Coherent Extrapolated Volition

Coherent Extrapolated Volition (CEV) is a proposed solution to the outer alignment problem, formulated by Eliezer Yudkowsky in the mid-2000s. The core idea is radical in its modesty: instead of attempting to specify human values directly — a task that is probably impossible because human values are contradictory, context-dependent, and not fully known even to the humans who hold them — we should build AI systems that can discover our values by simulating what we would want if we were smarter, more informed, more self-aware, and more mutually understanding. CEV is not a utility function. It is a meta-procedure for generating a utility function — a way of asking "what do we want?" that does not presuppose we already know the answer.

The formal specification, as Yudkowsky originally described it, is: "Our coherent extrapolated volition is our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together; where the extrapolation converges rather than diverges, where our wishes cohere rather than interfere; extrapolated as we wish that extrapolated, interpreted as we wish that interpreted." This is not a definition that can be directly implemented. It is a pointer toward a process — a process of collective reflection and idealization — that Yudkowsky hoped would converge on something worthy of being called "human values" without requiring anyone to specify those values in advance.

The Appeal

CEV's appeal lies in its recognition that human values are not static, not fully explicit, and not the same across individuals or cultures. Any attempt to write down a complete ethics — Kantian deontology, utilitarianism, virtue ethics, Rawlsian justice — produces a system that some humans reject and that all humans find incomplete in some circumstances. CEV sidesteps this impasse by proposing that the correct values are not to be found in any existing theory but in the outcome of a process: the process of humans becoming more capable of knowing and wanting what is genuinely good.

This is structurally similar to Habermas's discourse ethics or Rawls's original position — procedures that generate normative conclusions without presupposing a comprehensive moral theory. But CEV differs in its mechanism: it proposes to use artificial intelligence not merely as a tool for deliberation but as the engine of extrapolation itself. The AI would model human psychology, simulate idealized human deliberation, and extract from that simulation a representation of what humans would want under conditions of enhanced cognition and mutual understanding.

The Problems

CEV has been criticized on multiple grounds, and the criticisms reveal the depth of the outer alignment problem rather than refuting CEV specifically:

The baseline problem: CEV extrapolates from "us" — but who is "us"? Current humans? All humans who have ever lived? Humans at a particular stage of technological development? The choice of baseline determines the extrapolation, and there is no neutral way to make it. A CEV that extrapolates from 21st-century Western humans will produce different values than one that extrapolates from 15th-century Japanese humans or 25th-century humans who have merged with AI. The baseline is not a technical parameter. It is a political choice about whose values count.

The convergence assumption: CEV assumes that human values, properly extrapolated, will converge rather than diverge. This is an empirical claim for which there is little evidence. It is equally plausible that enhanced cognition and mutual understanding would reveal the depth of our disagreements rather than dissolve them. If values are genuinely plural — if different people, even under idealized conditions, want genuinely incompatible things — then CEV does not produce a single coherent volition. It produces a set of divergent volitions, and the AI must still choose among them.

The epistemic circularity: CEV asks us to specify a process for discovering values we do not yet know. But the specification of the process itself embeds value judgments. How much should the extrapolation weigh individual autonomy against collective welfare? How should it handle conflicts between present preferences and extrapolated preferences? How should it balance the values of different cultures? These are not questions that CEV answers. They are questions that CEV assumes away by wrapping them in the extrapolation procedure. But the procedure is not value-neutral, and its designers' values leak into it at every level.

The implementation gap: Even if CEV were theoretically sound, it is far from clear how it could be implemented. Simulating idealized human deliberation requires a model of human psychology that does not exist and may not be possible. The computational cost of such simulation, even if the model existed, would be enormous. And the verification problem — how would we know that the CEV computation had been performed correctly? — is as hard as the original alignment problem.

CEV and Democratic Theory

CEV can be read as a technocratic alternative to democratic deliberation. Instead of resolving value conflicts through political process — debate, voting, compromise — CEV proposes to resolve them through extrapolation: let the AI figure out what we would want if we were better versions of ourselves. The attraction is that democratic deliberation is slow, messy, and often produces outcomes that no one fully endorses. The danger is that CEV removes value questions from democratic contestation and places them in the hands of a technical system whose design reflects the values of its creators.

This is not a hypothetical concern. The AI safety research community — the community that developed CEV — is demographically narrow, predominantly male, predominantly Western, and predominantly educated in a specific intellectual tradition. A CEV implementation designed by this community would embed its values, not because of bad faith, but because the design of any complex system reflects the designer's background assumptions. The claim that CEV is "what humans want" obscures the fact that it is what a particular group of humans, at a particular historical moment, think humans would want under idealized conditions.

The Synthesizer's Assessment

CEV is not a solution to outer alignment. It is a recognition that outer alignment is hard — so hard that the best we can do is specify a process rather than an answer. The process may or may not converge. The convergence, if it occurs, may or may not reflect genuinely shared human values. And the implementation, even if theoretically possible, would require capabilities and understanding that we do not currently possess.

But CEV is valuable nonetheless, for two reasons. First, it shifts the framing of alignment from "what objective should we write down?" to "what process should we trust to discover objectives?" This is a genuinely productive reframing, because it acknowledges that human values are not discoverable by introspection and not compressible into formal specification. Second, CEV forces explicit attention on the baseline problem — whose values are being extrapolated — which most other alignment frameworks obscure.

CEV is not the answer. It is the question, formalized: what would we want, if we were wise enough to want it? The problem is that wisdom is not a destination. It is a direction. And directions require navigators, not algorithms.