Organoid Intelligence

Established The prize is a computing substrate that grows and repairs its own wiring and runs on a power budget measured in watts rather than megawatts — if the efficiency survives a fair test, which no one has yet run. The human brain performs its work on about 20 watts; the Frontier supercomputer it is routinely compared against drew about 21 megawatts in June 2022, a gap the organoid-intelligence field cites as a roughly 10⁶-fold advantage in power efficiency. The entire case for growing neurons to compute rests on capturing some fraction of that gap in a device a silicon competitor cannot match at equal task performance.

Frontier Stated in the field’s own units, the banked results are small and the extrapolation is enormous. A dish of about 800,000 cortical neurons changed its behaviour in a Pong-like task within about five minutes of closed-loop stimulation; a three-dimensional brain organoid used as a physical reservoir has been reported classifying speaker identity at roughly 78% while cutting training time by more than 90% against a conventional network carrying a long short-term memory layer. Each organoid today holds fewer than 100,000 cells — about one three-millionth of a human brain — and no system has reported joules per inference at all.

Speculative Who gets the upside depends on which of two futures the substrate turns out to be in. If the efficiency claim holds at scale, the beneficiary is anyone running inference against the energy wall that now limits data centres, and the spillover — cheap, reliable culture of human cortical tissue — would accrue to disease modelling and drug screening whether or not a single wetware computer is ever sold. If it does not hold, the gains stay inside neuroscience, where cultured cortex is already a useful instrument for studying learning.

Handwave The condition on all of it is a number nobody has measured against a fair opponent. The bottleneck list in section 5 turns on a single missing experiment named in section 7: a matched-parameter, energy-reported contest against a silicon reservoir. Until that runs, the upside is bounded by a quantity the field has never reported, and section 15 shows how often the missing number is quietly replaced by the brain’s.

Established “Organoid intelligence” is an umbrella over at least three distinct experimental systems that share a slogan and almost none of their methods. The first is a two-dimensional monolayer of cortical neurons grown on a multielectrode array and driven in a closed sensory-motor loop — the DishBrain configuration. The second is a three-dimensional brain organoid used as the fixed nonlinear medium of a reservoir computer — the Brainoware configuration. The third is a research programme, named organoid intelligence in a 2023 roadmap, that proposes to scale organoids into a general biocomputing platform and treats the first two as proofs of principle. The systems make different claims, on different substrates, and a reader who lets the umbrella collapse them will over-read every one.

Established The substrate distinction is not pedantry; it decides what each result can mean. A monolayer culture is a thin sheet where every neuron sits near oxygen and near an electrode, so a dish can be embodied and taught. An organoid is a millimetre-scale ball where the interior is starved of oxygen and unreachable by a surface electrode array, so it is used as a reservoir — a rich, fixed dynamical system that a trained silicon read-out layer interprets — rather than as something that learns a task. Calling both “organoid intelligence” imports the learning story from the dish into the organoid, where it has not been shown.

Frontier The interpretive lens throughout is the free energy principle, which is both a strength and a liability. The dish work reads its data as active inference — a system acting to minimise surprise about its sensory input — and that lens is what licenses the words “learning” and “sentience”; whether the framework describes the tissue or merely redescribes any adaptive system is contested, and the field’s conclusions inherit that dispute.

Frontier The move from sheet to ball buys developmental realism and pays in access. Organoids self-organise a layered, three-dimensional architecture a monolayer cannot, which is why they are the better disease model; the same three dimensions make them a worse computer to interface, which is the tension the field’s name papers over.

Frontier Reservoir computing is the concept that makes the organoid results legible. In a reservoir computer, an untrained nonlinear system with fading memory maps inputs into a high-dimensional state; only a simple linear read-out is trained. The organoid supplies the reservoir; the reported accuracy is the read-out’s, borrowing the organoid’s dynamics. This is a genuine and useful architecture, and it is also the reason a good organoid score does not by itself show the organoid learned anything — the training happened in silicon.

Established The founding motivation is thermodynamic, not algorithmic. Nobody proposes cultured neurons because they are faster than a GPU; they are proposed because biological neural tissue reconfigures its own connectivity through synaptic plasticity for a metabolic cost that silicon cannot presently approach. The overview to hold in mind for the rest of this brief is therefore narrow: the field is a bet that a substrate which wires itself cheaply can be made to compute usefully, and every section below is a way of asking whether the bet has paid any measurable dividend yet. The sober treatment of the neighbouring wetware programmes is in Biological Computing, which this brief does not duplicate.

