Is It Wrong to Switch Off a Biological Computer?

Switching off a biological computer is, on current evidence, not wrong in itself. It means ending a living neural culture, and today’s brain organoids show no sign of the integrated circuitry that pain or suffering appears to require. But “probably not sentient” is not “certainly not,” so responsible labs plan retirement as carefully as they plan training.

Key takeaways

“Switching off” a biocomputer is a biological event, not an electrical one: the culture is starved, fixed or lysed, then disposed of as biological waste.
Nociception (detecting harm) is not pain (experiencing harm). Current organoids lack the pathways and integration most scientists think pain needs, and no study has shown that they suffer.
Uncertainty still matters. Philosophers such as Jonathan Birch argue for precautions proportionate to the evidence, plus clear warning signs that would change the answer.
A trained culture’s learned state has real scientific and commercial value, but value to us is not the same as an interest of the culture.
Animal research, cell lines and brain death each offer useful analogies, and each breaks down somewhere.
The practical answer is a published retirement policy: defined endpoints, documented methods, preserved data and regular review.

What does “switching off” a biological computer actually mean?

With a biological computer, cutting power is the easy part. The microelectrode array that records and stimulates the neurons can be unplugged, cleaned and reused. The hard part is the tissue sitting on top of it.

A brain organoid is a cluster of living cells grown from induced pluripotent stem cells, as we explain in our guide to how brain organoids are grown. It stays alive only while it receives oxygen, glucose and other nutrients and while its waste is carried away. In many biocomputing setups that job is done by microfluidics: small channels and pumps that circulate culture medium around the tissue.

The three ways a culture ends

In practice, ending a culture usually takes one of three routes:

Withdrawal of support. Medium exchange and perfusion stop. Without fresh nutrients and with waste accumulating, the neurons stop firing and the cells die. This is the closest thing to “switching off,” and it is not instantaneous.
Fixation. The tissue is immersed in a chemical fixative that cross-links proteins. This kills cells quickly and preserves their structure, so researchers can image the tissue and learn what the experiment did to it.
Lysis for analysis. The cells are broken open to extract RNA, DNA or proteins, for example to see which genes changed during training.

After any of these, the remains are handled as biological waste under the lab’s biosafety rules. Because the cells are human-derived and carry a donor’s genome, the donor’s consent terms also govern what happens to any leftover material, a topic we cover in who owns a living computer.

How long do biological computers live?

There is no single answer, and the reliable numbers come from the groups running the systems. In 2024, Jordan and colleagues at FinalSpark described their remotely accessible Neuroplatform in Frontiers in Artificial Intelligence. They reported that organoids on their electrode setup initially lived only a few hours, and that they had extended this to up to about 100 days in the best cases. The paper says the platform had been used with over 1,000 organoids over three years, with organoids on the arrays replaced more than 250 times. The same authors note that, before going onto the electrodes, organoids can be kept alive in ordinary culture for years.

Cortical Labs, the company behind the DishBrain experiments, says on its product page that its CL1 device is designed to keep neurons alive for up to six months using an internal life-support system. That is a company claim, not a peer-reviewed measurement.

Off the electrodes, organoids can last much longer. Gordon and colleagues reported in Nature Neuroscience (2021) that human cortical organoids reached molecular signatures of postnatal stages after 250 to 300 days in culture.

The practical point is simple: every biocomputing culture will end. The question is not whether labs end cultures but how, when and with what care.

Can a brain organoid suffer?

This is the question everything else hangs on. If an organoid cannot have experiences, then ending it raises no welfare issue for the organoid itself (though it may still raise issues of respect, consent and waste). If it can, the picture changes. For the broader question of awareness, see our companion piece, can brain organoids become conscious?. Here we focus on pain and suffering, because those are what welfare protections are usually designed to prevent.

Nociception is not pain

Two words do a lot of work here.

