Connectome-based approaches to whole brain emulation start with a structural map of a brain, then simulate it computationally. The map is a snapshot. Neural activity changes by the millisecond; synaptic weights shift over hours and days; the connectome itself rewires over years. What an upload captures is a single moment, then freezes it.

A 2025 paper by Erik Hoel proposes a formal criterion for consciousness that this approach may structurally fail to satisfy. The criterion is continual learning: real-time, internal-state-modifying adaptation to ongoing experience. Hoel argues this is not a correlate of consciousness but a necessary condition for it, and he provides a formal proof that systems without it cannot satisfy non-trivial consciousness theories.

If the argument holds, the problem is not scan resolution or simulation fidelity. A perfectly faithful static replica may fail to be conscious.

The Core Shift

Hoel’s argument targets large language models as the primary case, but the logical structure applies to any fixed-parameter system, including a connectome snapshot simulated without ongoing adaptation.

The formal structure requires a consciousness theory to satisfy two constraints simultaneously: falsifiability (the theory makes predictions that could be wrong) and non-triviality (the theory does not classify clearly non-conscious systems as conscious). Hoel shows that any function-based theory claiming a frozen LLM is conscious must also claim a lookup table is conscious (non-triviality failure), while any theory adding properties beyond function to exclude the lookup table cannot be tested empirically (falsifiability failure).

Continual learning escapes this trap. A system that modifies its internal states in real time in response to experience is qualitatively different from one that processes inputs through fixed parameters. The difference is not in behavioral output but in the computational process: genuine self-modification versus static computation.

For mind uploading, the Bennett temporal co-instantiation argument provides a parallel constraint from a different direction: consciousness may require simultaneous co-presence of components, which sequential simulation violates. Hoel adds a temporal constraint of a different kind: consciousness may require continuous self-modification, which any snapshot-based approach violates at the moment of capture.

Eon Systems’ fruit fly connectome emulation produced emergent walking, grooming, and feeding behaviors from a static connectome map. Whether that system does anything consciousness-relevant is precisely what Hoel’s criterion makes precise: does it modify its own internal structure in response to its simulated experiences, or does it run the same connectome forward indefinitely?

theconsciousness.ai covers the full formal proof and its implications for LLM architecture in Continual Learning as Necessary Condition for Consciousness, including the specific mathematical structure of the non-triviality and falsifiability constraints.

Comparative Data

Architecture	Internal State Modification	Satisfies Hoel’s Criterion	Consciousness Probability (Hoel)	WBE Relevance
LLM (frozen weights after training)	None after training	No	Near zero	Direct analogy to snapshot upload
Connectome snapshot (static simulation)	None after capture	No	Near zero	Current default WBE approach
Connectome snapshot with online synaptic plasticity	Synaptic weights update during simulation	Partial	Uncertain	Requires plasticity rules captured and re-enabled post-upload
Biological brain during emulation process	Continuous throughout	Yes	Baseline (confirmed conscious)	Source material only
Neuromorphic chip with STDP	Local weight updates in real time at hardware level	Partial	Uncertain	Hardware pathway; requires full-connectome implementation
Gradual neuron replacement (Moravec approach)	Continuous throughout transfer	Yes	High if functionalism holds	Avoids snapshot problem entirely

The table identifies a structural advantage for the Moravec gradual replacement approach: by never creating a frozen snapshot, the biological system’s continual learning is never interrupted. The upload is a process rather than a copy, and the system remains adaptive throughout.

Practical Impact

Hoel’s criterion generates a specific engineering requirement: any upload process that creates a frozen structural map must also include a mechanism for restoring ongoing plasticity before claiming consciousness preservation.

The most plausible implementation route captures plasticity rules alongside connectome structure. A complete WBE would need three layers: the static wiring map (synapse positions), the dynamic weight state (current synaptic strengths), and the adaptive rules (how those weights change in response to experience). Current connectome efforts typically address only the first layer.

Building Brains on a Computer: The 2026 Core Requirements identifies structural fidelity, functional replication, and dynamic adaptability as three non-negotiable WBE capabilities. Hoel’s argument sharpens the third: dynamic adaptability is not optional. Without it, the other two may be achieved at full resolution while still producing a non-conscious system.

The Digital Consciousness Model evaluates consciousness candidates across nine theoretical stances. Under Hoel’s framework, a candidate that lacks continual learning would score poorly on stances requiring temporal self-modification (recurrent processing, predictive processing) while potentially scoring well on static behavioral criteria. That profile mismatch, high behavioral and low architectural, is exactly the pattern found in LLM evaluations under the DCM.

The whole brain emulation roadmap by Sandberg and Bostrom was written before the continual learning criterion was formalized. Updating that roadmap with Hoel’s constraint would add a fourth requirement to structural, functional, and dynamic fidelity: adaptive fidelity, the preservation of the brain’s real-time learning capacity.

Limitations and Open Questions

Hoel’s proof applies within a functionalist framework. Biological naturalists would reject the premise that function-based theories are the right tools for consciousness attribution. If consciousness requires specific physical processes (neurochemical dynamics, quantum effects, metabolic grounding), then neither static nor adaptive digital systems would satisfy the requirement.

The continual learning criterion is coarser than the neuroscience. Biological learning operates at multiple timescales simultaneously: millisecond spike-timing dependent plasticity, hour-scale long-term potentiation, day-to-week structural synaptic changes, year-scale connectome remodeling. Specifying which timescale is consciousness-relevant requires more theoretical precision than Hoel’s formal proof provides.

The strongest practical challenge is measurement. There is no current instrument for determining whether a simulated system is modifying its internal states in a consciousness-relevant way, as distinct from running standard weight updates that satisfy the formal criterion without achieving the phenomenal outcome.

---

Official Sources

• Erik Hoel. “Continual Learning as a Necessary Condition for Consciousness: A Disproof of LLM Consciousness.” 2025. https://erikhoel.substack.com
• theconsciousness.ai analysis: Continual Learning as Necessary Condition for Consciousness
• Michael Timothy Bennett. “A Mind Cannot Be Smeared Across Time.” AAAI 2026. arXiv:2601.11620
• Sandberg, A. & Bostrom, N. “Whole Brain Emulation: A Roadmap.” Future of Humanity Institute Technical Report, 2008.
• M. Schons. “Building Brains on a Computer: The Three Core Requirements.” Asimov Press, 2026. DOI: 10.62211/92ye-82wp