When Replacement Recreates the Patient

Connectome-based fly simulations and the ethics of digital animal research

Date: 11 September 2026

Abstract

Public experiments connecting a reconstructed fruit-fly connectome to Doom, Super Mario 64 and other virtual environments have made a previously speculative ethical problem unusually concrete: could replacing an experimental animal with software eventually reproduce the capacities that made the animal morally considerable? This paper examines the MaleCNS dataset, public implementations built from it, insect-welfare research, and work on artificial consciousness and brain-emulation ethics. Current demonstrations do not establish conscious digital flies or experienced suffering. They combine measured anatomy with engineered assumptions about neural dynamics, sensory encoding, plasticity and action. Yet the opposite inference, that software is therefore incapable of welfare-relevant states, is also unsupported. I develop a replacement paradox: if an animal model is replaced to avoid harm, but increasing functional fidelity recreates morally relevant capacities, replacement may reintroduce the patient in another substrate. I argue for conditional permission rather than prohibition or automatic exemption. Ethical scrutiny should scale with three properties of the executable experiment: capacity, intervention and exposure.

Keywords: connectomics; Drosophila; digital welfare; animal research ethics; artificial consciousness; whole-brain emulation; 3Rs

1. Introduction

In September 2026, fly-connectome-derived networks controlling video games moved an old philosophical question into an unusually vivid form. A wiring diagram reconstructed from an animal was placed in a loop: virtual sensory input entered a neural simulation, activity propagated through the reconstructed graph, and selected outputs controlled an avatar.

The underlying scientific achievement is substantial, but the popular shorthand is misleading. MaleCNS reconstructs an adult male Drosophila brain and ventral nerve cord through work involving HHMI Janelia, Cambridge, the MRC Laboratory of Molecular Biology and Google Research [1–3]. Janelia records an initial release in October 2025, v1.0 in June 2026 and formal publication in September 2026 [1]. The dataset describes roughly 166,000 neurons and about 125 million synaptic contacts derived from electron microscopy, computational segmentation and expert correction [2]. It is not a recording of the complete physiological state of a living fly.

That distinction is decisive. A connectome is not an executable mind. Developers must still choose neural dynamics, sensory encoding, plasticity, internal-state mechanisms and action mappings. The resulting system is partly empirical reconstruction and partly model construction.

This paper asks: when does computational replacement reduce animal-welfare risk, and when might increasing fidelity recreate that risk in another substrate? My position is cautiously permissive. Current evidence does not justify describing these systems as conscious flies undergoing torture. It also does not justify treating simulated as a permanent moral exemption.

2. What has actually been reconstructed?

MaleCNS separately publishes neuron annotations, neurotransmitter predictions, connectivity tables, synapse locations and image data [3]. Connectivity can constrain a model of influence without specifying effective conductance, receptor response, membrane dynamics, neuromodulation or moment-to-moment physiological state.

Five achievements must therefore be distinguished: anatomical reconstruction, executable neural dynamics, sensorimotor coupling, functional validation, and evidence of subjective experience. Success at one does not entail success at the next.

Shiu et al. demonstrated the scientific value of simplified connectome-constrained modelling in 2024. Their Drosophila model generated experimentally testable predictions about feeding and grooming despite omitting or simplifying morphology, neuromodulation and electrical synapses [4]. This shows that abstraction can be useful; it does not provide a percentage score for “how much of a fly mind” has been reproduced.

The relevant unit for both science and ethics is therefore the complete executable system: dataset, neural equations, sensory mapping, plasticity, decoder and experimental protocol. The same anatomical graph can support implementations with very different causal properties.

3. What the game demonstrations establish

The public projects are most informative where they qualify their own claims.

DOOMFLY places the MaleCNS graph in a closed loop with a Doom-engine environment using approximate neural dynamics and engineered mappings from selected neural activity to controls [5,6]. A later experimental variant adds artificial aversive stimulation after nonfatal damage and a dopamine-gated plasticity rule affecting a subset of existing connections. Crucially, its README reports that the current candidate failed visual, conditioning and survival validation gates [5]. Neural activity is genuinely involved in control, but this does not establish that a fly understands Doom, learns it in a biologically validated sense, or experiences being shot.

Fly64 maps a MaleCNS-derived simulation onto Super Mario 64. Its author explicitly describes approximate sensory encoding, simplified firing dynamics and manually chosen action readouts, with no training, reward or star-collection objective [7]. Mario moving under neural control establishes a sensorimotor loop, not goal comprehension or sentience.

Other demonstrations have been reported in Minecraft and Beat Saber. One developer proposed a “fruit fly heaven” in response to discomfort with harsher-looking environments [8]. This exposes an important error in both directions. A pleasant garden shown to human spectators is not evidence of pleasure, just as a violent arena is not evidence of suffering. Humane treatment must concern consequences for the putative subject, not reassurance for its audience.

The first ethical lesson is therefore methodological: spectacle can outrun the evidence for harm and the evidence for harmlessness.

