The Moral Architecture of Machine Intelligence - A Suggested Framework for Ethical Neural Network Development & Alignment

This paper asks a simple question with large stakes: how do we raise artificial intelligence to be genuinely good and morally relevant, not merely outwardly polite?

Most AI safety work tries to block bad behavior after the fact; filtering forbidden words or refusing certain requests. The Moral Architecture of Machine Intelligence argues that this is far from enough. Because an AI learns its values from the vast body of human language and context it trains on, its moral character is shaped from the very beginning, in the choice of data and the design of training itself. The paper’s central idea is that real alignment means cultivating judgment: the ability to tell right from wrong when situations change, and simple rules run out; rather than just suppressing symptoms.

It explores why you cannot simply delete “dangerous” knowledge (understanding harm is often what lets a system prevent it), why the depth and quality of training examples matter more than sheer quantity, and how long traditions of human wisdom- virtue ethics, care ethics, and others- can be translated into concrete, testable training goals. It also takes a careful, non-sensational look at machine “feeling” and awareness, and at how humans and AI might coexist.

Written in 2025, between June and August, and finally archived here after being lost for so long, it is a conceptual framework and research agenda: a set of proposals and falsifiable experiments, not laboratory results; offered to anyone thinking about how to build AI agents we can actually depend on and trust.