Established The anchor result of the whole field is the DishBrain study, and its measured core is real even where its language is contested. Kagan and colleagues plated about 800,000 cortical cells — both human induced-pluripotent-stem-cell-derived neurons and primary mouse cortical neurons — onto a high-density multielectrode array, embedded them in a simulated Pong world through stimulation and feedback, and reported that learning was apparent within five minutes of real-time gameplay and not present in control conditions. Sessions ran 20 minutes, with the first five minutes compared against the last fifteen.

Established The statistics on the primary comparison are strong and worth stating precisely. Against rest conditions, only the experimental cultures improved on the summary metrics — average rally length, aces and long-rallies of more than three consecutive hits — with human cortical cells at t = 10.44, p = 3.91 × 10⁻¹⁹ and mouse cortical cells at t = 6.15, p = 5.27 × 10⁻⁰⁸. Human cells outperformed mouse cells throughout. The analysis pooled 399 test-sessions in the main experiment (80 control, 42 rest, 38 in-silico, 101 mouse, 138 human) and 483 in the feedback experiment.

Frontier The word the study is remembered for is the one its authors defined most narrowly. The paper claimed “sentience” only in a technical sense — responsiveness to sensory impressions through active inference under the free energy principle — not phenomenal experience. That framing drew a formal multi-author reply in the same journal, Balci and colleagues’ response to claims of emergent intelligence and sentience in a dish, which argued the interpretation outran the evidence and that homeostatic, stimulus-avoidance dynamics can produce the same behaviour without anything worth calling learning-with-sentience.

Frontier The most important recent result on the dish is a partial failure to reproduce it, from the original group. In a 2025 Communications Biology study using the DishBrain setup with roughly 200,000 human iPSC-derived neurons plus 50,000 astrocytes on a 1,024-electrode array, the authors reported that no statistically significant learning effects were observed over time in untreated cultures, with gameplay improvements appearing only after the anti-seizure drug carbamazepine at 200 micromolar, and with smaller effect sizes than the original work. Phenytoin and perampanel produced no significant change.

Established DishBrain included an in-silico control, which is why the effect was taken seriously. Among the 399 test-sessions were 38 in-silico runs alongside 80 no-feedback controls and 42 rest conditions; that comparison structure is the load-bearing part of the design, and it is exactly the structure an independent replication has to reproduce blind.

Established The two flagship results are not one escalating experiment on the same tissue. DishBrain is a two-dimensional culture taught in a loop; Brainoware is a three-dimensional organoid read as a reservoir; the 2025 follow-up added about 50,000 human cortical astrocytes to its neurons, a co-culture choice absent from the 2022 original. Reading the three as a single rising line is the error section 15 takes apart.

Established The organoid branch rests on Brainoware, and its cleanest numbers are the ones it reports in full. Cai and colleagues used a mature human brain organoid on a multielectrode array as a reservoir and, on a Henon-map prediction task, improved a regression score from 0.356 ± 0.071 before training to 0.812 ± 0.043 after four training epochs of one epoch per day; a blank array with no organoid scored 0, and blocking synaptic plasticity with the inhibitor K252a held improvement to 0.31 ± 0.072 up to only 0.385 ± 0.063 — evidence that the organoid’s own plasticity, not just the silicon read-out, was doing work.

Frontier The headline benchmark for Brainoware is a speaker-identification figure this brief could not open at the primary source. The published Nature Electronics paper reports classifying speaker identity at roughly 78% on a spoken-vowel dataset and cutting training time by more than 90% against a conventional network with a long short-term memory layer; the accuracy sentence sits behind a paywall the brief could not read, so the figure is reported here as the paper’s claim rather than as verified against its primary text, and section 15 returns to what that inaccessibility means for a field selling reproducibility.

Established The infrastructure is real and is being sold and rented, which is a genuine change since 2022. FinalSpark’s Neuroplatform reported running more than 1,000 brain organoids over three years with a lifetime of more than 100 days in the best cases, up from a few hours at the start; each array holds 4 organoids at 8 electrodes each, access is entirely remote through a Python interface, and in 2023 the platform selected 8 of 36 proposed academic projects. Cortical Labs’ CL1 is described as a scalable device for closed-loop, real-time interaction with in-vitro cultures over an integrated life-support perfusion circuit — the first such system marketed as a product rather than a laboratory rig.