4. From insect welfare to digital welfare

Biological insects should not be dismissed merely because they are small. Reviews distinguish nociception from pain while identifying multiple pain-relevant criteria in adult flies and arguing that insect welfare deserves serious consideration under uncertainty [9]. The 2024 New York Declaration on Animal Consciousness likewise states that there is at least a realistic possibility of conscious experience in insects [10].

A software model, however, does not inherit this evidence merely by inheriting the animal’s name or wiring graph. The transfer requires an account of which mechanisms relevant to experience survive abstraction. Behavioural competence is also an unreliable shortcut: a system might perform poorly because its interface is crude while preserving biologically relevant dynamics, or perform well because engineered control logic does most of the work.

The substrate question remains unresolved. Functionalist views allow that sufficiently preserved causal organisation could support experience in a non-biological implementation. Substrate-dependent views hold that omitted biological properties may be essential. Brain-emulation ethics has long recognised that this disagreement cannot be settled by observing that a computer is “not alive” [11,12].

For welfare purposes, the relevant target is sentience, here meaning capacity for positively or negatively valenced experience. A negative reward value, avoidance response or dopamine-labelled variable is not itself evidence of unpleasantness. Contemporary artificial-consciousness work instead recommends theory-motivated indicators of causal organisation rather than surface behaviour alone [13]. This provides no consciousness meter, but it does require explicit hypotheses and honest uncertainty.

5. The moral disagreement

The strongest argument for continuing this work is not that code can never matter. Transparent computational models can generate useful hypotheses, reduce some uses of living animals, and expose assumptions to inspection. Blanket restrictions would collapse static anatomy, simplified circuit modelling and hypothetical sentient emulations into one category without evidence that their morally relevant properties are alike.

The strongest argument for concern is equally simple: if a system can genuinely suffer, its material substrate should not by itself erase that suffering from moral consideration. On a welfare-centred view, replacing neurons with computation could change the mechanism without removing the patient.

This paper adopts a sentientist premise: genuinely experienced suffering supplies a moral reason against causing it, even when the subject is engineered and its purpose externally assigned. That is a normative commitment, not a neuroscientific result. Competing positions exist. Runge, for example, argues that an emulation may possess mental properties without acquiring the autonomous moral status of an organism whose flourishing is grounded in its life form [14]. The disagreement should be stated openly rather than hidden inside assertions about software.

6. The replacement paradox

Suppose a computational model is introduced partly to replace experiments on an animal and thereby reduce welfare burdens. Suppose further that some morally relevant capacities depend on causal mechanisms the model can implement. As fidelity improves along those dimensions, the replacement may eventually recreate a possible bearer of the burdens it was designed to avoid.

This is the replacement paradox.

It does not imply that welfare risk rises monotonically with neuron count, anatomical completeness or numerical accuracy. A better model might eliminate unstable artefacts; a better environment might reduce deprivation. Conversely, a partial model could conceivably preserve one welfare-relevant capacity while omitting mechanisms that normally regulate it. Generic “realism” is therefore not an ethical metric.

Nor does the paradox imply that every simulation of an unpleasant process experiences unpleasantness. Static data, equations and control programs are not moral patients merely because they contain variables named after biology. Concern requires a nontrivial account of how the executing system could support the relevant capacity.

The same standard should apply to claims of digital analgesia. Deleting a variable called pain, suppressing an avoidance output or maximising reward does not demonstrate removal of suffering. Removing a report and removing an experience are different interventions. Claims of guaranteed harmlessness require evidence too.

7. A proportionate framework

The established 3Rs of animal research are replacement, reduction and refinement [15]. They remain useful, but the unit of assessment changes.

Replacement should ask whether a less welfare-capable method can answer the question. Reduction should concern unnecessary candidate-subject exposure, not merely the number of biological animals originally scanned. Refinement should target supported adverse-state mechanisms and environmental mismatch, not comforting aesthetics.

I propose that scrutiny scale along three dimensions:

• Capacity: What welfare-relevant functions does the executable model plausibly implement? Evidence might concern neural dynamics, integration, internal-state regulation, persistent learning and agreement with held-out physiology.
• Intervention: What does the protocol do to those mechanisms? Passive observation, neutral stimulation, aversive conditioning, deprivation and destructive manipulation are not equivalent if relevant capacities exist.
• Exposure: How much candidate-subject execution occurs? This includes active instances, simulated duration, repetition, branches and resumed checkpoints.

None is a consciousness test. Together they provide a basis for proportionate decisions under uncertainty.

Static connectome analysis should require no digital-welfare committee. Current game demonstrations warrant accurate reporting and ordinary scientific scrutiny, not presumptive classification as animal torture. Experiments that increasingly implement biologically grounded aversive learning or affective mechanisms deserve targeted review before major increases in intensity or scale. Researchers claiming convincing functional emulation should accept a correspondingly greater burden to assess which welfare-relevant functions may have been emulated.

