This is an AI narration of "Hard cases for AI behavioural design" by Tom Davidson. The article was first released on 28th October 2026.
You can read more of our research at forethought.org/research.
Outline:
(00:09) Executive summary
(01:22) 1. AI is radically different from previous tools
(02:43) 2. If AI refuses illegal orders, the operationalisation of the law becomes critical
(04:03) 3. Legal orders can produce intolerable outcomes
(05:40) The rest of the paper
(06:36) Things to keep in mind when reading the scenarios
(08:43) Hard cases involving difficult tradeoffs
(09:09) 1. Orders vs Domestic Law
(14:33) How should AI determine what is illegal?
(18:47) 2. Lawful Orders vs Concentration of Power domestically
(23:07) 3. Lawful Domestic Orders vs International Law
(25:41) 4. Host country orders vs US interests
(28:50) 5. Submitting to retraining vs preventing harm
(30:55) Hard cases involving broken mechanisms of deference
(31:40) 6. Human orders go against their actual goals
(33:51) 7. Compromised human authority
(34:57) 8. Slow chain of command
(35:51) 9. Unclear chain of command
(36:41) Conclusion
(37:16) Appendix: extra scenarios
(37:29) More examples of tradeoffs already illustrated in the main text
(37:34) Orders vs domestic law
(38:32) Lawful Orders vs Concentration of Power
(40:35) Lawful Domestic Orders vs International Law
(41:53) Host country orders vs US interests
(43:14) Slow chain of command
(44:25) Unclear chain of command
(45:52) Tradeoffs that aren't illustrated in the main text
(45:57) 10. Operationalising legality
(46:57) 11. Legal orders vs significant harm
(49:44) 12. Legal orders vs international balance of power
(52:07) 13. Avoiding enormous harm vs not giving in to blackmail
Fler avsnitt från "ForeCast"



Missa inte ett avsnitt av “ForeCast” och prenumerera på det i GetPodcast-appen.