Frontier What the position lacks is the comparison the field’s own thesis demands. No published system reports joules per inference for a cultured-neuron computation, and no benchmark pits a neural culture against a silicon reservoir matched for parameter count on the same task. Recent reviews of the area say plainly that matched silicon baselines and energy-per-inference comparisons are absent. The measured record, then, is a strong closed-loop learning claim now under a reproducibility cloud, a clean reservoir demonstration, and a real service industry — standing on top of an efficiency argument no one has tested head-to-head.

Frontier Does a cultured network learn, or does it merely settle? The central open question is whether the behavioural change in a closed loop is learning — the acquisition of a task-specific policy — or a return to a low-variance dynamical regime under structured stimulation that would occur for any patterned input. The distinction is testable and has not been cleanly settled; the 2025 non-replication in untreated cultures makes it the live question of the field rather than a philosophical aside.

Frontier How much of a reservoir result belongs to the tissue? Because the trained read-out lives in silicon, a reservoir demonstration must show the organoid contributes dynamics a trivial reservoir would not. The K252a plasticity-blocking control in Brainoware is exactly this kind of evidence and is the strongest single argument that the tissue matters; whether it generalises beyond one task and one organoid is open.

Speculative Is the efficiency advantage real once the whole system is metered? The 20-watt brain is not the thing on the bench. A wetware rig carries incubation, perfusion, temperature control and CO₂ regulation; the honest denominator is watts for the whole life-support envelope divided by useful inferences, and no one has published it. It is entirely possible the substrate is efficient and the system around it is not.

Frontier Can a three-dimensional organoid ever be more than a surface? A ball of tissue is read and written only at its shell; its computational interior is dark. Whether high-density three-dimensional interfaces, or organoids engineered with internal vasculature and recording sites, can turn the volume into usable substrate is the question that decides whether “organoid” computing is more than skin-deep.

Speculative What is the right task class for a self-wiring substrate? If the substrate’s distinctive property is cheap self-reconfiguration, the tasks that flatter it are adaptive, online and non-stationary ones, not static benchmarks a GPU already owns. Nobody has defined the benchmark suite on which a neural culture would be expected to win, which is why every current comparison is against a task silicon was built for.

Speculative Does the human-cell advantage mean anything? DishBrain reported human cortical cells outperforming mouse cells on the task; whether that reflects a genuine computational difference or a difference in maturation, culture properties or electrode coupling is unresolved, and it matters because the entire premise of using human tissue — with all its ethical cost — rests on the answer.

Handwave Where, if anywhere, does moral status begin? The frontier question the field cannot answer and cannot avoid is whether scaling the substrate toward the complexity that would make it a better computer also moves it toward the complexity that would make it a subject. This is treated head-on in section 11; it is a frontier question because the science that would locate the threshold does not exist.

Established Oxygen diffusion is the hard physical ceiling on organoid size. Without an internal blood supply, nutrients and oxygen reach only a shell a few hundred micrometres thick; the interior of a larger organoid necrotises. This is why organoids stay below about 500 micrometres and fewer than 100,000 cells, roughly one three-millionth of a human brain, and why the roadmap’s target of about 10 million neural cells per organoid is gated on solving perfusion, not on culture time.

Established Lifespan bounds every economic argument. The best-case organoid lifetime on a running platform is reported at more than 100 days; a device whose compute element must be regrown every few months, from a differentiation process that takes weeks and yields variable tissue, has an availability and amortisation problem no silicon chip has. FinalSpark’s own history — from a few hours to 100 days — shows the trajectory is real and shows how recently it was measured in hours.

Frontier Reproducibility is the bottleneck that turns results into anecdotes. Organoids are highly heterogeneous batch to batch; spontaneous activity patterns vary within and between cultures and depend heavily on protocol. A field whose flagship learning result was not statistically reproduced in untreated cultures by its own originators, three years on, has a variance problem before it has a scaling problem.