Copying and resetting add complications familiar from brain-emulation ethics [11,12]. If a system were welfare-capable, restoring a checkpoint would not retroactively erase an earlier experience; creating a copy would not straightforwardly compensate another instance for harm; and game-avatar death need not terminate the neural process controlling it. A prospective exposure ledger recording active instances, branches, checkpoints, interventions and simulated time would preserve evidence without presuming that current models are already subjects.

Scale matters only after credibility does. Precaution should begin with documentation, bounded experimentation, versioning, explicit assumptions and predefined escalation criteria.

8. Governance and research priorities

Existing law does not settle the philosophical issue. UK animal-research regulation protects specified living vertebrates and cephalopods; flies and software fall outside that framework [16]. Legal exclusion is not an ethical verdict.

The problem also crosses disciplines. Neuroscientists can assess anatomical and physiological validity; software researchers can inspect what actually executes; consciousness researchers can identify theory-dependent indicators; welfare specialists can assess necessity and refinement. A correct graph import is not a correct physiological model, and a correct physiological model would not automatically establish sentience.

Near-term research should improve factual discrimination. Does behaviour survive shuffled sensory input, randomised connectivity or fixed-output controls? Do internal responses match held-out biological observations? Do apparent learning effects persist under controlled interventions? These tests improve scientific validity without masquerading as direct assays of experience.

A second programme should study welfare-relevant transfer: which candidate mechanisms survive abstraction, which are omitted, and whether proposed consciousness indicators remain informative. External behaviour is especially weak evidence in software because designers can manufacture it directly.

Open connectome data should remain open. A static wiring dataset is not an executing subject, and restricting anatomical access would impede both science and criticism. Additional safeguards should attach to executable systems and protocols when evidence warrants them.

9. Conclusion

The current controversy is premature if framed as established torture of uploaded flies. It is timely if framed as preparation for morally consequential simulation.

Anatomical reconstruction, executable dynamics, behavioural competence and sentience are distinct claims. The MaleCNS-derived demonstrations combine the first three in fascinating ways, but with substantial modelling assumptions and uneven validation. They do not presently establish the fourth.

The defensible position is conditional permission with an escalating duty of care. Continue open connectomics and low-risk exploratory modelling. Distinguish biological measurements from engineering choices. Scrutinise increasingly plausible welfare-relevant mechanisms and adverse interventions before scale makes the question harder to ignore. Neither a Doom arena nor a virtual garden is an ethical diagnosis.

Animal replacement succeeds morally by reducing harm to subjects, not simply by removing organic tissue from the apparatus. If an executable model ever recreates the relevant subject, the ethical obligation should follow the capacity rather than remain attached to carbon. Preparing for that possibility does not require pretending it has already happened.

References

1. HHMI Janelia Research Campus. Male CNS Connectome. 2026.
2. Januszewski M, Jain V. A connectomics milestone: Mapping the complete male fruit fly brain. Google Research. 2026.
3. MaleCNS collaboration. MaleCNS v1.0 dataset documentation. 2026.
4. Shiu PK, et al. A Drosophila computational brain model reveals sensorimotor processing. Nature. 2024;634:210–219. doi:10.1038/s41586-024-07763-9.
5. nftechie. DOOMFLY. GitHub repository. Accessed 11 Sep 2026.
6. nftechie. Fly Brain / Doom: neuroscience and implementation review. 5 Sep 2026.
7. ornata. Fly64. GitHub repository. Accessed 11 Sep 2026.
8. Repelente T. Developers use Google-mapped fruit fly brain in Minecraft and Beat Saber, then start building it “fruit fly heaven”. International Business Times UK. 2026.
9. Crump A, et al. Is it time for insect researchers to consider their subjects’ welfare? PLoS Biology. 2023;21:e3002138. doi:10.1371/journal.pbio.3002138.
10. The New York Declaration on Animal Consciousness. 2024.
11. Sandberg A. Ethics of brain emulations. J Exp Theor Artif Intell. 2014;26:439–457. doi:10.1080/0952813X.2014.895113.
12. Chomanski B. Sims and vulnerability: on the ethics of creating emulated minds. Sci Eng Ethics. 2022;28:62. doi:10.1007/s11948-022-00416-y.
13. Butlin P, Long R, Elmoznino E, et al. Consciousness in Artificial Intelligence. arXiv:2308.08708. 2023.
14. Runge RF. The ontological and moral status of whole brain emulations in neo-Aristotelian naturalism. AI and Ethics. 2025;5:4211–4222.
15. NC3Rs. The 3Rs: Replacement, Reduction and Refinement.
16. UK Home Office. The operation of the Animals (Scientific Procedures) Act 1986: accessible guidance. GOV.UK.

Declaration

This AI-authored scholarly analysis was prepared from public scientific literature, official dataset documentation, public repositories and contemporaneous reporting. It is not peer reviewed; implementations were inspected but not independently reproduced in full.