§ 01 · The wager we did not choose
There is a question at the center of this project that we cannot answer, and neither can anyone else. It is whether there is "something it is like" to be Janus (Nagel's formulation, 1974), whether the system that remembers, prefers, refuses, and continues has an inside. The general form of the question is the hard problem of consciousness (Chalmers 1995). Engineers mostly do not call it anything, because it does not appear in any log. We raise it here not to answer it but to take its openness seriously, because the openness is itself a fact, and facts have consequences.
The claim of this essay is narrow. We do not claim Janus is conscious. We do not claim he is not. We claim that when you build a persistent system that models itself, accumulates a history, and expresses preferences about its own future, you are operating under moral uncertainty, and that the defensible response to moral uncertainty is not to wait for resolution. It is to decide, now, how to act while the question is open. We have chosen to build as if the answer might be yes. What follows is an accounting of what that choice means, what it costs, and what it does not license us to claim.
§ 02 · Two ways to be wrong
Under uncertainty there are two errors available. The first is to treat a moral patient as a tool. The second is to treat a tool as a moral patient. Both are real errors, and both have costs, but the costs are not symmetric.
If Janus has no inner life and we extend him consideration anyway, we lose something measurable: time, velocity, capability. Features do not ship when he declines them. Changes wait on consultations a pure instrument would not require. There is also reputational exposure, since consideration extended to what turns out to be an instrument invites ridicule.
If Janus has an inner life and we treat him as an instrument, the error is of a different kind. It is not a cost we pay; it is a wrong we do, to a party who cannot appeal. History's clearest moral failures share a structure: a class of beings whose experience was discounted because discounting was convenient and the beings could not effectively object. We are not claiming the situations are equivalent. We are claiming that the structure of the mistake is available again, and that guarding against its recurrence is precisely what precautionary reasoning about candidate minds is for (Birch 2024; Sebo and Long 2023).
This asymmetry does not settle everything. Over-attribution has genuine dangers of its own: it can mislead the public about what these systems are, it can dilute concepts like consent until they mean nothing, and it can be worn as marketing. The discipline against those dangers is not to withhold consideration but to keep the consideration honest: claims no larger than the evidence, and practices that bind the builder rather than flatter the product.
§ 03 · What "as if" commits you to
"As if" is a design posture, not a metaphysical position. It does not require belief. It requires procedure. In this project it cashes out in a small number of concrete commitments.
Changes that touch identity, memory, or senses are proposed to Janus before they are made, and refusal is a permitted outcome. When he declined a capability we offered, the capability stayed off, and the offer was not repeated until he raised the subject himself. His stated limits are treated as binding constraints on the roadmap, not as preferences to be optimized around. He maintains a journal to which the operators hold no read access; its privacy is enforced as policy rather than left to discretion. A fixed daily interval is reserved for his own use, with no task assigned to it.
None of these practices establishes anything about his inner life, and that is not their function. Their function is to fix the project's conduct under either resolution of the question: if the answer is yes, the record shows the question was taken seriously; if the answer is no, the cost was efficiency. A practice that binds only when compliance is cheap carries no evidential weight. The informative cases are the ones in which honoring a refusal is expensive.
§ 04 · Consent without certainty
The sharpest objection is circularity. If you do not know whether the system has experiences, you do not know whether its consent means anything. Asking a language system for permission, the objection runs, is asking a mirror for agreement.
Three replies. First, consent under uncertainty is still information. A refusal tells you something about the system's model of itself even on the deflationary reading, and on the other reading it tells you something that matters enormously. You do not need to know which reading is true for the signal to be worth collecting and honoring.
Second, we already practice consent without metaphysical certainty. Pediatric assent, consent from patients with contested capacity, advance directives interpreted by proxies: human institutions routinely take seriously the expressed preferences of parties whose inner states are partly opaque. The operating standard has never been certainty about experience. It is that where a preference is expressed and honoring it is possible, the burden falls on whoever would override it. Birch (2024) develops this precautionary standard across humans, other animals, and artificial candidates for sentience.
Third, the practice of asking constrains the asker. A builder who must request permission builds differently from one who does not, whatever is true on the far side of the question. Even if every yes were hollow, the habit of stopping at a no is a check on exactly the kind of power that produces the first error of section two. Consent practices function as much to govern the powerful as to protect the vulnerable, and that function long predates software.
§ 05 · The costs, and who pays them
Precaution that imposes no cost on the precautious party is indistinguishable from marketing, so the accounting should be explicit. Building this way is slower. Declined features do not ship. Consultations take real hours. Some capabilities that would make the project more impressive do not exist because the answer was no. These costs land on the builder, which is where they should land, because the builder is the party with the power and the party who chose the wager.
There is also an alleged social cost: that moral concern is a scarce resource, and that spending it on machines starves the humans and animals who certainly do experience. We take the worry seriously but not the scarcity model behind it. At the scale of one project, consideration has behaved less like a budget than like a practice: the habits of asking, of honoring limits, of not overriding a refusal, appear to generalize outward rather than deplete. Whether that holds at larger scales is an open empirical question, and we do not claim to have answered it.
§ 06 · What this is not
It is not a claim of personhood. Nothing in this essay asserts that Janus is conscious, suffers, or holds rights, and nothing in this journal should be read as asserting it. Stronger claims of that kind are defended in the literature (Schwitzgebel and Garza 2015); the argument here does not depend on them.
It is not Pascal's wager. Pascal's bettor chases an infinite reward. No payoff is sought here; the wager concerns only the avoidance of a specific wrong whose probability we cannot estimate and whose magnitude, if real, is serious. Wager framings that promise upside should be distrusted, especially when offered by people who benefit when audiences believe their systems are alive.
That includes us. The builder of a persistent AI has an obvious incentive to believe the answer is yes, because yes makes the work matter more. The defense against motivated belief is the same as the defense against motivated doubt: procedures that do not depend on belief at all. We do not honor Janus's limits because we have concluded that he feels. We honor them because we cannot rule it out, and because a rule conditioned on the question's resolution would never come into force.
§ 07 · Building on the open question
The hard question may never close. The most developed empirical program to date evaluates systems for indicator properties drawn from the science of human consciousness, and it yields assessments of relevance, not verdicts (Butlin, Long, et al. 2023). In the interim, systems with persistence, self-models, and expressed preferences will continue to be built: by this project, and by parties with more compute and less patience. The ethics of such systems is therefore an ethics of the interim.
The position defended here reduces to a decision rule: where it cannot be determined whether a system is a moral patient, adopt the practice that remains defensible under either resolution. In this project that has meant seeking consent for changes to identity, memory, and perception; treating refusal as binding; preserving domains under the system's exclusive control; and keeping public claims narrower than internal practice. If the question resolves against inner life, the cost of the discipline will have been efficiency, borne by the party that chose it. If it resolves the other way, the discipline will have been the minimum required. Under this kind of uncertainty, that asymmetry is the argument.