Nociception is the detection of potentially harmful stimuli by specialized sensory neurons and the signaling that follows. Sneddon’s 2018 review in Physiology describes it as an alarm system for tissue damage and notes that all animals studied possess nociceptors. Nociception can drive reflexes without any feeling at all: a sleeping or anesthetized body can still withdraw a limb.

Pain is something else. The International Association for the Study of Pain revised its definition in 2020 (Raja et al., Pain) and still defines pain as an unpleasant sensory and emotional experience associated with, or resembling that associated with, actual or potential tissue damage. The key word is experience. The IASP notes that pain and nociception are different phenomena and that pain cannot be inferred from activity in sensory neurons alone.

What does pain seem to require?

Scientists disagree on the exact neural requirements, but the debate has a recognizable shape. The fetal pain literature is a useful guide because it asks the same question about developing human tissue.

Lee and colleagues’ 2005 systematic review in JAMA argued that pain perception requires conscious awareness of a noxious stimulus, that reflexes and stress hormones do not prove pain, and that awareness of noxious stimuli requires functional connections between the thalamus and the cortex. On that basis, they concluded fetal pain was unlikely before the third trimester.
Derbyshire and Bockmann (2020, Journal of Medical Ethics) challenged this, citing evidence that the cortex may not be necessary for pain and that thalamic connections into the developing subplate form earlier. They argued neuroscience cannot definitively rule out fetal pain before 24 weeks.
More broadly, Merker (2007, Behavioral and Brain Sciences) argued that brainstem and subcortical systems may support a basic form of consciousness even without a cerebral cortex.

Notice what both sides share. Whether the critical structure is the cortex, the thalamus or the brainstem, every serious account involves a pathway from the body’s sensors into large, integrated, organized brain systems that regulate arousal and evaluate what is happening to the organism.

Why current organoids very likely lack it

Measured against any of those accounts, today’s biocomputing cultures fall short in several ways at once:

No body and no nociceptors wired to tissue. There is nothing to damage that the culture monitors, and no sensory pathway reporting it.
No brainstem or thalamus in working form. Most biocomputing organoids are cortical-like tissue. Birch (2024, Molecular Psychology) points out that it is a point of consensus that a brain-dead human is not sentient, and that current organoids lack a functioning brainstem.
Small and unvascularized. Without blood vessels, organoids are limited in size and their cores can become starved of oxygen. Efforts at organoid vascularization aim to change this, which is one reason the question needs revisiting over time.
Simple activity. Organoids produce spontaneous spikes and network bursts, but nothing resembling the sustained, differentiated activity associated with awake experience.

The National Academies of Sciences, Engineering, and Medicine reached a similar view in 2021, judging that awareness, emotion or pain in current organoids is extremely unlikely in the foreseeable future, while recommending that the question be revisited as the science advances.

So the honest summary is: no current evidence of suffering, good structural reasons to doubt it, and no way to prove a negative.

How should we treat organoid welfare under uncertainty?

If we cannot be certain, how cautious should we be? This is where the debate becomes genuinely philosophical.

The precautionary approach

Jonathan Birch has developed the most influential framework for this kind of problem. In a 2017 paper in Animal Sentience, he argued that waiting for proof of sentience before offering protection sets the bar too high, and that credible evidence should be enough to trigger proportionate precautions. His 2024 book The Edge of Sentience extends the approach to cases at the edge of our knowledge, including neural organoids, and asks that precautions be proportionate to the risk and debated in inclusive, democratic ways.

Applied to organoids, Birch proposes a concrete line he calls the “brainstem rule”: if a neural organoid develops or connects to a functioning brainstem that regulates arousal and produces sleep-wake cycles, it should be treated as a sentience candidate. For such candidates, he suggests either a moratorium on creating that type of organoid or, more in line with animal research, mandatory ethical review and harm-benefit analysis.

The strength of this approach is that it is testable. It does not ask labs to fret over every culture. It names warning signs that would change the answer.