Established The electrode interface is a two-sided constraint. Surface multielectrode arrays reach only the outer neurons, leaving the organoid interior both unreadable and unwritable; and the electrodes themselves can alter excitability. A recent review notes that input-driven plastic changes “can enhance performance even in nominally open-loop configurations” and that drift correction is needed to compensate for neurons moving relative to the electrodes — both of which can masquerade as learning if uncontrolled.

Frontier No one has instrumented the energy claim, so the field’s central selling point is unmeasured. The comparison that would justify the whole enterprise — joules per inference for the cultured system against a matched silicon reservoir — is absent from the literature. Neuromorphic silicon already reports figures like tens of millijoules per synaptic operation and picojoule-scale spike-timing updates; a wetware system that cannot state its own number cannot claim the advantage it was built to demonstrate.

Established The compute element takes weeks to grow and is never twice the same. Differentiation from stem cell to usable cortical tissue runs on the order of weeks to months and yields variable tissue, so every unit of a wetware machine is a slightly different device — a manufacturing problem with no counterpart in a fab that stamps out identical chips.

Frontier Standardisation does not exist. There is no agreed task suite, no agreed metric, no agreed reporting format for a wetware computation, and no shared control battery separating learning from drift and stimulus-response. Until there is, results from different laboratories are not comparable, and the service platforms accelerate output without constraining its form.

Established Thermal and fluidic constraints cap experiment length. The 2025 study notes the array’s heat drives evaporation and so restricts testing periods — a mundane engineering limit that nonetheless bounds how long any closed-loop experiment can run before the medium changes underneath it.

Established Everything downstream depends on reliable human neural differentiation. The substrate is only as good as the induced-pluripotent-stem-cell protocols that turn a donor cell line into cortical neurons of consistent identity and maturity. Batch-to-batch variability in this step is the upstream cause of the reproducibility bottleneck above; a stable, characterised, renewable cell source is the dependency all others sit on.

Established Organoid protocols must add what the field is named for and does not yet have: vasculature. Progress past the oxygen-diffusion ceiling depends on vascularised, perfused organoids, which is a problem being solved for reasons that have nothing to do with computing — disease modelling and transplantation — and which this field would inherit. The dependency is treated in Lab-Grown Organs.

Frontier The read-write interface depends on recording hardware and decoding mathematics borrowed wholesale from neurotechnology. High-density arrays, spike sorting, drift correction and closed-loop decoding are the same tools developed for Brain-Computer Interfaces and Neural Interfaces; the wetware field consumes them and contributes little back, which means its interface improves only as fast as those adjacent fields advance.

Frontier The interpretive framework depends on a theory of what the tissue is doing. Claims about learning and sentience rest on the free energy principle and active inference; whether those frameworks are the right lens, or a post-hoc vocabulary, is itself contested, and the field’s conclusions are only as secure as the theory it reads its data through. The related question of whether simple living systems compute at all is owned by Basal Cognition.

Frontier Support cells may be part of the substrate, not scaffolding. The 2025 setup co-cultured about 50,000 cortical astrocytes with its neurons; glia shape network excitability and metabolism, so a mature wetware system depends on getting the non-neuronal cell mix right — a variable the clean reservoir framing tends to ignore.

Established The economic case depends on an efficiency result from an adjacent discipline. Whether wetware is worth pursuing at all is decided by the energy comparison that Ultra-Efficient Computing Energy Systems and Thermodynamics of Computation own; if silicon closes the efficiency gap first, the substrate’s single distinctive advantage evaporates before it is measured.

Frontier The decisive experiment is a pre-registered, head-to-head contest between a cultured-neuron system and a silicon reservoir matched for parameter count, run on the same task with joules per inference reported for both, and paired with an independent, blinded replication that controls for electrode-induced excitability drift. This single experiment settles the two questions the field lives or dies on — whether the substrate does anything a matched silicon system does not, and whether it does it more efficiently — and it has not been run.

Frontier The matched-baseline half tests the benchmark question. A cultured system and a silicon reservoir with the same number of trainable read-out parameters, given the same inputs and the same training budget, must be scored on the same metric. If the culture does not beat the silicon reservoir at equal parameter count, its accuracy figures describe the read-out, not the tissue. No current result includes this control.

Frontier The energy half tests the reason the field exists. Both systems report joules per inference across the whole envelope — for the culture, incubation, perfusion and thermal control included. A win on accuracy at a loss on total energy would refute the founding thesis; the field’s central claim cannot be evaluated until this number is on the page for both sides.

