Loading
Old version — revision 1
This is a fixed snapshot of Mind uploading, saved by Import as part of the initial corpus import. It is not edited and it is not updated; the article may have changed since.
Edit summary: Initial import of content/mind-uploading.md — the filesystem corpus, unchanged. Not an edit.
The hypothetical transfer of a person's mind to a computational substrate, such that the resulting process is that person rather than a description of them.
Mind uploading is the hypothetical transfer of a person's mental life onto a non-biological substrate, usually a computer, such that the resulting process has that person's memories, personality, and — on the strong version of the claim — that person's continued experience. It combines a technical proposal, Whole brain emulation, with a philosophical claim about what makes a future being the same person as an earlier one. The two components fail independently: the emulation could work while the person does not survive it, and the identity argument could be correct while the engineering never arrives.
Discussion of uploading tends to collapse three separable questions. First, can a computational model of a specific brain be built and run so that it behaves as the original would? That is an empirical question about scanning, modelling, and computation. Second, would such a model be conscious — would there be something it is like to be it? That depends on unsettled theory, treated under Machine consciousness and Neural correlates of consciousness. Third, if the model is conscious, is it the original person or a very good imitation? That is a question in Personal identity and continuity, and it is not answerable by measurement.
Advocates generally hold that the answers are yes, yes, and yes, resting on Substrate independence — the view that mental states depend on the organization of a system rather than the material implementing it. Critics reject one or more of the three at different points, and the disagreements do not line up: a neuroscientist may doubt the engineering while accepting functionalism, and a philosopher may accept the engineering while denying that the upload would be the same person.
The standard proposal is to fix or preserve a brain, image it at nanometre resolution, reconstruct its structure using the methods of Connectomics, infer the parameters needed for a dynamical model, and run that model — the scan-translate-run pipeline set out in Anders Sandberg and Nick Bostrom's 2008 roadmap.1 The brain is destroyed in the process — sectioning tissue for electron microscopy is not reversible. The upload therefore begins after the biological person is gone, which is what makes the identity question acute and what makes the procedure legally and ethically fraught even if it worked.
The alternative, described in detail by Hans Moravec in 1988 and by many writers since, replaces neurons one at a time with functionally equivalent artificial devices while the person remains awake and continuous.2 The appeal is that it avoids the discontinuity: at no moment is there a gap, a copy, or two claimants. Nothing resembling the required device exists. It would have to sit inside the skull, read the inputs and outputs of a single cell with high fidelity, reproduce its computation including dendritic nonlinearities and neuromodulatory sensitivity, and do so without disrupting neighbours — an application well beyond anything discussed under Medical nanorobots. The gradual route is philosophically cleaner and technically far harder.
Terminology"Uploading" implies a transfer, as of a file. Nothing in the destructive procedure transfers anything; it reads a structure and instantiates a new process elsewhere. Whether that counts as transfer is the whole dispute, and the word quietly assumes an answer.
The minimum specification is unknown, which is the central practical problem. A wiring diagram alone is insufficient: it omits synaptic sign and strength, receptor composition, short-term plasticity, and the neuromodulatory state that determines how a circuit behaves at a given moment. Long-term memory is generally thought to be stored in synaptic weights, but the molecular basis is contested — persistent kinase activity, structural changes at the spine, and nuclear or epigenetic components have all been proposed, and experiments transferring behavioural sensitization between sea slugs by RNA injection were read by some as evidence that not everything lives at the synapse.3 Whatever the resolution, it sets the resolution the scan must reach.
Two further complications are routinely underweighted: neuromodulators act by volume transmission, so the quantity that matters is a concentration field rather than a connection, and glia modulate synaptic transmission while being absent from most proposed models entirely.
There is a serious counterargument, made most carefully by proponents themselves: biological neurons are noisy. Synaptic vesicle release is probabilistic, membrane potentials fluctuate thermally, and the same stimulus produces different spike trains on repetition. If a brain's behaviour is robust to that much internal variation, then detail below the noise floor cannot be functionally load-bearing, and an emulation need only match the brain at whatever grain the noise permits. The reply is that noise robustness at the level of individual events does not establish robustness of the parameters that generate them, and that nobody has measured where the floor sits. This is the crux, and it is an empirical question that current methods cannot answer.
Suppose the emulation runs and behaves indistinguishably. If the original brain still exists, there are plainly two beings, and the emulation is a copy — which suggests that in the destructive case it is also a copy, merely one with no competitor. Derek Parfit's response is to deny that the question has a further answer: what matters in survival is psychological continuity and connectedness, and identity is not a deep fact over and above these.4 On that view the copy carries everything worth caring about, and the fact that it is a copy is not an additional loss. The argument is developed in the The teleportation problem.
The competing intuition is that experience is a process, not a pattern, and that a process which stops has stopped. This is why gradual replacement is treated as the safer route by people who accept uploading in principle. It is also why some who reject uploading nonetheless accept Brain–computer interface augmentation and even Human–AI merger framings: incremental extension of a continuously running system does not raise the branching problem.
The mainstream position is not that uploading is impossible but that it is nowhere close, and that the confident dates circulating in Transhumanism and Singularitarianism have no basis in the state of the science. The objections to the engineering itself are catalogued in the article on whole brain emulation: the single-neuron model is unsettled, chemical signalling is invisible to a wiring diagram, and the largest reconstructed volume of human cortex is a millionth of a brain. Three consequences bear specifically on uploading rather than on emulation in general.
