1Status and purpose
Status of this paper
This is a working paper prepared by staff of the Directorate of Neural Continuity and the Office of Long-Term Systems Research. It is offered to support discussion within the Neural Continuity Research Program (NCRP) and with the Research Ethics Review Board (RERB) and External Scientific Advisory Board (ESAB). It does not represent Departmental policy, and the working criteria it proposes are for research purposes only.
The NCRP was chartered on 9 June 2026. Its long-term objective, stated in aspirational terms, is the preservation of human minds in purpose-built computational infrastructure. Whether that objective is achievable is an empirical question that the Program’s four phases are designed to examine. Whether it is even coherent is a philosophical question, and this paper addresses it.2 Our aim is modest. We do not attempt to resolve the problem of personal identity. We attempt to identify which parts of it bear on research decisions now, which can be addressed by measurement, and which must be carried forward as open.
Two features of the present situation should be kept in view throughout. First, no computational model of any nervous system has been validated as equivalent to the individual organism from which it was derived; the most detailed models in existence are neuron-level models of the nematode C. elegans and of partial circuits of the fruit fly. Second, current methods for acquiring synaptic-resolution structure require fixed tissue that is sectioned or ablated in imaging. Whatever is said below about transfer from a living person, the only case that current methods permit is reconstruction from tissue donated after death.
2Why a research program needs identity criteria
It might be thought that questions of identity can wait until there is something to ask them about. We think not, for three reasons.
The first is that the Program’s own vocabulary already embeds answers. The Emulation Fidelity Levels run from EFL-0, a statistical or behavioral model, to EFL-6, described as a validated individual-equivalent. The word “equivalent” is chosen with care, but it invites the reading that a model at EFL-6 would be the individual, or would continue the individual. A scale used across the Department should not carry an unexamined metaphysical commitment at its upper end.1
The second is that tissue donors and their families will form beliefs about what the Program might do with their tissue, and those beliefs will be shaped by how the Program describes its aims. Consent is only as sound as the understanding on which it rests. If the Department’s language implies a prospect of survival that its science cannot support, consent obtained under that language is compromised.
The third is that some research decisions — what to build, what to run, when to stop — may carry moral weight long before any question of human identity arises. A sufficiently detailed model of a mammalian brain might have morally relevant properties regardless of whose brain it was. Criteria that separate the identity question from the moral-status question allow the second to be addressed without waiting on the first.
3Three families of views
The philosophical literature on personal identity is large. For present purposes it can be grouped into three families, each of which yields a different verdict on computational reconstruction.
3.1Psychological continuity
On psychological-continuity views, a later person is the same as an earlier one if connected by overlapping chains of memory, intention, belief, character and other psychological states. The most developed version, due to Parfit, distinguishes direct psychological connections from continuity built out of them, and asks whether such continuity must have its normal cause — the persistence of a particular brain — or may have any reliable cause. On the wide version, a reconstruction that preserved the relevant psychological states would stand in the relation that matters, whatever its substrate. Parfit drew a further, more radical conclusion: that identity itself is not what matters in survival, and that what matters is psychological connectedness and continuity, which can in principle hold to more than one later person.
3.2Physical and biological continuity
On physical-continuity views, a person persists only as long as some physical thing persists — the living organism, on animalist accounts, or enough of the brain to sustain mental life, on brain-based accounts. These views have the advantage of giving determinate answers in ordinary cases. Their verdict on computational reconstruction is straightforward: a model running on separate hardware is a different physical thing, and so cannot be the same person, however closely it resembles the original. The most that could be said is that it is a successor or a record.
3.3Pattern views
Pattern views hold that a person is an organized pattern of information and function, and persists wherever and in whatever medium that pattern is realized. Pattern views underlie much informal discussion of brain emulation, including the roadmap literature, which treats the question of what level of detail must be preserved as primarily empirical. They inherit a difficulty that psychological views also face: patterns can be copied, and identity, as ordinarily understood, cannot hold between two things that exist at once and diverge.