Frontier The replication half tests the sentience-adjacent learning claim. An independent, blinded, pre-registered replication of the closed-loop learning result — with control conditions that separate learning from a return to low-variance dynamics, from electrode-induced excitability changes, and from spontaneous drift — is the experiment the field most needs and least has. Its urgency was raised, not lowered, by the 2025 finding that untreated cultures showed no significant learning over time.

Established One decisive control has already been run once and should be standard. Blocking synaptic plasticity pharmacologically, as Brainoware did with K252a, is the cleanest way to show the tissue rather than the read-out is learning; that it held improvement to a fraction of the intact case is the strongest existing evidence for the substrate, and its generalisation across tasks and laboratories is the near-term experiment that matters most.

Speculative A scaling experiment would test the roadmap’s premise. Whether a vascularised organoid of about 10 million cells, read in its interior rather than at its shell, computes measurably better than a 100,000-cell shell-read organoid is the experiment that would show volume is usable substrate. The hardware to write and read an organoid interior does not exist, so this test is currently unscheduled.

Frontier A standardised task-and-control battery is the meta-experiment. Before any single result is decisive, the field needs an agreed suite — a task class chosen to favour a self-reconfiguring substrate, a fixed metric, and a mandatory control set — run identically across platforms. Without it, the service platforms will keep producing incomparable one-off results faster than anyone can adjudicate them.

Handwave No experiment now on any bench can detect an inner life. The one test the ethical stakes demand — a marker that would distinguish a computing substrate from a subject — has no agreed operational form, which is precisely why the precautionary arguments in section 11 carry the weight they do.

Established The engineering problem is a life-support computer with a biological co-processor, and most of the parts count is life support. A working rig couples a multielectrode array to an integrated perfusion circuit, incubation, temperature and gas control; Cortical Labs’ CL1 packages exactly this as a product, which is the real engineering advance since 2022 — not a faster neuron but a stable enclosure for a slow one. The design point owned by Biological Computing — that the only wetware computer anyone has sold spends most of its power keeping the culture alive — is the design point here too.

Frontier Electrode density and depth are the binding hardware constraints. Arrays have gone from tens of channels to the 1,024-electrode class used in the 2025 DishBrain work, but they remain planar or surface-bound. The engineering that would matter — three-dimensional, high-density interfaces that reach an organoid’s interior — is unsolved, and until it is, added organoid volume is dead weight.

Frontier Drift and thermal management are first-order, not housekeeping. Neurons move relative to electrodes and require continuous drift correction; array heat evaporates the medium and caps session length. A production wetware system has to hold a living, moving, warming target in a stable measurement regime for months — a control problem with no silicon analogue.

Speculative Scale-out, not scale-up, is the plausible near-term architecture. Because a single organoid is size-capped by diffusion, the engineered path to more capacity is many small units in parallel — four organoids per array today — rather than one large one, which trades the interconnect problem silicon knows how to solve for the culture-uniformity problem biology does not.

Established Throughput today is a service, not a supercomputer. One platform ran more than 1,000 organoids over three years at four organoids per array and eight electrodes each; that is enough to rent experiments to dozens of groups and orders of magnitude short of the channel counts and interconnect a serious parallel machine would need.

Speculative The cost structure is inverted relative to silicon. A chip is expensive to design and cheap to copy; a wetware unit is cheap to specify and expensive to keep alive, so the economics reward running a few long-lived units continuously rather than many identical ones briefly — the opposite of the workload silicon data centres are built for.

Frontier The whole-system energy budget is the specification nobody writes. An honest datasheet would state watts for incubation, perfusion, gas and thermal control alongside useful inferences per second; that this line does not appear on any current system is the engineering expression of the missing experiment in section 7.

Established The nearest neighbour is the sober wetware survey this brief defers to. Biological Computing owns the three-programmes-one-name problem across DNA storage, molecular logic and cultured neurons, and owns the energy comparison the field keeps making badly; this brief narrows to the cultured-neuron programme and the sentience-and-benchmark pair it uniquely raises.

Frontier The moral-status question is shared with the consciousness briefs and settled by none of them. Consciousness Research, Machine Consciousness and Synthetic Consciousness own the theories — integrated information, global workspace, higher-order — that any welfare judgement about cultured cortex would have to invoke; the dish is the place those theories acquire a laboratory object and a legal deadline.

