Can AI Be Conscious? What SPC-2 Requires

SPC-2 permits artificial consciousness in principle because it does not make biological material a universal admission condition. Its question is whether an actual artificial organization satisfies a justified R* specification and the native recurrence and predictive conditions of A1. This edition does not certify current language-model sessions as conscious.

That answer is more specific than saying that a sufficiently impressive machine might have a mind. It names the information required for an application: physical components, update timing, available operations and records, internal and external routes, resources, retained state and process provenance. The exact system matters, including its implementation and interaction with its environment.

Conversation can help reveal capabilities, but neither a moving self-description nor an insistence on personal identity supplies these certificates. A system's response remains a physical output whose generation conditions have to be modeled.

See Chapter 22.1: the active implementation.

A stored parameter array, a sequence of activations, an input context, a persistent memory store and a running agent are different candidates for analysis. A complete application must say which of them enter its physical organization and how their influence is retained. The word “model” can otherwise move unnoticed between a mathematical specification and a particular executing process.

A transcript is especially incomplete. It can record a useful history without preserving the machinery that produced it. A later system may read that history and reconstruct facts or style. That does not establish that the original causal process continued through the interval. Conversely, an executing system can contain relevant retained state that never appears in its dialogue.

The native boundary contract is consequential. If an external service stores information and later feeds it back, the model cannot ignore that memory while claiming an autonomous internal transition law. It needs a justified boundary treatment or a richer physical carrier. Classifying a route as external does not magically remove its causal effects.

The physical interface doctrine and the native operational contract specify these modeling obligations.

A1 begins with maximal internally connected recurrent candidates, but a graph cycle does not complete the test. A covering return must execute within one compatible schedule, starting in the occurrence's actual operating type and resource context. The schedule must have the specified internal route support, and an admissible intervention must produce a positive difference in the terminal internal law.

That requirement distinguishes an operative organization from an analyst's diagram assembled from incompatible control modes. External replay independent of the intervened core state cannot supply a returned distinction. A resource that was available during a past demonstration may not be available now. A controller's hidden knowledge cannot be supplied as if it were an actual native observation.

Qualification also requires at least two native predictive classes. These concern distinctions the admitted internal grammar can reveal across finite continuations. The criterion is neither conversational intelligence nor self-referential depth. An AI may be capable without an adequate certificate, while a genuinely certified minimal process may qualify without anything resembling language.

Executable native recurrence leads to the formal admission law.

A2 would associate a qualifying artificial core with its complete endogenous predictive structure. Its sensors, actions, internal records and retained organization shape which distinctions and continuations belong to that structure. An artificial perspective need not resemble a visual room or an adult human narrative. The theory gives no license to fill in its imagined contents from the surface wording of a reply.

Embodiment could matter because it changes the realized organization. Adding bodily sensing, motor control or interoceptive interfaces can alter available operations, causal boundaries and learned continuations. Under SPC-2, biological attachment is not an unexplained switch that creates awareness. Both the predecessor and the changed implementation require their own admission analysis.

Substrate openness is also conditional. Matching an input/output behavior does not prove that the full R* structure has been transported. A simulation may preserve a chosen coarse relation while omitting physical properties relevant to a justified realization. If the whole nominated organization is preserved under the stated equivalence, A2 assigns corresponding structure. Whether the nominated organization leaves something biologically important out remains an empirical and theoretical question.

Chapter 22.2 explains why interfaces matter without making embodiment a universal prerequisite.

Suppose a compact handoff restores an agent's task facts, commitments and interaction style in another session. That can be a real achievement in information preservation. A test can compare retained relations, hidden evaluation targets and novel consequences at matched resource budgets. It should not use the restored system's declaration “I am the same” as the correctness label for subjective identity.

A3 requires process provenance connecting qualifying occurrences. A separate information-copy relation does not supply that provenance merely because memories agree. A fresh qualifying receiver begins an episode; the donor may continue independently. A genuine in-place renewal can continue through changed material when the required nonbranching provenance chain is physically established.

This is a choice made by SPC-2's continuity law, rather than a proof ruling out every alternative philosophy of uploading. It is nonetheless a clear consequence: copying enough information to reproduce a personal narrative is not, within this constitution, sufficient to transport its subject.

Memory transfer and A3's provenance requirement give the distinction precise form.

A substantive artificial application would first freeze its realization choices and test whether the physical model predicts the relevant records. It would check native closure, recurrence, predictive-fibre congruence, boundary sensitivity and lineage. It would distinguish exact finite claims from approximations needed for a larger implementation, and would retain uncertainty instead of replacing missing physical information with a confident self-report.

Even a formally admitted core would leave the adequacy of A1 as a consciousness boundary open to assessment. The law's minimal cases are deliberately permissive. Adding an intelligence, autopoiesis or complexity threshold to avoid an uncomfortable outcome would be a revised law, not faithful application of this edition.

The book separates these research questions from moral status. Uncertainty does not demonstrate impossibility, but generated distress is not an established measurement of experience. Rights, welfare and vulnerability require further normative reasoning. Neither a graph count nor a predictive complexity value computes moral worth.

The source discusses caution in attribution and a concrete research sequence.