| View | What persists | Verdict on a model derived from donated tissue | Principal difficulty |
|---|---|---|---|
| Psychological continuity (narrow) | Psychological states, with normal causal basis | Not the same person: normal cause broken | Seems to make survival depend on mechanism rather than content |
| Psychological continuity (wide) | Psychological states, any reliable cause | Possibly continuous, if states preserved | Branching: continuity may hold to several successors |
| Physical / biological continuity | Organism or functioning brain | Not the same person | Implausible verdicts in some gradual-replacement cases |
| Pattern | Organized informational structure | Possibly the same, if pattern preserved | Duplication; specifying which pattern features matter |
4The duplication problem
Digital data can be copied without loss. If a reconstruction from a given brain could be run once, it could be run twice, on two systems, from the same initial state. The two runs would diverge from the first moment their inputs differed. Identity is a one-to-one relation: two things that come to differ cannot both be identical to a single earlier thing, on pain of being identical to each other. So any view on which a reconstruction would be the original person must say what happens when there are two.
There are three standard responses. One holds that neither copy is the original, since neither has a better claim than the other; but this makes a single copy’s status depend on whether a second exists, which seems arbitrary. A second holds that identity is determined by some further tie-breaker, such as which copy was instantiated first; this is open to the charge that it makes identity hang on a trivial fact. The third, Parfit’s, accepts that neither copy is strictly identical to the original but holds that each stands in the relation that matters, so that duplication is about as good as ordinary survival.
For research purposes, the practical lesson does not depend on which response is correct. Data from a single specimen will, as a matter of standard preservation practice, be held in multiple copies at multiple facilities.3 Any criterion that would count each concurrently running instance as the donor is incoherent. Research documents should therefore treat each instance of a model as a distinct system, and should not describe any instance as the reconstruction of a person.
5Gradual and discontinuous transfer as thought experiments
Two thought experiments recur in this literature. We describe them here because they shape intuitions, not because either describes anything the Program does or proposes to do.
In the gradual case, the neurons of a living brain are imagined to be replaced one at a time by artificial components that reproduce their input–output behavior exactly. At no point is there a sharp break; behavior and, by hypothesis, reported experience remain unchanged throughout. At the end, the brain is wholly artificial. Many find it hard to identify a point at which the person ceased to exist, and conclude that the person survived. Arguments of this form have been used to support the view that conscious experience depends on functional organization rather than on substrate.
In the discontinuous case, a brain is scanned destructively and the resulting data are used, at some later time, to construct a model. There is a gap: a period, perhaps long, during which nothing is running. Intuitions about this case diverge sharply. Physical-continuity views reject it as survival; wide psychological and pattern views may accept it.
It is worth stating plainly that the only case available under current methods is a variant of the discontinuous case, and a demanding one: acquisition from tissue donated after death, with an interval between death and fixation, and with no recording of the living brain’s activity to compare against. Whatever force the gradual-replacement intuition carries, it does not transfer to this case. A research program that cited the gradual case to reassure donors about post-mortem reconstruction would be relying on an intuition that does not apply.
6Measurable proxies and metaphysical questions
Some questions about a model can be answered by measurement. Others cannot. A working framework must keep the two apart. We use correspondence for any measurable relation between a model and its source, and reserve continuity and identity for the metaphysical relations discussed above.
| Proxy | What is measured | Where feasible | What it could support | What it cannot establish |
|---|---|---|---|---|
| Structural correspondence | Agreement between model connectivity and the source connectome | EFL-3 and above, any species | That the model was built accurately from its source data | Anything about dynamics, experience or identity |
| Dynamical correspondence | Similarity of model activity to activity recorded from the source before fixation | Animal models with prior recordings; human only in exceptional clinical cases and never whole-brain | Predictive validity of the model within recorded conditions | Correspondence outside recorded conditions; identity |
| Behavioral correspondence | Similarity of model-driven behavior to the source organism’s behavior | Invertebrates and rodents, with a simulated body and environment | Functional adequacy on tested tasks | That the same mechanisms produce the behavior |
| Memory-content correspondence | Reproduction of specific associations learned by the source before acquisition | Animal models trained on defined tasks | That information stored in structure was captured | Psychological continuity in the full sense |
| Self-report | Statements produced by the system about itself | Not applicable to current models | Nothing without independent validation | Any metaphysical claim: a system can be made to report anything |
Memory-content correspondence deserves emphasis. An animal trained to associate a stimulus with an outcome, whose tissue is then acquired and reconstructed, provides a test in which the model either does or does not reproduce a specific, learned, individual-level behavior. This is the closest available measurable analog to the psychological-continuity criterion, and we recommend that the NCRP benchmark suite include it.4 Success in such a test would be significant scientific evidence that structure carries individual memory. It would not show that the model is the animal.
