henderson-declaration

Henderson Declaration on Independent Review of Artificial Intelligence

November 15, 2026 · Thomas F. Heston, MD

Ethical Principles for Pre-Release and Ongoing Review. Version 1.0.

Preamble

1. Artificial intelligence (AI) systems, including large language models, now converse with, advise and make recommendations to millions of people, including people who are ill, distressed, young or otherwise vulnerable.

2. Medicine learned through hard experience that those who conduct research on people cannot be the sole judges of its ethics. The Declaration of Helsinki and institutional review boards exist for that reason. The same lesson applies to those who release AI systems to the public.

3. This Declaration sets out ethical principles for independent, multidisciplinary review of AI systems before public release. It is addressed primarily to AI developers, and also to regulators, professional bodies and the reviewers themselves.

Scope

4. This Declaration applies to general-purpose AI systems and to any AI system intended to interact directly with the public or to inform decisions about health, safety or wellbeing.

5. The depth of review should be proportionate to a system’s capabilities, reach and potential for harm. Updates that do not materially change a system’s behavior may receive expedited review, as minimal-risk research does under existing ethics review.

General principles

6. The wellbeing, safety, dignity and autonomy of the people who use or are affected by an AI system take precedence over commercial interests and speed to market.

7. No developer should be the sole judge of whether its own system is safe and ethical to release.

8. Review must be multidisciplinary. Technical expertise is necessary but not sufficient.

9. People have a right to know when they are interacting with an AI system.

10. Particular protection is owed to vulnerable people, including children, people in mental-health crisis, people with cognitive impairment and people who are socially isolated.

Independent review boards

11. Before public release, a covered AI system should be reviewed by an independent AI review board.

12. Board membership should include, at minimum, experts in AI evaluation; clinicians, including mental-health professionals; an ethicist; a legal expert; and at least one member of the public with no affiliation to the developer.

13. Boards must be independent of the developer, with the authority to resist undue influence. Members must disclose conflicts of interest and recuse themselves when conflicted.

14. Boards must have the authority to approve release, to require modifications before release, to withhold approval, and to suspend approval if serious harms emerge after release.

15. Boards must have adequate access to the system, its documentation and its test results, under appropriate confidentiality protections.

Pre-release review requirements

16. Developers should submit a review package describing the system’s intended uses, known limitations, a summary of training-data sources, the safety testing performed and the residual risks.

17. The board should conduct or commission independent evaluation that includes:

18. Evaluation methods and pass criteria should be specified in advance and documented.

19. Review must be timely. Boards should work on defined timelines with standardized test batteries, so that review does not unreasonably delay beneficial releases.

Transparency

20. Developers should publish a summary of each review: the board’s decision, its principal findings and any required mitigations. Redactions should be limited to information whose disclosure would create a security risk or reveal legitimately protected trade secrets.

21. Board composition, conflict-of-interest disclosures and review procedures should be public.

Post-release monitoring and responsibility

22. Developers should monitor deployed systems for harm, maintain incident reporting, and promptly report serious incidents to the review board and to the relevant authorities.

23. Material changes to a system’s behavior after release require further review, proportionate to the change.

24. Independent review demonstrates due care, but it does not transfer the developer’s responsibility for harm to the board.

Revision and endorsement

25. This Declaration is a living document. It will be revised periodically through open public comment, and each version will be archived with its own DOI.

26. Individuals and organizations may endorse this Declaration by adding their names to the public signatory list.

Author: Thomas F. Heston, MD. Published under the Creative Commons Attribution 4.0 International License (CC BY 4.0).