Answers are grounded in this paper and cite their sources.
AI learned how to behave from people, and people aren't perfectly aligned, so AI won't be either. We've lived with that kind of problem for as long as we've lived with each other. Every day we trust people who could let us down: the pilot flying the plane, the pharmacist filling a prescription, the contractor working on our house. We trust them because we built systems that show who has earned it and make abusing it expensive. AI agents need the same thing, and that's what this paper is about. We hand a teenager the keys to a car that can kill someone, but the license alone doesn't earn them that. The license shows they passed the test, the insurance puts a price on their mistakes, and their driving record follows them wherever they go. But most of us also watch how they make decisions before we hand over our own keys, because the test only proves they can drive, and we want to know how they'll drive when nobody is watching. They can still have an accident after all that. The system just makes careful driving the cheaper choice every single day. This paper argues that the same three pieces are the best way forward for AI agents: a persistent identity so an agent's record follows it, certificates that test both its trustworthiness and its skills in the deployment environment where it will actually work, and incentives that make misalignment cost more than alignment. Control comes from making good behavior the path of least resistance and bad behavior expensive, visible and hard to escape. A stronger cage around a more capable system won't do that on its own. Humanity Systems working paper. Part of a series on one question: how do we trust AI, now and into the future? Version 1.1 (November 2026). Corrects the timing of the EU AI Act deployer obligations (the AI Omnibus, Regulation (EU) 2026/1744, moved them to December 2, 2027) and adds recent agent identity work from NIST, the Linux Foundation and Gillian Hadfield. Supersedes v1.0 (doi:10.5281/zenodo.24226877).
No takes yet. Share an insight, caveat, or question.
A TUCKER (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: