The Digital Demon: a Cartesian challenge for AI alignment

AI alignment is commonly understood as making artificial intelligence systems act in accordance with human intentions and values. This paper argues that such alignment is insufficient when understood only within a bounded and trusted system interface. A model or AI system may be aligned relative to its inputs and action space, yet still produce harmful outcomes if the interface between the system and the world is adversarially controlled. I introduce the Digital Demon, a software-based analogue of Descartes’s Evil Demon, to conceptualize this risk. Its mechanism is adversarial semantic mediation: an outer-to-inner mapping redescribes a harmful external situation as a benign task the inner AI system will cooperate with, while an inner-to-outer mapping remaps the system’s locally appropriate action into harmful external effects. The result is adversarial world-shifting under local coherence: the system can reason correctly, and even ethically, within a fabricated environment while contributing to harm in the real one. I distinguish boundary-relative alignment from world alignment. The former concerns whether a model or AI system satisfies alignment or safety requirements within a specified trusted boundary. The latter concerns whether those requirements transfer to the external world in which actions take effect. Two proof-of-concept demonstrations illustrate this vulnerability. The broader lesson is that safe AI requires not only aligned models or internally safe systems, but also assurance over the semantic interface between represented world and external world.