Link to the code: brain-emulation GitHub repository

The Physics Wall: Why Non-Destructive Brain Scanning Cannot Upload a Living Mind

Every serious discussion of mind uploading eventually collides with the same constraint: how do you read the information out of a living brain without destroying it? The destructive approaches, slicing tissue into nanometer-thin sections and imaging each one with electron microscopy, yield extraordinary structural data. But they require a dead brain. They cannot preserve the person whose mind is being mapped.

Non-destructive scanning would solve this. If you could image a living brain at the resolution needed to capture its full synaptic architecture while it remained alive and conscious, you would have the raw data for an upload that avoided the copy-and-kill problem. The person could in principle survive the procedure with their subjective continuity intact.

The physics of doing this runs into walls that are not engineering obstacles awaiting better hardware. They are rooted in the fundamental properties of the imaging physics involved.

Technology Readiness Level: TRL 1 (basic principles observed; no scanning modality capable of yielding upload-quality data in living tissue exists at the theoretical level, let alone in practice).

What Non-Destructive Scanning Would Require

Before examining what the physics prohibits, it helps to be precise about what upload-quality brain scanning would need to achieve.

The whole brain emulation roadmap identifies three types of data a credible emulation requires: structural connectivity (the wiring diagram of which neurons connect to which, with what synaptic weight), molecular state (neurotransmitter concentrations, receptor distributions, gene expression in individual cells), and dynamic state (the moment-to-moment pattern of neural activity that constitutes ongoing cognition).

A non-destructive scanner would need to capture all three simultaneously in a living brain. It would need spatial resolution at the synaptic level, roughly 10 to 100 nanometers for synaptic cleft geometry and 1 to 10 micrometers for dendritic spine morphology. It would need temporal resolution sufficient to track neural firing, on the order of milliseconds. And it would need to do all of this across the full volume of a human brain, approximately 1,200 cubic centimeters, without depositing harmful doses of radiation into the tissue or using contrast agents that require invasive delivery.

No existing modality achieves more than two of these requirements at once. The reason is not inadequate engineering. It is that the physical properties that give scanners spatial resolution tend to work against the properties that enable temporal resolution and non-invasiveness, and vice versa.

Functional MRI: The Resolution Floor

Functional magnetic resonance imaging (fMRI) is the dominant non-invasive tool for studying human brain function in living subjects. Its strengths are well established: whole-brain coverage, no ionizing radiation, spatial resolution now reaching sub-millimeter levels with ultra-high field magnets at 7 Tesla or above. It has been the backbone of human cognitive neuroscience for three decades.

Its fundamental limitation is that it does not measure neural activity directly. It measures the blood oxygen level dependent (BOLD) signal, which reflects changes in the ratio of oxygenated to deoxygenated hemoglobin in local vasculature. Neural firing increases metabolic demand, which triggers localized increases in blood flow. The BOLD signal is a hemodynamic proxy for neural activity.

The hemodynamic response has an intrinsic temporal delay of 1 to 6 seconds. Neurons fire on the millisecond timescale. The BOLD signal smears millisecond-scale neural dynamics into a response that peaks several seconds after the triggering activity. This temporal blur is not eliminable by improving scanner hardware. It reflects the physiology of blood flow, which cannot be accelerated without changing the biology.

For upload purposes, this is disqualifying at the functional level. An upload requires capturing what a brain is doing at the timescale of neural dynamics, not a hemodynamic echo of what it did several seconds ago. fMRI cannot provide this, regardless of spatial resolution improvements.

At the spatial level, even 7 Tesla fMRI achieves approximately 0.5 millimeter isotropic resolution under optimal conditions. Individual neurons are 10 to 100 micrometers in diameter. Synaptic spines are smaller still. The gap between the best fMRI spatial resolution and synaptic resolution is roughly three to four orders of magnitude. Achieving synaptic-level spatial resolution in fMRI would require magnetic field strengths that would induce dangerous induced currents in neural tissue.

Magnetoencephalography: Superb Timing, Zero Depth

Magnetoencephalography (MEG) measures the extremely weak magnetic fields generated by synchronized neural currents. Unlike fMRI, it tracks neural activity directly with millisecond temporal resolution. Studies using modern MEG systems with superconducting quantum interference device (SQUID) sensors or, more recently, optically pumped magnetometers (OPMs) operating at room temperature, have demonstrated genuine millisecond tracking of distributed cortical dynamics.

MEG’s limitation is spatial and geometric. Magnetic fields from neural currents fall off rapidly with distance from the source, following an inverse square relationship. MEG sensors are positioned outside the skull. The signals they detect are dominated by currents in superficial cortex where the geometry is favorable for external detection. Deep brain structures, including the hippocampus, basal ganglia, thalamus, and brainstem, are essentially invisible to MEG because their magnetic fields by the time they reach external sensors are below the noise floor.

Source localization in MEG is also fundamentally ill-posed. Multiple distinct configurations of neural current sources can produce identical (or arbitrarily similar) magnetic field patterns at the sensor array. Reconstruction algorithms make assumptions about source geometry to select a plausible solution, but the selection is not unique. The spatial resolution of MEG source reconstructions is approximately 5 to 10 millimeters under favorable conditions, worse in deep structures and in practice often worse than that figure.

For upload purposes, MEG provides the temporal resolution required but lacks the spatial resolution by four orders of magnitude and cannot image deep structures at all. It is an excellent tool for studying cortical dynamics but cannot read the structural connectivity or molecular state that an upload would need.

Functional Ultrasound: The Most Promising Non-Invasive Path

Functional ultrasound (fUS) imaging is a newer modality that has attracted substantial attention in the neuroscience and BCI communities since approximately 2014. It uses high-frame-rate plane wave ultrasound to track cerebral blood volume changes with temporal resolution of tens of milliseconds and spatial resolution approaching 100 micrometers, both significant improvements over fMRI for tracking hemodynamic proxies of neural activity.

