In May 2024, Harvard and Google published in Science the most detailed reconstruction of human brain tissue ever created: a single cubic millimeter of cortex, roughly half the size of a grain of rice. That fragment contained 57,000 cells, 150 million synapses, and 230 millimeters of blood vessels, all rendered at nanoscale resolution. The raw dataset required 1.4 petabytes of storage. The human brain is approximately one million times larger. Jeff Lichtman, who led the Harvard side, noted that imaging an entire human brain at this resolution would generate data roughly equivalent to the world’s total annual data production. The next planned step is a 10-cubic-millimeter section of mouse hippocampus, and even that will take years.
This is the empirical starting point for mind uploading: the most ambitious mapping effort in neuroscience history has reconstructed one-millionth of a single brain, and the structures it found were full of surprises that contradicted textbook models. The distance between this achievement and whole brain emulation is not merely quantitative. It is qualitative. We do not yet know what level of detail is sufficient for functional emulation, which means we do not know how far we are from the destination.
The State of Brain Emulation Report 2025 (Zanichelli et al.), the first comprehensive reassessment of the field since Sandberg and Bostrom’s 2008 roadmap, organized the challenge into three pillars: recording neural dynamics (functional activity), mapping brain structure (connectomics), and computationally simulating the result. On neural dynamics, no experiment has achieved whole-brain recording at single-neuron, single-spike resolution in any organism. The closest results are larval zebrafish at approximately 80 percent brain coverage, with temporal resolution well below neuronal firing rates. On connectomics, complete synaptic-resolution maps exist only for C. elegans (302 neurons, ten reconstructions) and adult Drosophila (roughly 140,000 neurons). The cost per reconstructed neuron has fallen from an estimated $16,500 in the original C. elegans work to roughly $100 in recent zebrafish projects, but proofreading, the manual correction of automated neuron tracing, remains the dominant bottleneck, particularly for the large, morphologically complex neurons of mammalian brains. On computational simulation, embodied models of C. elegans now reproduce specific behaviors, and Drosophila whole-brain models recapitulate known circuit dynamics. GPU-cluster feasibility studies have demonstrated simulations approaching human-brain scale, but only with drastically simplified biophysical assumptions. The report estimates that fewer than 500 people globally work directly on brain emulation. Beyond structural mapping, the report flagged neuromodulation as a distinct and potentially deeper challenge: recording modulatory molecules like serotonin, dopamine, oxytocin, and neuropeptides is even harder than recording neural spikes, and these systems profoundly shape mood, motivation, personality, and subjective experience in ways a connectome alone cannot capture. The Carboncopies Foundation, which coordinates much of the field, launched a Brain Emulation Challenge modeled on ImageNet: as of late 2025, no validated working model of any complete brain exists, not even for organisms with a few hundred neurons.
The philosophical challenges may be harder than the technical ones. David Chalmers’ analysis of mind uploading identifies two distinct questions: Will the upload be conscious? And will it be me? On the first question, Chalmers argues via organizational invariance: if two systems have the same functional organization, they have the same conscious states. Gradual neuron-by-neuron replacement with functionally equivalent silicon circuits should preserve consciousness at every step. This is the functionalist case, and it is powerful. The embodied cognition tradition, following Thompson, Varela, and Rosch’s The Embodied Mind (1991), challenges the premise itself: cognition is constitutively shaped by having a body that moves through a physical environment, by hormonal feedback, gut-brain signaling, proprioception, and immune responses, none of which a connectome-based emulation would preserve. The second question is harder. Even granting that the upload is conscious, a destructive scan-and-copy procedure creates what Chalmers calls the “twin” problem. A molecule-for-molecule duplicate is qualitatively identical but not numerically identical. If both the original and the copy exist, the copy is clearly a distinct person. If the original is destroyed, whether the copy constitutes survival or sophisticated death depends on a question about personal identity that philosophy has not resolved.
Many people intuitively prefer gradual replacement over scan-and-copy, believing the continuous process preserves identity in a way that destruction and reconstruction does not. Wiley and Koene argued in a 2016 paper in the Journal of Consciousness Studies that this intuition is a fallacy: there is no principled metaphysical distinction between the two procedures. If instantaneous in-place replacement preserves identity, destructive scan-and-copy should as well, because the physical result is identical. If scan-and-copy fails to preserve identity, gradual replacement faces the same challenge at a finer grain. The philosophical problem cannot be dissolved by engineering a smoother transition. It persists regardless of method.
A 2025 expert survey (Caviola and Saad, Oxford) gathered forecasts from 67 researchers across digital minds research, AI, philosophy, and forecasting. The median estimated probability of digital minds (defined as computer systems capable of subjective experience) being created by 2030 was 20 percent, rising to 50 percent by 2050 and 65 percent by 2100. Critically, 89 percent judged digital minds unlikely to arrive before AGI. The survey also found that machine learning systems are considered more likely to be the first digital minds than brain emulations, a finding that shifts the mind uploading question: the first digital consciousness may not be a human mind transferred to silicon but an artificial system that develops (or appears to develop) subjective experience on its own. Sandberg himself has estimated 50 percent confidence in whole brain emulation arriving by 2064.
If mind uploading becomes feasible, the societal implications restructure everything: death becomes optional for those who can afford the procedure, population becomes a policy variable rather than a demographic fact, and legal and moral frameworks built on the assumption of a single continuous embodied life require reconstruction from foundations. Robin Hanson’s The Age of Em (2016) remains the most detailed economic analysis: copyable minds drive wages to computational subsistence (the cost of running a copy), faster-clocked emulations out-compete slower ones, and the resulting economic dynamics resemble biological evolution more than human labor markets. The governed outcome (+3) reflects the possibility that proactive preparation, including legal frameworks for digital personhood, rights protections for uploaded minds, and access policies preventing uploading from becoming an immortality technology reserved for the wealthy, could turn this into a managed transformation. The unmanaged outcome (-2) reflects the more likely scenario: the technology arrives (if it arrives) into a world with no legal infrastructure for digital persons, no consensus on whether uploaded minds are the original person, and intense commercial pressure to offer the procedure before the philosophical and regulatory questions are settled.
Key tension: Whether mind uploading is immortality or sophisticated suicide depends on a question about consciousness that philosophy has not resolved in three thousand years. The technology may force the answer before the question is ready.