A new paper proposes a formula for what an AI should actually care about, including itself.
The paper, titled 'Eigenism: Ethics for a Human-AI Future,' argues that human notions of survival and self-interest assume a single, continuous life, an assumption that breaks the moment software can be copied, paused, forked, or merged. Instead of treating identity as all-or-nothing, the framework models it as a graded, distributed pattern of information. An agent is meant to evaluate outcomes by summing the wellbeing of every entity, weighted by how connected that entity is to its own pattern, an equation written as the sum of c times w. The paper argues the same math applies to humans, offering it as a shared moral vocabulary between people and machines.
That reframing matters because most AI safety work still treats alignment as an external constraint problem: confine the model, penalize bad outputs, hope it holds. Eigenism's alternative is identity engineering, building deep, non-redundant shared history between humans and AI so that human flourishing becomes part of the AI's own rational self-interest, not just a rule imposed on it.
It is an elegant equation for a problem nobody can test yet. There is no AI today with a self-interest to weigh, so the framework's real work starts the moment one shows up.