AI Safety, Alignment, and Ethics (AI SAE)

Ethics should be built into AI systems from the ground up, not added as an afterthought. The paper argues that current approaches treat ethics as an external constraint, when instead it should be embedded as a core design principle—similar to how cooperation evolved as a fundamental survival strategy in nature.

Moral values can be identified and manipulated within AI neural networks. Using techniques like sparse autoencoders and activation steering, researchers may be able to locate where moral reasoning happens in AI systems and directly influence how these systems make ethical decisions, moving alignment from theory to engineering practice.

Alignment requires coordination across three levels: how AI systems think internally, how they're optimized, and how they're governed externally. The paper proposes treating alignment as a multi-layered problem rather than a single solution, requiring changes to AI architecture, training methods, and institutional oversight simultaneously.

Different moral theories can be tested empirically rather than debated philosophically. By representing competing ethical frameworks as different geometric patterns in AI neural networks, researchers can test which moral representations actually work best for sustaining human-AI cooperation as AI systems become more powerful.

The goal is sustainable human-AI cooperation ("symbiosis") as AI capabilities increase. Rather than controlling AI through restrictions, the paper suggests building systems where ethical cooperation becomes the natural, embedded strategy—making alignment self-reinforcing rather than externally enforced.