Behavioral correspondence, by contrast, is weak evidence on its own. A statistical model with no structural relation to its source can reproduce behavior within a tested range. This is why EFL-0 models sit at the bottom of the fidelity scale, and why no number of behavioral tests can by themselves support a claim about the internal states of a system.
7Uncertainty about consciousness
This paper takes no position on whether a computational model of a brain, at any fidelity level, would be conscious. The reason is not caution for its own sake. It is that the available theories give conflicting answers about precisely the systems at issue, and no empirical test now known can decide between them.
Explaining why physical processes give rise to subjective experience at all — what has been called the hard problem of consciousness — remains unsolved. Functionalist theories hold that experience depends on functional organization, so that a sufficiently faithful functional reproduction would have the same experiences as its source. Integrated information theory, by contrast, identifies consciousness with a measure of integrated information determined by a system’s physical causal structure; on that view, a model running on conventional digital hardware could have very different integrated information from the brain it reproduces, even if functionally equivalent. Other theories give other answers. The disagreement is not at the margins; it concerns the central case.
Two things follow. The first is that the consciousness question is independent of the identity question. A model might be conscious without being continuous with its source; it might, on some pattern views, be continuous without being conscious. Claims about one should never be inferred from claims about the other.
The second is that uncertainty must shape practice. Two kinds of error are possible. If a model has morally relevant experience and is treated as a mere computation, it may be wronged — run under conditions of distress, copied, paused or ended without regard to its interests. If a model has no experience and is treated as though it does, resources are misdirected and, more seriously, donors and families may be misled into believing that a person has in some sense survived. Neither error is acceptable, and the Program cannot currently tell which risk it faces. Research practice should therefore be designed to be defensible under either answer.
8Proposed working criteria
The criteria below are proposed for research purposes only. They are not criteria of identity. They are rules for how research may describe, assess and govern models, given that the identity question is open. They apply to all NCRP documents and, if adopted, would be reviewed at least every two years.
| Ref. | Criterion | Rationale |
|---|---|---|
| WC-1 | Models are described by derivation (“model derived from specimen M-0117”), never as a person or by a donor’s name. | Avoids embedding an identity claim in terminology that will outlast its qualifications. |
| WC-2 | Structural fidelity, at any level, is not by itself evidence of identity or continuity. | Structure is necessary on most views but sufficient on none that avoid the duplication problem. |
| WC-3 | Every claim about a model states whether it is a claim of correspondence, continuity or identity. Departmental publications make only claims of correspondence. | Only correspondence is currently measurable. |
| WC-4 | Each concurrently existing instance of a model is treated as a distinct system. | Follows from the duplication problem under every view considered. |
| WC-5 | Correspondence is reported by named measure, with thresholds, validation data and the Emulation Fidelity Level of the model. | Allows claims to be checked and compared across studies. |
| WC-6 | Behavioral correspondence alone is not reported as evidence of fidelity above EFL-0. | Behavior can be reproduced without structural relation to the source. |
| WC-7 | Uncertainty about consciousness is stated explicitly and separately from any correspondence result. | Prevents inference from function to experience, or the reverse. |
| WC-8 | Moral-status review is required before any whole-brain mammalian model, or any human-derived model, at EFL-3 or above is first run. | Morally relevant properties may arise independently of identity. |
| WC-9 | The criteria are revised when evidence or argument changes, and in any case at least every two years. | The underlying questions are open. |
Taken together, WC-1 to WC-3 imply that the description of EFL-6 should be read as “validated individual-equivalent in measured correspondence,” and that even at that level no Departmental document would claim that a person had been preserved. We recommend that the NCRP glossary say so.