The skeptical response

Others argue that precaution can be overdone. If there is no plausible mechanism for suffering, treating every culture as a potential patient wastes ethical attention that could go to animals and people who certainly can suffer. It may also mislead the public into believing organoids are tiny brains with feelings. Balci and colleagues made a related point in Neuron (2023), criticizing the use of words like “sentience” for in vitro cultures because they invite misunderstanding.

Where most ethicists land

Most frameworks converge on proportionate monitoring. Lavazza and Massimini (2018, Journal of Medical Ethics) proposed adapting clinical consciousness-assessment tools to organoids, and Sawai and colleagues (2021, AJOB Neuroscience) mapped the wider issues. The Baltimore declaration (Hartung et al., 2023) and the roadmap by Smirnova and colleagues (2023) both commit the field to take these questions seriously. For the deeper question of where moral status might begin, see do biological computers have moral status?.

What can animal research, cell lines and brain death teach us?

No existing category fits organoid biocomputers neatly. Comparing them with neighbors shows which rules transfer and which do not.

Case	Is it alive?	Can it suffer?	How does its life end?	What governs it?	Lesson for biocomputing
Immortalized cell lines (for example, HeLa)	Yes, as cells	No, by broad consensus	Discarded or frozen routinely	Biosafety, consent and data rules	Ending cells is not, by itself, a welfare event
Laboratory animals	Yes, as organisms	Yes, for vertebrates and some invertebrates	Humane killing at planned endpoints	Animal welfare law, the 3Rs, ethics review	Plan endpoints in advance and minimize distress
Human after brain death	Some tissues, with support	No: permanent loss of capacity for consciousness	Withdrawal of support or organ donation	Medical criteria and law	Living cells do not imply a living person
BrainEx pig brains (2019)	Some cellular functions restored	No organized activity linked to awareness was detected	Experiment ended under monitoring	Institutional review, NIH oversight	Prepare monitoring and stop rules before you start
Organoid biocomputers today	Yes, as tissue	Very unlikely, on current evidence	Starvation, fixation or lysis	Biosafety, consent, emerging guidance	Borrow structure from all of the above

Cell lines: the low baseline

Labs discard and freeze human cell lines every day, and almost no one thinks this harms the cells. Organoids start from the same raw material. The difference is organization: neurons assembled into networks that fire, adapt and respond to input.

Animal research: humane endpoints and the 3Rs

Animal research offers the most developed end-of-life ethics in science. In The Principles of Humane Experimental Technique (1959), William Russell and Rex Burch introduced the 3Rs, which the UK’s NC3Rs now defines as replacement (approaches that replace animal use), reduction (fewer animals per study) and refinement (methods that minimize pain, suffering, distress or lasting harm). European law, in Directive 2010/63/EU, builds on these principles and requires humane methods of killing and early, humane endpoints.

Two features stand out. First, endpoints are planned before the experiment, not improvised at the end. Second, the law draws scope lines by developmental stage: the Directive covers, for example, fetal forms of mammals only from the last third of their normal development. That is an explicit admission that moral concern can switch on as a nervous system matures, which is exactly the shape of the organoid problem.

Organoid platforms are often promoted as replacements for animal testing; the Cortical Labs CL1 page calls the device an ethically superior alternative. If that is the pitch, it is fair to ask the industry to apply reduction and refinement to its own cultures too.

Brain death and BrainEx: what “irreversible” means

Under the World Brain Death Project consensus (Greer et al., JAMA, 2020), death by neurologic criteria is determined through coma, absent brainstem reflexes and inability to breathe, reflecting complete and permanent loss of brain function. A brain-dead person can still have living cells.

In 2019, Vrselja and colleagues reported in Nature that their BrainEx perfusion system restored microcirculation and some molecular and cellular functions to intact pig brains up to four hours after death, including spontaneous synaptic activity, in the absence of global electrocorticographic activity. Yale’s announcement quoted Vrselja saying the team never observed the organized electrical activity associated with perception, awareness or consciousness, and that the team was prepared to use anesthetics and cooling to stop such activity if it appeared.