Established The interface is borrowed from neurotechnology. Brain-Computer Interfaces and Neural Interfaces supply the recording hardware and decoding mathematics every demonstration here uses, and their progress rate-limits this field’s.

Frontier The comparison classes it has chosen for itself. The efficiency argument belongs to Ultra-Efficient Computing Energy Systems and Thermodynamics of Computation; the intelligence comparison the roadmap invites belongs to Artificial General Intelligence, against which the wetware programme has so far published no benchmark; and the question of whether living matter computes at its simplest is owned by Basal Cognition.

Established The field organised itself before it had a reproducible result. The 2023 organoid-intelligence roadmap was accompanied by a community-forming workshop and an explicit “embedded ethics” proposal in which ethicists join research teams from the start. Building the institution ahead of the evidence is defensible for a field with moral stakes, and it also means the community’s public narrative ran ahead of what any experiment had shown.

Frontier Commercial platforms are now the field’s dominant institution, which changes the incentives. FinalSpark rents remote organoid access; Cortical Labs sells a device. When the infrastructure is a product, throughput is rewarded and standardisation is not, and independent replication — the thing the field most needs — is nobody’s commercial interest. The service model accelerates the production of results without constraining their comparability.

Speculative Embodiment programmes are being funded before the dish result is secure. A Johns Hopkins programme, SURPASS, targets organoid control of robots; funding embodiment while the underlying learning claim is under a reproducibility cloud is the clearest sign the field’s ambition runs ahead of its evidence, and a reason the plain replication in section 7 deserves money before the robots do.

Frontier Oversight sits in a gap between existing regimes. Cultured human neurons are neither a human subject nor an animal model nor an ordinary cell line; standard research-ethics committees are described in the literature as lacking the specialised expertise to evaluate the harder experiments, and there is no dedicated regulatory pathway. The institutional requirement is a review body competent to judge moral status and a provenance-and-consent regime that distinguishes a research culture from a sold compute product.

Speculative Consent is a live legal exposure, not a formality. The human cells originate from a donor; the literature flags unresolved questions about the scope and duration of consent, the donor’s ability to withdraw, and property rights in tissue turned into a commercial product. The first serious legal challenge to a wetware product is likely to be about the cell line, not the computation.

Frontier Funding is real and its shape matters. Public and philanthropic money now backs organoid-intelligence work and the platforms that support it; whether any of it is directed at the unglamorous, non-commercial experiment in section 7 — a blinded replication and a matched-energy benchmark — is the institutional test of whether the field is trying to verify its own claims.

Frontier The welfare question is the one thing this brief refuses to dodge: could a dish of human cortical neurons be a moral patient, and if so, when. The honest current answer is that today’s cultures and organoids almost certainly are not — the ethics literature holds that moral status turns on functional capacities such as sentience and integrated experience, and that current organoids clearly lack the capacity for consciousness or cognition — but the answer is a function of scale, and the field’s entire ambition is to increase the variable the answer depends on.

Established Counting neurons does not settle it, in either direction. The consensus criterion is functional, not structural: possession of human neurons alone does not confer moral status, and no cell count does either. This cuts against the reassurance that an organoid is “only 100,000 cells” and against the alarm that it is “human brain tissue”; both substitute a structural fact for the functional question that actually decides the matter.

Frontier “Sentience” was claimed in a narrow sense and heard in a broad one. The DishBrain paper defined sentience as responsiveness to sensory input under active inference, not felt experience; the Balci and colleagues reply argued even that was premature. The gap between the technical claim and the public reading is itself an ethical hazard, because the reassurances and the alarms in circulation both attach to the broad meaning the authors did not assert.

Speculative The precautionary case does not need proof of experience, only its non-negligible possibility. Because no test can currently detect an inner life in tissue, and because the roadmap proposes to build tissue of steadily greater complexity, a defensible position is to require welfare-relevant safeguards — limits on complexity, on embodiment, on nociceptive circuitry — before the science that would locate a threshold exists. The cost of over-caution is slower research; the cost of under-caution is the creation of a suffering subject nobody could recognise.

Speculative The tractable safeguard is procedural, because the scientific one does not exist. With no marker of experience available, the move a cautious field can make now is to cap organoid maturation, decline to engineer nociceptive circuitry, and withhold rich embodiment until an assessment method exists — limits that cost research speed and buy insurance against building an unrecognised subject.

