Last Challenge (5). Could a brain be reproduced in software? What would it take to emulate a brain?

Series “The Last Challenge”, article 5. Whole brain emulation aims to reproduce a particular brain in software. Learn why a wiring map is not a working mind, what a successful test would require, and why memory, identity, and consciousness remain open questions.

Last Challenge (5). Could a brain be reproduced in software? What would it take to emulate a brain?
Illustration generated with ChatGPT. Can we reproduce not just the structure of a brain, but the processes that make intelligence possible?

Who should read it?

AI practitioners, technical leaders, and curious non-specialists

Key message

Can we reproduce not just the structure of a brain, but the processes that make intelligence possible?

Intro

Imagine a computer model that talks like someone who has died, recognizes their friends, and recalls a favorite place.

  • "Would it be the person returning, or a new system that resembles them?"

We do not know how to build such a model of a human brain.

Bostrom discusses whole brain emulation as a possible route to machine intelligence. Rather than invent an intelligent system from first principles, researchers would try to reproduce the processes of a particular biological brain in software.

The proposal is speculative. A successful human emulation, preservation of its memories, continuity of the original person's experience, and consciousness in the resulting system are all unresolved. Those questions are related, but answering one would not automatically answer the others.

A scan is the start of a model, not a mind

A brain contains nerve cells, called neurons, which communicate through connections including synapses. A connectome maps connections in a nervous system, helping researchers trace potential signal pathways, much as a circuit diagram shows connected components.

A map alone does not tell us what happens when a signal travels through the circuit. Connections can differ in their effects, and cells have changing internal states. A functional emulation would need enough relevant biological structure and dynamics to reproduce the modeled brain's behavior.

Which details are necessary, and whether a scan can recover them, remain research questions.

Anders Sandberg and Bostrom's 2008 whole brain emulation roadmap separates the proposed work into scanning, translation, and simulation. The sequence sounds tidy. Each stage introduces uncertainties that the next one inherits.

Scanning would capture a brain at sufficient resolution for the model. One proposed approach would preserve the tissue, cut it into thin sections, and image them. That would destroy the biological brain, though not every conceivable scanning method must be destructive. Preservation could also lose needed information.

Translation would reconstruct cells and connections from images and estimate properties affecting their behavior. Imagine a faint gap in a neuron across two adjacent sections. An incorrect join could create a connection that was never there, and the error may be hard to find if the model still produces plausible activity.

Simulation would run the reconstructed components and their interactions. It needs computing resources and a way to provide inputs and receive outputs, perhaps through a simulated body or environment. Faster hardware cannot compensate for a missing connection or an incorrect biological rule. A detailed dataset does not establish that a model can run at a useful speed.

A useful test asks what the model preserves

Progress in mapping is real, but it is not a successful upload. In a 2024 paper in Nature, the FlyWire team reported a wiring diagram of an adult fruit fly brain. The authors describe limitations, including connections that may be missing. It is a research map, not a demonstrated emulation of that individual fly's mind.

A map may reveal a pathway without reproducing an individual's behavior. A simulation may reproduce some responses without retaining learned skills. Similar output is evidence for a particular claim, not a certificate of complete fidelity.

Consider an invented experiment with an animal trained to respond to a cue. Researchers record its responses before scanning, build a model, and test it with that cue and unfamiliar variations. Agreement on the variations would offer stronger evidence of preservation beyond a memorized response. Researchers would still need controls against accidental matches, comparisons across tasks, and an account of what the test did not measure.

Memory is especially difficult. A scan or reconstruction may miss relevant features. A correct answer might reflect a preserved state, information supplied during testing, or a plausible guess. Researchers would need comparisons to distinguish these possibilities. Whether a human emulation could retain particular memories remains unknown.

When assessing an upload claim, ask what was built, which measurements were held back for evaluation, and which uncertainties entered during reconstruction.

Recognition does not settle identity or consciousness

Suppose a future system passes many behavioral tests, describing a shared holiday and recognizing an old friend. That would not resolve whether the original person's subjective experience continued in the simulation. If the biological original remained alive beside the model, both might claim the same past while having different experiences after copying. Behavioral similarity alone cannot tell us which, if either, is the original in the sense of continuous experience.

Consciousness is a separate uncertainty. A system that talks about pain may or may not have an experience of pain. We do not have an agreed test that settles whether a brain emulation would be conscious. Nor should we assume that lack of certainty proves it could not have experiences.

These distinctions change the ethical stakes. If emulations could have experiences, creating copies to perform work, interrupting them, or deleting them might affect beings with interests of their own. If they are not conscious, the same actions would have a different moral status. Either way, the person whose brain supplied the data has interests in consent, privacy, and the possible uses of that data. A promise of digital immortality would go far beyond what the technology or the identity argument currently establishes.

Mind Crime

The term mind crime describes possible harm to simulated beings with morally relevant experiences. It depends on whether such beings can exist and what moral status they have. Bostrom's voluntary slavery scenario is different:

  • Selected or modified copies would willingly work for subsistence.

Their willingness, not the original donor's agreement, is central to the scenario. Selection or modification also raises questions about exploitation and meaningful consent. These are ethical problems to consider, not claims that human uploads exist.

The animated series Pantheon treats uploading as fiction. Its characters confront a destructive scanning procedure and an uploaded person's confinement to work. It illustrates the concerns. It is not evidence that the technology works.

Copying intelligence would also copy responsibility

An emulation could be valuable as a research model even if it fell short of reproducing a person. Projects should specify consent, data access, independent validation, and limits on use. Brain-derived data may contain sensitive information about a person's life.

Bostrom also considers copying digital minds and running them on computing hardware, conditional on successful emulation, which has not been demonstrated for humans. Copies would still require hardware, energy, and a working environment. They would not be free or guaranteed to run faster than biological brains. Nor would a human-like starting point automatically make a system safe or superintelligent.

The appeal of whole brain emulation is that it might let researchers reproduce intelligence without first explaining every aspect of it. The cost is a different set of hard questions:

  • "What the scan missed?"
  • "what the model actually preserved?"
  • "what obligations we would have if the result could experience the world?"

Previous

Article 4: Should an AI answer, act, or run the show?

Next

Article 6: How quickly could AI improvement accelerate?