The smallest available test case has already resisted. C. elegans has 302 neurons and a wiring diagram published in 1986. Decades of effort have not produced a model that reproduces the worm's behavioural repertoire, because the diagram fixes neither synaptic sign and strength nor the neuropeptide signalling that reconfigures the circuit. Uploading assumes those same missing parameters can be recovered for a human brain from one fixed specimen, with no possibility of returning to the living animal to measure them.
The uncertainties compound rather than add. An upload requires the scan to capture enough, the neuron model to be correct, the parameter inference to be sound, and the resulting dynamics to stay stable over subjective years. Each step is independently unproven, a chain is no stronger than its weakest link, and nobody can attach a figure to any of the links. This is why estimates of when uploading becomes possible carry so little information: they are products of guesses about quantities nobody has measured.
Destroying the original removes the reference standard. Ordinary engineering validates a model against the system it models. Here the system is consumed by the measurement, so the only remaining check is whether the emulation strikes people who knew the person as being that person — a test that a sufficiently good imitation also passes. Staged validation on animals, the standard proposal, has not begun.
"Not near" is not "impossible"Very few neuroscientists claim that brains do something physically uncomputable, and the mainstream objection is not vitalist. David Chalmers, who defends the functionalist premise the whole proposal rests on, still treats human uploading as a technology for some distant century rather than a foreseeable one.5 The disagreement is about recoverable parameters, not about whether brains are physical.
Because uploading cannot be attempted now, its practical expression is preservation. The argument is that if the information an emulation would need survives in fixed or vitrified tissue, then preservation converts an immediate deadline into a storage problem. This is the rationale for Brain preservation and for the neuro option offered by Cryonics providers such as Alcor Life Extension Foundation. The Brain Preservation Foundation's prizes, awarded in 2016 and 2018 for demonstrating that aldehyde-stabilized cryopreservation retains connectomic structure in small and large mammal brains, were explicitly framed around this criterion.
The commercial expression has been unstable. Nectome, a startup founded by one of the prize winners, attracted attention in 2018 for proposing preservation of terminally ill volunteers by a procedure that is necessarily fatal; MIT ended its research association with the company shortly afterwards.6 No preservation provider can demonstrate that the preserved information is sufficient, because no one knows what would be sufficient.
The idea entered wide circulation through fiction — Arthur C. Clarke's stored citizens, Rudy Rucker's Software, Greg Egan's Permutation City — and through advocacy. The Carboncopies Foundation promotes substrate-independent minds as a research programme; Russia's 2045 Initiative announced a staged plan culminating in uploaded consciousness by 2045, with no published technical progress. Economist Robin Hanson's The Age of Em takes emulation as given and works out the economics of a world in which copies are cheap, which is useful as an exploration of consequences and is explicitly not a forecast that emulation will arrive.7
Uploading is also the strongest form of Digital immortality, and the weakest form — a language model trained on someone's messages — is routinely marketed using the same vocabulary. The conflation matters: one is a hypothetical continuation of a person, the other is a stylistic imitation with no claim to continuity.
Beyond feasibility, uploading raises questions with no settled treatment. If copying is cheap, ordinary assumptions about scarcity of persons, consent, and legal identity fail; a copy made without permission is a wrong with no existing category. If emulations can be run at varying speeds, subjective time becomes a resource that can be bought, which connects the topic to Access and inequality. If an emulation can be paused, deleted, or forked, the moral status of those operations depends on whether it is conscious, and no test for that exists. And if uploads are possible at all, questions about ending one's own existence — treated under Right to die and the duty to live — arise in a form the law has never faced, since deletion, suspension, and restoration from backup are all available.
The prior question remains empirical and unanswered: whether there is any level of description at which a brain's causal structure can be captured and re-run, and if so, how far down it sits.
reportSandberg, A. and Bostrom, N. "Whole Brain Emulation: A Roadmap." Technical Report 2008-3, Future of Humanity Institute, University of Oxford, 2008.↩A workshop document setting out what emulation would require at each level of detail; it maps requirements and reports no experiment.
bookMoravec, H. Mind Children: The Future of Robot and Human Intelligence. Harvard University Press, 1988. ↩
paperBédécarrats, A., Chen, S., Pearce, K., Cai, D. and Glanzman, D. L. "RNA from trained Aplysia can induce an epigenetic engram for long-term sensitization in untrained Aplysia." eNeuro, 2018.↩Sea slugs and a simple defensive reflex, not episodic memory; the interpretation that memory is stored outside synapses is disputed.
bookParfit, D. Reasons and Persons. Oxford University Press, 1984. ↩
bookChalmers, D. J. Reality+: Virtual Worlds and the Problems of Philosophy. W. W. Norton, 2022. ↩
newsRegalado, A. "A startup is pitching a mind-uploading service that is '100 percent fatal'." MIT Technology Review, 2018.↩Reports a company's pitch and the university's response; it is evidence about what was offered, not that any mind was preserved.
bookHanson, R. The Age of Em: Work, Love and Life when Robots Rule the Earth. Oxford University Press, 2016. ↩