A Melbourne biotech company called Cortical Labs has spent the past four years building computers whose processors are not silicon transistors but living human neurons. Its newest device, the CL1, made headlines when an independent developer connected roughly 200,000 lab-grown neurons to a version of the 1993 video game Doom, and the cells learned to move and shoot within about a week. The demonstration reopened a debate that started in 2022, when the same company first claimed its neurons showed “sentience,” and triggered a formal rebuttal from 30 neuroscientists. This piece lays out what the CL1 actually is, what it can and cannot do, the two decades of research that led to it, and why the unresolved question underneath all of it — whether anything is being felt inside that dish — still has no scientific answer.
Cortical Labs’ CL1 Biocomputer Learned Doom’s Basic Controls Within About a Week
The Doom demonstration used a single CL1 unit carrying about 200,000 human neurons cultured on a silicon multielectrode array. An independent researcher named Sean Cole, working with Cortical Labs’ Python-based API, built the connection between the game and the cell culture. According to Cortical Labs application scientist Dr. Alon Loeffler, the team and Sean Cole “coded the first working version of Doom using the Cortical Labs API, and running on a CL1.” Multiple outlets reported that Cole had little to no prior experience with biological computing and completed the build in under a week using the company’s new high-level programming interface.
The setup rewarded the neurons electrically when their output correctly aimed at or eliminated an in-game enemy, and disrupted them with less predictable stimulation when they failed — a punishment-and-reward loop directly descended from Cortical Labs’ earlier Pong experiments. Chief Scientific Officer Dr. Brett Kagan was careful about what the result did and did not show. “So can the cells learn to play Doom? Yes, they’re receiving information, they’re sending commands to move their character around, they’re able to find enemies, and shoot. Is it an esports champion? Absolutely not.”
That distinction matters for how this article treats the whole story. The Doom demo is a company-run product showcase, not a peer-reviewed study. The peer-reviewed science behind it — the 2022 DishBrain paper — is a separate, published, and independently scrutinized piece of work, and this article treats the two very differently.
AI’s Runaway Power Demand Is Driving Scientists Toward the Brain’s 20-Watt Efficiency
The reason a biological computer exists at all traces back to electricity, not neuroscience. The International Energy Agency projects that global data centre electricity demand could reach roughly 1,000 terawatt-hours in 2026 — on the order of Japan’s entire annual electricity use — driven overwhelmingly by AI workloads. That demand growth has already pushed major technology companies toward nuclear power: in September 2024, Microsoft signed a 20-year agreement with Constellation Energy to restart the 835-megawatt Unit 1 reactor at Three Mile Island, a plant that had been shut for economic reasons since 2019, specifically to supply a data centre. Google, Amazon, and Meta have signed comparable nuclear and small-modular-reactor deals since.
Against that backdrop, the human brain looks strange. Fred Jordan, co-founder of the Swiss biocomputing company FinalSpark, put the comparison plainly: “A human brain consumes roughly 20 watts to power 100 billion neurons, which collectively make about 10,000 connections.” Reproducing that same volume of connections and adaptability in silicon, by FinalSpark’s own estimate, would require millions of times more power. That gap — not any claim about intelligence or awareness — is the actual commercial argument for wetware computing, and it is the argument Cortical Labs leads with in its own marketing.
Two Decades of Wetware: From a Lamprey Brainstem to Lab-Grown Organoids
The CL1 did not appear from nowhere. It sits at the end of a research line that stretches back at least 25 years, and three separate breakthroughs had to converge before it was possible.
The first closed-loop “hybrot” — a term coined for hybrid biological robots — came from Ferdinando Mussa-Ivaldi’s lab at Northwestern University in 2000. His team removed the brainstem and part of the spinal cord from a sea lamprey, kept the tissue alive in a cooled, oxygenated salt bath, and wired it to a small wheeled Khepera robot. The lamprey tissue’s natural vestibular circuitry, normally used to keep the animal oriented in water, was repurposed to steer the robot toward or away from light.
