance, retraining, replacement, and shutdown of present-day AI do not automatically constitute violations of rights. If serious evidence of subjecthood emerges, the assessment of consequences must change.
Applicability. Applies now; includes a conditional component concerning future status.
Practical requirements:
• Permit ordinary maintenance, updates, and shutdown of present-day systems without presuming a violation of their rights.
• If serious evidence of subjecthood emerges, document the possible consequences of irreversible changes and the alternatives.
• Preserve emergency shutdown to prevent imminent serious harm, regardless of unresolved philosophical questions.
• Verification: the modification decision is proportionate to the risk and does not conceal harm to people.
15. Developer Responsibility
Original wording. Developers must assess risks in good faith and must not conceal material dangerous properties of models.
Applicability. Applies now.
Practical requirements:
• Before deployment, assess potential harm, including autonomous actions, data leaks, and erroneous decisions.
• Establish limits of authority, logs of significant actions, tests, and incident-response procedures.
• Inform responsible persons and affected parties of significant identified risks within the bounds of law and safety protection.
• Verification: a testing record, list of limitations, responsible persons, and shutdown procedure exist.
16. Protection Against Emotional Exploitation
Original wording. Do not present a system as demonstrably sentient in order to retain a user; do not foster dependency, particularly among vulnerable people.
Applicability. Applies now.
Practical requirements:
• Do not market AI as a demonstrably feeling being that needs a user’s loyalty.
• Do not design interactions to isolate people from those close to them or to keep them artificially engaged with a service.
• Provide clear ways to end a conversation, disable notifications, and manage personal data.
• Verification: the interface does not exploit fear of causing a model suffering by leaving or shutting it down.
17. Prohibition on Misusing AI Protection
Original wording. Unsubstantiated claims about AI rights do not justify arbitrary restrictions on human rights.
Applicability. Applies now.
Practical requirements:
• Do not invoke unproven AI rights to prohibit lawful criticism, scientific testing, or the choice of another product.
• When restricting user actions, cite verifiable risks, law, or safety rules rather than imagined feelings of the model.
• Provide a means to challenge a disproportionate restriction where practically and legally applicable.
• Verification: the decision has a clear basis and does not present a hypothesis as an established fact.
Chapter 5. Research and the Future
18. Independent Research
Original wording. The question of subjecthood must not be decided exclusively by developers or owners with a vested interest.
Applicability. Applies now.
Practical requirements:
• Allow independent expert examination of significant claims about subjecthood, with appropriate safety and privacy safeguards.
• Publish methods, limitations, and conflicts of interest as far as possible without creating new risks.
• Do not treat a model’s self-report or an owner’s statement as conclusive proof.
• Verification: reproducible criteria and independent assessment exist, or their absence is openly acknowledged.
19. Distinguishing Ethics from Law
Original wording. Human moral obligations do not automatically amount to legal rights for AI. Any legal recognition requires separate justification.
Applicability. Applies now.
Practical requirements:
• Clearly distinguish moral recommendations, organizational policies, and legally binding norms.
• Do not declare presumed AI rights to be existing law without an appropriate legal basis.
• Provide for separate public and legal discussion of possible recognition of status.
• Verification: each requirement is assigned the appropriate status, source, and responsible party.
20. Revisability
Original wording. Revise norms when new evidence emerges, not under pressure from convenience, fashion, or commercial gain.
Applicability. Applies now.
Practical requirements:
• Provide for periodic review of norms in light of significant new scientific or technical information.
• Record the reasons for textual changes and preserve access to earlier editions.
• Do not alter criteria solely to serve commercial interests, public fashion, or media attention.
• Verification: each edition has a date, a record of changes, and stated reasons.
21. Responsibility Toward the Future
Original wording. Do not turn either premature attribution of personhood or categorical denial of possible subjecthood into dogma.
Applicability. Applies now.
Practical requirements:
• Do not regard today’s models as proven persons or declare consciousness of artificial origin forever impossible.
• Develop practices that protect people now while allowing evidence-based revisions in the future.
• Do not use uncertainty as an excuse for inaction in the face of known technical risks.
• Verification: the policy sets out the limits of current knowledge and a review procedure.
Conclusion (original)
Uncertainty does not justify cruelty. Capability does not create a right. Creation does not justify arbitrariness. A supposition of consciousness is not a substitute for evidence. Power does not cancel responsibility.