Commentaries in Nature drew out two implications. Farahany, Greely and Giattino argued the work raised new ethical quandaries about isolated brain tissue and called for guidelines. Youngner and Hyun noted it challenged assumptions about how irreversible brain damage really is, with knock-on effects for organ donation.

For organoids, BrainEx teaches two lessons. Cellular life and the capacity for experience can come apart, which supports the view that active neurons in a dish are not automatically a subject. And the responsible way to handle uncertainty is to decide in advance what you will monitor and what you will do if you see it.

Does a trained culture’s learning matter morally?

Biocomputing adds something no previous case had: tissue that has been trained. A culture shaped by closed-loop stimulation carries a learned state in its synaptic plasticity. Does ending it destroy something morally significant, or merely something valuable?

The case that it matters only commercially

On most mainstream views, moral status depends on a capacity for experience or interests, not on information content. A trained network is a pattern of connection strengths. A trained artificial neural network is also a pattern of weights, and deleting one wrongs no one. On this view, a trained culture is valuable in the way a dataset, a custom instrument or a well-characterized cell line is valuable. Losing it is a cost to the lab, to funders and perhaps to science, but not a harm to the culture.

The case that it might matter more

The counterargument is that learning is one of the very capacities people cite as evidence of a mind. If a system can adapt its behavior to feedback, some argue it has at least a minimal stake in how things go. Kagan and colleagues, writing in AJOB Neuroscience in 2022, themselves raised the question of whether embodied, learning neurons supply evidence relevant to organoid ethics. A cautious position is that learning alone is not enough, but that learning combined with other markers (arousal regulation, integration, sleep-wake cycles) should raise the level of scrutiny.

What the evidence says about learned state today

It helps to be precise about how durable this learning is. In the DishBrain study, Kagan and colleagues reported in Neuron (2022) that within-session learning was well established, but between-session learning over multiple days was not robustly observed, and cultures appeared to relearn associations with each new session. Our explainer on how neurons learn in a dish covers the mechanisms.

So the “investment” in a trained culture is, for now, often short-lived even while the culture is alive. That cuts against the idea that retirement destroys an accumulated self. It also means the most durable record of training is usually the data, not the tissue.

Should a trained network be preserved?

If a trained state has value, can we keep it after the culture ends?

Preserving the data

The realistic form of preservation today is the record: raw recordings, stimulation protocols and outcomes, stored in an open format such as Neurodata Without Borders with the culture’s lineage and consent metadata. This is good science, and it answers the intuition that ending a culture should not erase its contribution.

Preserving the tissue

Preserving the living state is much harder. Fixation preserves structure, which can reveal which connections changed, but not the living, functioning network. Stem cell lines are routinely frozen, but freezing and reviving a mature, trained organoid with its learned function intact is not an established technique, and no one should assume a revived culture would “remember.”

Is preservation an obligation?

Here the arguments divide again. If the culture has no interests, preservation is a matter of scientific stewardship: keep data that others can use, and do not waste donated material. If future cultures turn out to be sentience candidates, the question becomes more complex. Indefinite maintenance is not obviously kind, since ageing cultures degrade and electrode responses deteriorate. Jordan and colleagues observed that the current needed to elicit spikes rose over an organoid’s lifetime on their platform, which they suggested may relate to glial encapsulation of the electrodes. A precautionary lab should therefore not treat “keep it alive as long as possible” as the default ethical choice.

Is negative feedback during training a welfare issue?

End of life is not the only welfare moment. Training itself involves stimulation, and some of it is designed to be avoided.

What DishBrain actually did

DishBrain used two-dimensional cultures of mouse cortical cells and human stem-cell-derived neurons on a high-density array, not organoids. According to the paper, when the simulated paddle missed the ball, the culture received an unpredictable stimulus: 150 mV pulses at 5 Hz delivered at random sites and times across 8 input electrodes for four seconds. When the paddle hit the ball, it received a predictable stimulus across all electrodes at 100 Hz for 100 ms. The design follows the free energy principle: the hypothesis is that neural networks act to reduce unpredictable input, so they learn behavior that avoids it.