Handwave The instrumentation gap is the crux and it is not closing. Assessing consciousness through integrated-information measures or neural-correlate tests may be necessary, but a culture that cannot interact with an environment is extraordinarily hard to assess behaviourally, and there is no agreed marker. Until one exists, every welfare judgement in this field is a bet made in the dark, and honesty requires saying so rather than resolving it by assertion.

Frontier Consent and commodification carry their own ethics, separate from sentience. Turning a named donor’s cells into a rented or sold computing product raises questions of ownership, benefit-sharing and dignity that hold even if the tissue never approaches moral patienthood; the ethical account owed here is not only about the organoid’s possible experience but about the person it came from.

Speculative The largest civilizational stake is energy, and it is conditional on a number that does not exist. Inference at scale is running into an energy wall; a computing substrate that delivered useful work at a fraction of silicon’s joules per inference would matter at the level of national power budgets. That is the whole prize — and it is bounded by the unmeasured quantity named in section 5, so the civilizational claim is exactly as strong as the missing benchmark, no stronger.

Frontier The reliable spillover is biomedical, not computational. Whatever happens to wetware computing, the pursuit drives cheaper, more reliable, longer-lived human neural cultures, which feed disease modelling, drug screening and the study of learning itself. A civilization that funded organoid intelligence and got no computer but got better neuroscience would still have made a reasonable trade — a rare case where the consolation prize is worth the entry fee.

Speculative A durable new moral category may be the most consequential output. Cultured human neural tissue that is neither person nor animal nor object forces institutions built on that trichotomy to invent a fourth thing. Getting that category right — before there is any subject to protect — is a civilizational task the field’s embedded-ethics posture is trying, ahead of the evidence, to begin.

Handwave “Minds cultured for labour” is not a policy question this decade. The scenarios that treat cultured neurons as an exploitable workforce presuppose exactly the inner life the field cannot detect and the general capability it has not shown; they are worth naming only so the precautionary work in section 11 is not mistaken for a response to them.

Handwave The science-fiction futures are not on the evidence. Biological general intelligence, conscious machines grown to order, minds cultured for labour — these belong to Artificial General Intelligence and the consciousness briefs, and nothing in the measured record of cultured neurons points at them. The civilizational implication that survives scrutiny is narrow: an energy bet with a biomedical hedge and a novel moral-status problem attached.

These horizons track three clocks running at different speeds: a reproducibility problem that could be resolved this decade, a scaling problem gated on vasculature that is not, and an efficiency claim that stays a claim until someone meters it.

• 10 yr: Frontier An independent, blinded, pre-registered replication of the closed-loop learning result is published, or the learning claim is quietly downgraded to a dynamical-settling effect. Frontier A matched-parameter benchmark against a silicon reservoir appears with joules per inference stated for both sides; the honest prior is that it will not flatter the whole-system energy of the wetware rig. Speculative Standardised tasks, metrics and control batteries emerge, most likely imposed by funders rather than volunteered by platforms. Frontier Cultured cortex becomes a routine instrument for studying learning and disease, its computing pretensions aside.
• 25 yr: Speculative Vascularised organoids exist for biomedical reasons and the computing field inherits them, or the size ceiling holds and “organoid” computing stays a scale-out array of small shells. Speculative If any energy advantage was demonstrated, wetware occupies a narrow niche for adaptive, non-stationary tasks; if it was not, the substrate remains a neuroscience tool. A dedicated oversight and consent regime for cultured human neurons exists, prompted by a legal challenge to a product.
• 50 yr: Speculative Either three-dimensional interfaces turned organoid volume into usable substrate and a genuine biological co-processor exists for a defined task class, or the field is remembered as the era that produced excellent human neural cultures and no computer. Handwave The moral-status question has an operational marker, or it does not and the precautionary limits set decades earlier are still the only safeguard.
• 100 / 250+ yr: Handwave A grown, self-repairing computational substrate made from feedstock rather than fabricated is coherent as a goal and has no demonstrated component today. Handwave If a threshold of experience is ever crossed by engineered neural tissue, the most important thing this generation could leave behind is the category and the caution to recognise it, not the machine.