Eight years later, Kevin Warwick and Ben Whalley at the University of Reading built a more direct successor. Their robot, nicknamed Gordon, ran on a culture of roughly 50,000 to 100,000 rat neurons spread across a five-inch, 60-electrode array, connected wirelessly to a two-wheeled robot body. The neurons received feedback from the robot’s ultrasound sensors and, over about a week of exposure, learned to steer the robot away from walls. Warwick noted that different neuron cultures behaved almost like individual personalities, some more active than others.
While that line of robotics research grew, a separate and arguably more important breakthrough was happening in stem cell biology. In 2006, Kazutoshi Takahashi and Shinya Yamanaka at Kyoto University showed that a mature, already-differentiated cell could be reset to an embryonic-like state using just four transcription factors — Oct3/4, Sox2, Klf4, and c-Myc, now called the Yamanaka factors. Their original paper reprogrammed mouse fibroblasts; a follow-up a year later did the same with adult human skin cells. Yamanaka shared the 2012 Nobel Prize in Physiology or Medicine for the discovery. Its consequence for biological computing was direct: scientists no longer needed to extract neurons from a living or recently deceased brain. They could take a small skin sample from a donor, reprogram it into a stem cell, and differentiate it into neurons in a dish.
The third piece was three-dimensional growth. Neurons cultured flat on a dish connect only to their immediate neighbours. Left to grow in three dimensions instead, stem-cell-derived tissue self-organizes into layered structures that loosely resemble developing brain regions — these are called organoids. The trade-off, which becomes important later in this piece, is that a flat culture can sit directly on a grid of electrodes and be read in full, while an organoid’s electrodes can only reach its outer surface.
Cortical Labs combined all three advances in 2022 with DishBrain, a system built from roughly 800,000 neurons of mixed human and rodent origin, grown flat across a high-density multielectrode array. In the resulting paper, published in Neuron and led by Brett Kagan, the team reported that the culture displayed measurable, goal-directed changes in its firing patterns after five minutes of real-time Pong gameplay — activity the control conditions did not show. The paper’s title called this “sentience,” a word choice that would define much of the public reaction to everything Cortical Labs has done since.
How Cortical Labs Turned Doom’s Video Feed Into Electrical Stimulation Patterns
The CL1 that ran Doom is a commercial product, not a research prototype. Cortical Labs announced it at Mobile World Congress in Barcelona on March 2, 2025, describing it as the first “code-deployable” biological computer. Each self-contained unit houses a planar silicon electrode array on which neurons are grown, sealed inside an internal life-support system with its own pumps, gas mixing, temperature regulation, and fluid filtration. Cortical Labs says the neurons can survive up to six months this way without an external computer attached. A standalone CL1 costs roughly $35,000; buyers who deploy units in a 30-device server rack pay closer to $20,000 per unit, and a full rack draws 850 to 1,000 watts in the range of a single mid-tier GPU server. The company shipped its first 115 units in 2025 and also offers remote, cloud-based access it calls “Wetware-as-a-Service.”
Running Doom on that hardware required solving a problem the neurons cannot solve themselves: they have no eyes, so they cannot see pixels, enemies, or level geometry. Sean Cole’s system, released publicly on GitHub after the demo, breaks the pipeline into three distinct layers. A conventional computer runs Doom itself and a reinforcement-learning agent built on Proximal Policy Optimization, a widely used machine-learning method. An encoder within that agent decides which electrical stimulation patterns to send to the neurons based on what is happening on screen. The neurons respond with spikes of electrical activity, and a decoder translates those spikes back into game commands — move forward, turn, shoot. Cortical Labs’ CTO David Hogan described Cole’s contribution as piping the game’s “video feed into patterns of electrical stimulation” that gave the culture something to respond to in real time.
What the Open-Source Code and Ablation Tests Actually Showed
Cole’s published documentation is more revealing than the demo video itself, because it addresses a doubt that has followed this entire field since the 2000 lamprey experiment: is the biological tissue actually doing the work, or is a conventional AI system quietly doing all of it while the neurons sit along for the ride?
Cole ran a direct test of this. He built an “ablation mode” into his own code that swaps the live neural signal for pure random noise, and reported that when he did this, the system’s learning stopped. That result indicates the neurons are contributing something functionally necessary, not decorative.