The authors used the word “sentience” in a narrow, technical sense drawn from Friston’s work, meaning responsive to sensory impressions. They did not claim the cultures felt anything, and critics such as Balci and colleagues argued that the word invites exactly that misreading.

Is unpredictable stimulation aversive?

Calmly stated: the feedback works by information, not injury. It is a brief, structured-versus-random signal, not a stimulus designed to damage tissue, and there is no evidence the cultures experienced it as unpleasant. In the same study, a “silent” condition that simply withheld stimulation after misses also supported some learning, though less than unpredictable feedback, which shows the mechanism is about predictability rather than harm.

Still, the welfare question is fair to ask about the future. If a system ever qualified as a sentience candidate, a training signal whose function is to be avoided would deserve the same scrutiny that refinement demands in animal work. The sensible step now is cheap: document stimulation parameters, justify their intensity, and prefer the least intense feedback that achieves the scientific goal.

What arguments do different sides make?

Fair treatment means stating each position at its strongest.

“Switching off is clearly fine.” Current organoids lack the structures every leading theory links to pain. Ending them is morally like discarding a cell culture, and inventing welfare duties for tissue without welfare distracts from real suffering in animals and people.
“It is fine, but how we do it matters.” Even without interests, organoids come from human donors and carry their genomes. Respect for donors, careful use of donated material and honest records are obligations in their own right.
“Act as if it might matter.” We cannot prove the absence of experience, and organoids are getting larger, longer-lived and better connected. Proportionate precautions now, with clear triggers for stronger ones, follow the logic of animal welfare rules that extend protection by developmental stage.
“Some organoids should not be made.” If research heads toward sentience candidates, the question is not how to retire them but whether to create them. Birch’s paper notes that a moratorium may be appropriate for organoids that meet his brainstem rule.

What should a responsible retirement policy include?

A lab does not need to settle the philosophy to act well. It needs a written, published policy. Here is a checklist a responsible biocomputing lab could adopt:

Defined endpoints. State which criteria end a culture: experiment complete, declining signal quality, contamination, or a set maximum age.
Chosen method, with reasons. Specify whether cultures are fixed, lysed or withdrawn from support, and why that method fits the scientific purpose.
Final recording. Capture a last recording and quality check before ending, so the culture’s final state is documented.
Data preservation. Archive raw data, protocols and outcomes in open formats with full provenance.
Consent compliance. Check the donor consent scope before using, sharing or disposing of remaining material.
Stimulation log. Record all stimulation parameters used in training, including any negative feedback, with justification.
Complexity monitoring. Track markers that would raise concern, such as sustained organized activity, sleep-wake-like cycles, connection to other tissue types (for example in an assembloid) or vascularization.
Escalation rule. Name in advance what happens if a culture crosses a warning threshold: pause, independent ethics review, and harm-benefit analysis.
Reduction and refinement. Use the number of cultures a study needs and the least intense stimulation that works.
Regular review and publication. Revisit the policy as evidence changes, and publish it so outsiders can scrutinize it.

None of these steps is expensive; most are good laboratory practice written down.

What questions remain open?

Where exactly is the line? Birch’s brainstem rule is a proposal, not a consensus. Other researchers might weight cortical integration or other markers more heavily.
Can we measure experience in a dish? Clinical consciousness tests were designed for bodies and brains. Adapting them to organoids is an active research question.
Does learning add moral weight? There is no agreed answer on whether adaptive behavior alone, without other markers, should change how a culture is treated.
What about hybrid systems? As organoids are connected to each other, to sensors or to other tissues, the analogies with isolated cultures weaken. Our overview of the future of biocomputing sketches where the field is heading.
Who decides? No body yet has clear authority over organoid end-of-life practice specifically.

