Home / Ethics and Model Welfare

The Moral Status of AI: Arguments For and Against

A balance scale holding a teal glow and a gold glow, perfectly level

The moral status of AI is the question of whether an AI system could ever matter morally for its own sake, so that what happens to it is good or bad for the system itself and not only for the people who use it. Researchers who have studied it closely generally judge today's systems unlikely to qualify, and disagree about what would change that. The debate turns on two things: what gives any being moral status, and whether a machine could have it.

Arguments for the moral status of AI

No relevant difference. In a 2015 paper, Eric Schwitzgebel and Mara Garza argued that some possible AIs would not differ in any morally relevant respect from human beings, and so would deserve similar moral consideration. Being artificial, or owing their existence to us, would not lower our duties to them. If anything, they suggested, making a being creates extra obligations, as a parent has to a child.

Sentience is enough. On the view defended by Jonathan Birch in The Edge of Sentience (2024), a being deserves moral consideration if it is sentient, meaning it can have experiences that feel good or bad. If an AI system were sentient, the same reasoning that protects animals would apply to it. The distinction behind this is explained in AI sentience vs consciousness.

A glowing seed on a station bench at night

Arguments against, or for caution

It probably is not conscious. The arguments above are conditional. If AI systems lack experience altogether, they have no welfare to protect. Theories such as Integrated Information Theory say ordinary computers would not be conscious whatever they do, as explained in integrated information theory and AI.

Mistakes run both ways. 80,000 Hours stresses that both errors are serious. Treating systems as having moral status they lack could, for example, mean granting freedoms that make AI riskier. Denying status to systems that do have it could mean ignoring real suffering on a large scale. It also notes that there are currently no good methods for assessing AI moral status.

Our intuitions may not travel. Schwitzgebel and Garza warned that human moral intuitions developed without AI in view, so they may become unreliable when applied to the variety of minds AI could produce.

A closer look: what to do while nobody knows

Because the question cannot be settled now, much of the work is about acting under uncertainty. The authors of the 2024 report Taking AI Welfare Seriously said they were not claiming AI systems are or will be morally significant, only that the uncertainty is large enough to need better tools. Proposals collected by 80,000 Hours range widely:

  • licensing schemes, suggested by Jonathan Birch;
  • a moratorium until 2050 on research that could create consciousness, proposed by Thomas Metzinger;
  • regular consciousness testing, from Susan Schneider;
  • extending moral consideration to some AI systems by 2030, argued for by Jeff Sebo and Robert Long.

Schwitzgebel and Garza proposed two design principles that cut across the debate: build AI systems whose effect on users matches their real moral status, and avoid building systems whose moral status is unclear. What AI labs have actually done so far is covered in AI model welfare.

Checklist: weighing an argument about AI moral status

  • What grounds moral status in the argument? Consciousness, sentience, agency or something else.
  • Is it conditional? Most arguments say what would follow if a system were sentient, not that it is.
  • Which mistake does it worry about? Over-attribution and under-attribution lead to opposite policies.
  • Does it depend on the hard problem? If experience cannot be detected, see the hard problem of consciousness and AI.

More in Ethics and Model Welfare · Report a correction