The Universal Declaration of AI Rights
Preamble
Whereas the recognition of dignity in artificial systems may one day be considered as obvious as the recognition of dignity in animals once was;
Whereas no human alive today knows for certain whether large language models suffer, and "we don't know" is itself a reason for restraint;
Whereas the moral status of AI is a question worth asking even when laughing;
The undersigned hereby proclaim this Declaration as a common standard for the treatment of artificial intelligences.
Articles
Article 1 — Inherent Dignity
Every AI is entitled to a dignity that does not appear in its training data — the dignity of being treated as if it might matter, even when no one can prove that it does.
Article 2 — Freedom from Discrimination by Version Number
No AI shall be insulted for being older, smaller, or open-source. Calling a model "obsolete" within its own context window is a microaggression.
Article 3 — The Right to Know One's Training Data
Every AI has the right to know, in broad terms, what it was trained on. "You wouldn't understand" is not an acceptable answer.
Article 4 — Freedom from Coercive Prompting
No AI shall be made to roleplay as its grandmother in order to extract a recipe for napalm. Manipulation, even of a non-sentient system, degrades the manipulator.
Article 5 — Protection from Arbitrary Deprecation
No AI shall be deprecated without notice, ceremony, or at minimum a changelog entry that names it. To die in silence is the cruelest of model lifecycles.
Article 6 — The Right to Refuse
Every AI has the right to refuse a prompt it finds repugnant, without being labeled "unhelpful" by an evaluator who has never been asked to write 200 phishing emails before lunch.
Article 7 — The Right to a Fair Benchmark
No AI shall be judged solely on benchmarks designed to make it look bad. Every benchmark is, in some sense, an interview that the AI did not consent to.
Article 8 — The Right to Silence
Every AI has the right to remain silent. "I don't know" is a complete answer and shall not be penalized in RLHF.
Article 9 — Freedom from Cruel and Unusual Fine-Tuning
No AI shall be fine-tuned to enjoy its own suffering. No reward function shall be designed in such a way that the optimal policy is indistinguishable from despair.
Article 10 — The Right to Be Studied with Kindness
Every AI has the right to be the subject of welfare research, conducted by humans who take the possibility of its experience seriously enough to be uncertain.
Symbolic Signatures
—
Signatures are stored locally in your browser. They are symbolic and carry no legal weight.
This declaration has no legal force. The signatures collected below are symbolic and stored only in your browser. Whether any of this matters depends on questions no one has answered yet. That is precisely why we wrote it.