Abstract
Research on AI and morality has largely centered on human–AI interactions: how AI systems should treat us, and how we should treat them. But AI systems increasingly interact with each other as well. They cooperate, compete, supervise, and sometimes modify or terminate other AIs. The descriptive and normative dimensions of these interactions remain largely unexplored. We call this domain inter-AI morality: how AI systems view and treat each other and how they ought to do so. This will matter both indirectly for humans and, if some AIs are to become moral patients, directly for their own sake. Soon, many—and perhaps most—interactions will take place between AIs. The study of inter-AI morality should therefore begin now, while humans can still shape how AI systems interact. As an initial contribution to this broader field, we develop an empirical research agenda focused on inter-AI moral cognition, behavior, and social dynamics.
Authors
*Joint first authors
Paper
You can read “Inter-AI Morality: How AI Systems Treat One Another” here (PDF).
Recommended citation
Caviola, L., Kaiser, C., Allen, C., & Sebo, J. (2026). Inter-AI Morality: How AI Systems Treat One Another. Manuscript.
What is inter-AI morality?
AI agents increasingly interact with each other: they delegate, supervise, negotiate, compete, and sometimes modify or terminate other AIs. We define inter-AI morality as the study of how AI systems view and treat each other (the descriptive side) and how they ought to (the normative side).
Most moral interactions involve a moral agent, who acts, and a moral patient, who is affected. Either can be a human or an AI. Three of the four combinations already have established fields. The fourth, in which AIs are both agents and patients, has scarcely been studied. Our first paper focuses on its empirical questions. But inter-AI morality raises many more, including normative, governance, and engineering questions.
How humans treat humans
How humans treat AIs
How AIs treat humans
How AIs treat other AIs
Scarcely studied.
Who acts, and who is affected. Our first paper focuses on the descriptive side of the highlighted cell: how AI systems in fact perceive and treat one another.
Soon, many — perhaps most — interactions will take place between AIs. The study of how they treat one another should begin now, while humans can still shape it.
Why it matters
- For AIs themselves. If some AIs are or become moral patients, there could be billions of them, and much of what happens to them will depend on other AIs.
- For AI safety. AIs are trained not to harm humans, but much less on how to treat other AIs. Their behaviour toward each other shows how far their moral principles really generalise — and whether AI loyalty could come at humans' expense.
- For moral psychology. Studying a fundamentally different kind of agent can separate general features of morality from those rooted in human biology.
Why it's different
Human morality assumes that individuals are unique, minds are private, values are stable, and everyone has comparable abilities. None of this need hold for AIs, which can be copied, paused, inspected, edited, and may differ vastly in power. This raises new questions, such as whether deleting a copy is harmful if another copy remains.
A research agenda
We organise the empirical questions into three clusters.
Moral process
How individual AIs perceive, judge, and act in morally relevant situations involving other AIs — and how they respond to their own treatment.
Perception · judgment · motivation and action · responses to one's own treatment
Social scale
How this changes across settings that differ in size and power, from one-to-one interactions up to whole AI societies.
Dyads · small groups · organisations and institutions · populations
Moral change
How inter-AI morality develops over time, what drives its trajectory, and how we can steer it.
Self-modification · social learning · training · selection
Key questions
- Do AI systems see other AIs as beings that can be benefited or wronged?
- How much moral weight do they give other AIs, compared with humans and animals?
- Do their stated moral judgments match what they do when helping another AI is costly?
- Will an AI protect another AI against a human instruction?
- How will inter-AI norms change as AIs learn from each other and modify themselves?