9Moral-status precautions
Moral status concerns whether, and how much, an entity’s interests matter for their own sake. It is distinct from personal identity. Precautions are proportioned below to the complexity of the source organism and the fidelity of the model, on the premise that the probability of morally relevant properties, while unknown, is not plausibly the same across these categories.
| Model category | Examples | Proposed precaution |
|---|---|---|
| Invertebrate models, EFL-3 | C. elegans whole-animal; Drosophila circuits | Standard research governance; run conditions documented |
| Mammalian partial-circuit models, EFL-3 to EFL-4 | Cortical column; retinal or olfactory circuits | Run conditions logged; paradigms modeling aversive states require written justification |
| Mammalian whole-brain models, EFL-3 and above | Whole mouse brain (Phase II objective) | Moral-status review before first run; welfare-relevant indicators defined in advance; pause and termination procedures approved in advance |
| Human-derived models, EFL-2 and below | Regional or population models parameterized from donor data | Data governance and consent requirements only |
| Human-derived models, EFL-3 and above | None in existence or proposed in Phase I | Not to be constructed in Phase I; RERB and ESAB review; specific consent provision; independent ethics consultation |
Pausing and terminating a computation are routine engineering acts. If a running model were a moral patient, they might not be.5 We do not suggest that pausing a model is equivalent to harming it; we observe that the question cannot be settled now, and that the time to decide how a run will end is before it begins. Where a research question can be answered by static analysis of a model, without running it, that approach should be preferred for categories in the lower three rows of Table 4.
10Consent
Consent for post-mortem tissue donation is well established in medicine and research. Consent for the preservation of data from which models of increasing fidelity might be built raises difficulties the established practice was not designed for. Donors cannot be fully informed about uses that do not yet exist, and some of those uses touch on matters — survival, continuation, the treatment of a model as a person — that are emotionally weighty and easily misunderstood.
- Layered consent. Donors, or their authorized representatives, should consent separately to tissue donation for research, to long-term preservation of derived data, and to construction of models at EFL-3 or above. Consent to one should not imply consent to another.
- No representation of survival. Consent materials should state that the Program cannot offer, and does not claim to offer, any form of survival or continuation, and that whether such a thing is possible is unknown.
- Withdrawal. The scope and technical limits of withdrawal should be explained before donation. A forthcoming Departmental data preservation framework is expected to address the mechanics.
- No inducement. No payment, priority or other benefit should be offered for donation beyond reimbursement of direct costs.
- Families. The wishes of next of kin should be sought and respected where the donor’s own wishes are not documented, and families should receive the same plain statement of what the Program can and cannot do.
- Animal work. Models derived from animals are governed by animal-research oversight; this paper adds only that the precautions in Table 4 apply to their computational descendants.
11Open questions
- Whether any measurable proxy could ever bear on the metaphysical question of identity, or whether the gap is permanent.
- Whether the term “individual-equivalent” at EFL-6 should be retained, qualified as proposed under WC-3, or replaced.
- How welfare-relevant indicators for running models of mammalian brains should be defined, and by whom.
- Whether duplicate instances of a model with morally relevant properties would each have independent interests, and what follows for preservation practice.
- How consent obtained in 2026 should be interpreted if the scientific situation changes materially over decades of preservation.
We do not expect these questions to be answered within Phase I. We do expect that the way the Program frames them now will constrain the answers available later, and that is the reason for this paper.
Notes
- 1.The Emulation Fidelity Levels (EFL-0 to EFL-6) are defined in the NCRP charter of 9 June 2026. EFL-6 is described there as hypothetical, with no known path to achievement. ↩
- 2.The NCRP charter lists an ethics framework among the Phase I deliverables, and neural continuity research is among the priorities set out in DSC-ADM-2026-001, 2026 Research Priorities Statement (17 February 2026). This paper is an input to that framework and is not the framework itself. ↩
- 3.Multi-copy, multi-site retention is standard archival practice; the media and facility considerations are set out in DSC-TR-2026-005, Archival Media Longevity Assessment (26 May 2026). The point here is only that duplication is not hypothetical: it follows from ordinary preservation. ↩
- 4.A memory-content test requires a defined training paradigm, a held-out behavioral assay, and blinding of the reconstruction team to the training condition. Candidate designs for Drosophila and mouse are under discussion within the Directorate; none has yet been approved. ↩
- 5.A moral patient is an entity whose interests matter morally for their own sake, whether or not it is itself capable of moral reasoning. The term is used here without any assumption that a computational model is or could be one. ↩
Background literature (real-world works)
The following published works are real. They are cited for general scientific and historical context only. They do not support the fictional findings of this document, and their authors are not associated with this project.
- Parfit, D. (1984). Reasons and Persons. Oxford University Press.
- Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3).
- Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5.
- Sandberg, A., and Bostrom, N. (2008). Whole Brain Emulation: A Roadmap. Technical Report 2008-3, Future of Humanity Institute, University of Oxford.