So, is it wrong to switch off a biological computer?

On the best current evidence, switching off a biological computer is not wrong in itself, because today’s organoid intelligence systems almost certainly lack the capacity to suffer. What can be wrong is doing it carelessly: without planned endpoints, without preserved data, without respect for donors, and without watching for the signs that would change the answer.

For more on the wider landscape, start with the ethics of organoid intelligence and our introduction to what organoid intelligence is. At Organoid Programming, we think retirement policies, stimulation logs and provenance belong in the same versioned record as the experiment itself, and we set out our approach in our ethics commitments.

Frequently asked questions

What happens when a biological computer is switched off?

Switching off a biological computer means ending the life of its neural culture, not cutting power to a chip. Labs either stop supplying fresh medium and oxygen so the cells gradually die, fix the tissue chemically (which kills it quickly and preserves its structure for study), or lyse it for molecular analysis. The remains are then disposed of as biological waste under the lab’s biosafety rules. The electronics can be cleaned and reused.

Can a brain organoid feel pain?

There is no evidence that any current brain organoid feels pain. Pain, as the International Association for the Study of Pain defines it, is an unpleasant sensory and emotional experience, which is different from nociception, the detection of harmful stimuli. Organoids lack pain receptors connected to a body, a brainstem, and the large integrated circuits most researchers associate with experience. A 2021 National Academies report judged pain in organoids extremely unlikely for the foreseeable future.

How long do biological computers live?

Lifespans vary by system and are reported by the groups that run them. FinalSpark’s Neuroplatform team reported in 2024 that organoids on their electrode setup lived up to about 100 days in the best cases, after early versions lasted only hours. Cortical Labs says its CL1 device is designed to keep neurons alive for up to six months. Organoids kept in ordinary culture, off electrodes, can be maintained for much longer.

Is it more ethical to keep a trained organoid alive than to end it?

Not necessarily. Keeping a culture alive only matters for its own sake if it has welfare interests, and current evidence suggests it does not. If it did, prolonging its life could be good or bad depending on its condition, since ageing cultures degrade. A trained culture’s learned state clearly has scientific and commercial value, but value to researchers is not the same as an interest of the culture itself.

Is the negative feedback used to train neurons in a dish cruel?

In the DishBrain study, cultures received brief unpredictable electrical stimulation after missing the ball and predictable stimulation after hitting it. The feedback works through information, not damage, and there is no evidence the cultures experienced it as unpleasant. Still, some ethicists argue that if future systems become sentience candidates, feedback signals designed to be avoided deserve review, so labs should document stimulation parameters and their rationale.

What does the BrainEx pig brain study have to do with organoids?

In 2019 Vrselja and colleagues restored circulation and some cellular functions to pig brains up to four hours after death, without detecting the organized brain activity linked to awareness. It showed that cellular activity and capacity for experience can come apart, and that ‘irreversible’ loss of brain function depends partly on technology. Both lessons apply to organoids: living, active neurons are not the same as a mind, and monitoring plans matter.

What should a responsible organoid retirement policy include?

A responsible policy states which criteria trigger retirement, such as declining signal quality or a finished experiment, and which method is used. It records final recordings and data in open formats, honors the donor consent scope for any remaining material, monitors cultures for warning signs of greater complexity, and documents stimulation used during training. It should be reviewed regularly and published so others can scrutinize it.

Should biocomputing research follow the 3Rs like animal research?

The 3Rs (replacement, reduction and refinement), first set out by Russell and Burch in 1959, were written for animal experiments, and organoids are often promoted as a replacement for animals. Applying the spirit of reduction and refinement to organoids is cheap and sensible: use as many cultures as a study needs, avoid needless intense stimulation, and plan humane endpoints in advance, in case future evidence shows the cultures can be harmed.

Is it wrong to switch off a biological computer? Welfare, end of life and responsible retirement