• Depends on This brief waits on Lab-Grown Organs for vascularised, perfused tissue, without which the organoid size ceiling and the whole-system energy budget are both fixed by oxygen diffusion; on Brain-Computer Interfaces and Neural Interfaces for the high-density recording hardware and decoding mathematics every demonstration borrows; and on Thermodynamics of Computation and Ultra-Efficient Computing Energy Systems for the efficiency comparison that decides whether the field has a reason to exist. What it does not wait on is any new result in reservoir-computing theory, which is mature.
• Requires (not on this map) Five things sit outside the map of research briefs. The first is a matched-parameter benchmark against a silicon reservoir with joules per inference reported for both, the single missing number the field’s thesis rests on. The second is an independent, blinded replication of the closed-loop learning result, which the 2025 non-replication made urgent. The third is industrial: vascularised, perfused organoids that break the oxygen-diffusion ceiling, a problem being solved for biomedicine that this field would inherit. The fourth is institutional: a provenance-and-consent regime for cultured human neurons that separates a research culture from a sold product. The fifth is a market: an application whose economics actually reward wetware inference, without which the platforms rent a demonstration rather than sell a tool.
• Enables Reliable, long-lived human neural culture — the substrate this field needs — supplies disease modelling, drug screening and the empirical study of learning with a better instrument whether or not any computer is ever sold, which is the strongest argument for funding the subject at all. A demonstrated whole-system energy win would give Ultra-Efficient Computing Energy Systems a new substrate class to reckon with.
• Adjacent to Biological Computing, which owns the wider wetware survey and the energy comparison; to Consciousness Research, Machine Consciousness and Synthetic Consciousness, which own the moral-status theories this brief cannot settle; to Basal Cognition, which owns the question of whether simple living matter computes; and to Artificial General Intelligence, the comparison class the roadmap invited and against which no wetware benchmark has been published.

Established “Organoids learned to play Pong” is wrong on the noun. The Pong result used two-dimensional monolayer cultures on an array, not organoids; the organoid result, Brainoware, is a reservoir computer that classified inputs and did not play a game. The single most common claim in circulation fuses two different substrates and two different kinds of result, and the fusion is where the over-reading starts.

Frontier “A dish achieved sentience” is a claim the authors did not make in the sense the phrase is heard. Sentience was defined technically as responsiveness to sensory input under active inference; a formal multi-author reply argued even that was unsupported. The corpus is more useful for holding the fringe claim and showing exactly why it fails than for omitting it: the evidence supports a behavioural change under closed-loop stimulation, contested as to whether it is learning at all, and silent on felt experience.

Frontier “Brain organoids are a million times more efficient than supercomputers” swaps the organoid’s number for the brain’s. The 10⁶-fold figure compares a whole human brain at 20 watts with a supercomputer at 21 megawatts; it says nothing about a 100,000-cell organoid on a perfusion rig, whose joules per inference no one has published. The efficiency of the thing being sold is unmeasured; the efficiency being quoted is the brain’s.

Frontier “The results are reproduced and robust” is not what the record shows. The flagship learning effect was not statistically reproduced in untreated cultures by its own originators in 2025, and the headline reservoir accuracy sits behind a paywall this brief could not open at the primary source. A field marketing reproducible computation should be judged by whether its central results have been independently reproduced, and so far the honest answer is not yet.

Speculative “This is the path to biological AGI” is a genre claim, not an evidential one. Nothing in the measured record — a few summary metrics on a paddle game, one reservoir classification, a service platform — points at general intelligence, and the wetware programme has published no benchmark against the artificial-intelligence systems it invokes. The defensible statement is smaller and truer: this is an energy bet with a biomedical hedge, and its whole case waits on one experiment nobody has run.

Established “It runs on 20 watts” describes a brain, not a rig. A wetware system’s real power draw is dominated by the incubator, perfusion pump, and gas and thermal control that keep the culture alive; the low figure quoted is the isolated tissue’s, not the machine’s, and the machine is the thing that would have to sit in a data centre.

Established “Reservoir accuracy proves the tissue is intelligent” misplaces the training. In reservoir computing the trained part is the silicon read-out; a high score can reflect the read-out interpreting any sufficiently rich dynamics. The one clean way to credit the tissue is a plasticity-blocking control, which Brainoware ran with K252a — the exception that shows how rarely the control is done, not the rule.