The Non-Identity Problem

In the [YOU] on AI Field Guide

The cycle built around [YOU] on AI asks what we owe to the world that AI is making. The non-identity problem delivers the precise answer to why that question is harder than it looks: the AI development choices that will most profoundly shape the future—the pace of deployment, the safeguards adopted or skipped, the concentrations of power permitted, the global order established—will reach into the conditions under which billions are conceived and born, and will thereby determine which billions those are. A reckless trajectory that produces a damaged but survivable future will not have harmed any specific future person in the ordinary sense, because those specific people owe their existence to the recklessness. The non-identity problem tells us, with unwelcome rigor, that we cannot ground our obligations regarding AI in the most natural concept available to us.

This is why the thinkers in the cycle who care most about long-run AI risks—the longtermists, many of whom regard Parfit as a founding figure—are working in territory he mapped as dangerous. His problem licenses the conclusion that something can be deeply wrong about a future we choose even when no particular future person is wronged. It does not license the confidence with which that conclusion is sometimes deployed. Parfit himself reached no tidy answers and warned that our best impersonal principles collapse when followed to their logical ends.

Origin

Parfit introduced the problem in Reasons and Persons (1984) through a pair of cases. The first is a fourteen-year-old girl who chooses to have a child now: because she is so young, the child has a worse start in life than the child she would have had if she had waited a few years. Our intuition is firm—she did something wrong. But Parfit shows that if she had waited, this child would never have existed at all; the later child would have been a different person. So her actual child cannot complain of being harmed, because the only alternative for this child was non-existence. As long as the child’s life is worth living, no particular person was made worse off.

Parfit then scales the case to the Depletion-Conservation choice, where different policies produce not just different outcomes but entirely different future populations. The point generalizes: any choice that significantly alters the conditions of society over enough time produces different people, not different versions of the same people. This means the ordinary “person-affecting” principle—an act is wrong only if it is worse for some person—cannot ground our most important obligations to the future. The conclusion is clearly wrong—it would permit strip-mining the future—but the argument is valid. The framework must be wrong.

Key Ideas

The person-affecting constraint. The standard principle of welfare ethics holds that an outcome can only be worse if it is worse for some person. This is intuitive, humane, and anchored in respect for actual individuals rather than abstractions. It also fails completely when applied to choices that alter which individuals will exist. Parfit’s contribution was to show that the failure is not a technical defect to be patched but a structural feature of person-affecting reasoning that no patch can repair.

The AI extension. AI development is the most consequential non-identity choice of the current era. The pace of deployment, the concentration of access, the governance structures established now will alter the conditions of billions of conceptions across decades and centuries. The people who will live in the world those choices create are not worse-off versions of people who would otherwise have existed; they are different people whose existence is contingent on the choices. The non-identity problem tells us we cannot ground our obligations in harm to them—and simultaneously that this is not permission to be reckless, because something can be deeply wrong without a specific harmed individual.

The impersonal turn. The only way past the non-identity problem is to adopt a principle that judges futures by the quality of the lives they contain rather than by harms to nameable individuals. This is the impersonal ethics Parfit was attempting to build when he encountered the Repugnant Conclusion. The impersonal turn is necessary. It is also treacherous: every impersonal principle that avoids the Repugnant Conclusion generates its own unacceptable result. Parfit spent decades searching for Theory X and did not find it.