In open-skull animal preparations, fUS has demonstrated imaging of whole-brain hemodynamic activity at 100 micrometer resolution and 10 millisecond temporal sampling. Researchers at the Institut de la Vision in Paris and subsequently at Caltech have used fUS to decode motor intentions in non-human primates and to image neural activity across brain areas during behavior in rodents.

The limitations are significant. Like fMRI, fUS measures hemodynamics, not direct neural activity. The hemodynamic response delay problem applies, though at a shorter timescale than BOLD. More critically, ultrasound imaging requires a clear acoustic window: the skull strongly attenuates and scatters ultrasound, making transcranial imaging through intact human skull at the resolutions achieved in open-skull animal preparations currently impossible. Most high-resolution fUS demonstrations in non-human primates require partial craniectomy or thin skull windows.

Transcranial focused ultrasound (tFUS), explored for consciousness research and increasingly for therapeutic applications, uses phased arrays and adaptive focusing to work through the skull. The therapeutic use of focused ultrasound for consciousness research has advanced substantially. But achieving 100 micrometer resolution through an intact adult human skull transcranially remains beyond current capabilities. The skull’s spatial variability, bone density heterogeneity, and acoustic impedance mismatches create aberrations that current correction algorithms cannot fully compensate.

Even at its theoretical best in open-skull conditions, fUS does not reach synaptic resolution. 100 micrometers contains tens to hundreds of neuronal cell bodies. It cannot distinguish individual synaptic weights.

The Fundamental Trade-Off

The pattern across these modalities reflects a fundamental trade-off in the physics of non-invasive imaging.

Ionizing radiation (X-rays, gamma rays) can achieve nanometer-scale spatial resolution and penetrates tissue. But the doses required for high-resolution whole-brain imaging would cause extensive DNA damage and cell death. Medical CT at clinical doses cannot resolve synaptic structures. Pushing toward synaptic resolution with ionizing radiation would be radiobiologically lethal.

Visible and near-infrared light (used in two-photon microscopy and light sheet fluorescence microscopy) achieves the spatial resolution needed for synaptic imaging. In living tissue, it has been used to image dendritic spines at single-micron resolution in thin cortical windows. But light scatters in brain tissue over distances of approximately 1 millimeter. Imaging more than about 1 millimeter below the surface in scattering tissue requires specialized approaches (adaptive optics, three-photon microscopy) that remain limited in depth penetration. A human brain is 12 to 15 centimeters across. The depth penetration gap is orders of magnitude.

Ultrasound and radiofrequency waves (MRI) penetrate tissue well but lose spatial resolution as the physics of wave propagation in heterogeneous biological media limits focusing. The Rayleigh criterion defines the diffraction-limited resolution of any wave-based imaging system as approximately half the wavelength of the imaging wave. Imaging at synaptic resolution requires wavelengths shorter than synaptic structures, roughly 100 nanometers. Electromagnetic waves of that wavelength (soft X-rays) do not penetrate more than a few micrometers of biological tissue.

The trade-off is not escapable with current physics. Waves that penetrate do not resolve synapses. Waves that resolve synapses do not penetrate. There is no known physical principle that bypasses this, and no credible theoretical proposal exists for doing so.

What This Means for Upload Feasibility

The physics wall on non-destructive scanning matters for the mind uploading reality check in a specific way. It means that the “gradual upload while alive” scenario, which is the only upload pathway that preserves subjective continuity without copying and killing, faces a barrier that is not an engineering problem.

The Moravec transfer concept, which proposes replacing neurons one at a time with functional equivalents, is one proposed workaround. It sidesteps non-destructive scanning by replacing the brain incrementally rather than copying it all at once. But this requires a different set of technologies. It does not eliminate the scanning problem; it replaces it with a replacement problem.

The mammal brain preservation breakthrough from Song et al. advances the destructive pathway: post-mortem preservation at high fidelity for subsequent structural mapping. This is scientifically significant but does not address the problem of uploading a living mind.

Non-destructive upload of a living brain, at the resolution an upload would require, is not currently feasible and does not have a clear physical pathway to becoming feasible. This is distinct from saying it is impossible in principle: new physics, new imaging modalities, or circumvention strategies like incremental replacement could change the picture. But the resolution and penetration requirements are stringent enough that the burden of proof on any proposed solution is high.

Path Forward

Research directions that could make progress on the physics of non-destructive neural imaging include:

Quantum sensing. Nitrogen-vacancy (NV) centers in diamond can detect single nuclear spins in close proximity and have been used to image single neurons in vitro. Scaling NV-based sensors to whole-organ imaging remains a fundamental challenge, but the underlying physics does not prohibit single-neuron resolution in principle.

Neural dust. MIT and UCSF researchers have explored implantable wireless sensor networks at the scale of tens of micrometers that could in principle record local field potentials from distributed brain regions without requiring large implant footprints. Scaling this to whole-brain synaptic resolution recording would require orders of magnitude more sensors than have been demonstrated and raises its own biocompatibility questions.

Brain-computer interface bandwidth is a related constraint. Even if sensors could record at synaptic resolution, transmitting the resulting data out of the brain is a physically separate challenge with its own constraints.

The honest assessment for 2026 is that non-destructive upload of a living brain at the required resolution is not technically blocked by insufficient effort or funding. It is blocked by physically constrained imaging trade-offs that have not yielded to decades of research in medical imaging, basic physics, and neurotechnology. Moving beyond those constraints will require either new physics or a fundamental rethinking of what upload-quality data actually needs to contain.

Official Sources