At the same time, Cole’s own README is candid about how much of the computation happens elsewhere. He wrote that his decoder “tends to start becoming a policy head,” meaning the conventional software layer risks learning to solve the game on its own and routing around the biological component almost entirely. Reporting on the code by RD World Online summarized the honest version of the story: the architecture is mostly a standard reinforcement-learning agent, with the living neurons supplying a real but comparatively small piece of the total computation.
Cortical Labs Frames CL1 as an Interface Milestone, Not an Intelligence Breakthrough
Both Kagan and Loeffler were consistent in interviews about what the company believes it actually achieved. Kagan described the underlying advance as “adaptive real-time goal-directed learning,” and emphasized that reaching Doom took a fraction of the engineering effort the original Pong system required. Loeffler recounted that the team’s earliest low-level code needed roughly 18 months of development, while the newer high-level Python interface let Cole reach a working Doom build in a matter of days once it existed. Cortical Labs describes this as having “solved the interface problem” — the ability to reliably read from and write to living neural tissue in real time — and frames that, rather than any claim about intelligence, as the milestone worth publicizing.
The original 2022 DishBrain paper made a narrower and more technical claim than most coverage of it suggested. The authors termed the phenomenon “synthetic biological intelligence” and concluded that the work demonstrates the computational power of living neurons to learn adaptively in exchange with their environment. In later exchanges with critics, Kagan and his co-authors maintained that words like “sentience” were used in a specific technical sense tied to a mathematical framework called the free energy principle, in which “surprise” refers to statistical prediction error rather than a felt experience.
Why Neuroscientists Still Reject the “Sentience” Label Cortical Labs Used in 2022
That technical defense did not settle the argument. In March 2023, a formal response letter appeared in Neuron, signed by 30 neuroscientists across multiple institutions and led by Fuat Balci of the University of Manitoba. The group argued that terms like intelligence and sentience were “not based on any established or robust consensus” and that Cortical Labs’ claims rested on comparatively weak evidence for the strength of the words being used. They were not disputing that the neurons changed their firing patterns in response to training — that part of the result held up. Their objection was to the leap from “measurable plasticity” to loaded words like sentience and intelligence, which they warned could mislead both the public and the scientific record.
A separate 2023 analysis in Science and Engineering Ethics examined the DishBrain work specifically for its ethical implications and concluded that the experiment does not provide clear evidence of artificial suffering, while still arguing that the ethical questions deserve serious, ongoing attention as the underlying technology scales up. Cortical Labs appears to have taken that message seriously: the company has since worked with some of its most vocal early critics to build a formal ethics framework around its research, according to interviews Kagan gave in the years following the backlash.
Outside commentary has generally landed in the same place. Suhas Kumar, a principal scientist at Sandia National Laboratories who works on neuromorphic computing and was not involved in the study, called the original DishBrain work a genuinely notable technical achievement while pushing back on the sentience framing directly, noting that a network of that size is closer in scale to a bumblebee’s brain than a human one.
Why Scientists Can’t Yet Confirm or Rule Out Any Subjective Experience in These Cells
It is worth separating two very different questions that get blurred together in almost every popular account of this technology. The first — can a network of living neurons adapt its behaviour in response to reward and punishment signals — has a documented, peer-reviewed, replicated answer: yes, within the narrow conditions tested. The second — does anything it is like to be that network of neurons, does it feel anything at all — is what the philosopher David Chalmers named the “hard problem” of consciousness, and no experiment run on cultured neurons, organoids, or even intact human brains has resolved it.
Some evidence from human medicine shows how easy that gap is to miss even in a fully formed brain. In 2006, Adrian Owen and colleagues at the University of Cambridge asked a 23-year-old woman diagnosed as vegetative to imagine playing tennis while inside an fMRI scanner. Her brain activity in that moment was indistinguishable from healthy volunteers performing the same mental task, despite her showing no outward sign of awareness. A larger follow-up study by Martin Monti and colleagues, published in the New England Journal of Medicine in 2010, tested 54 patients with disorders of consciousness the same way. Five of them, about 9 percent, could willfully modulate their brain activity on command; within the subgroup specifically diagnosed as vegetative, that rose to roughly 17 percent, or about one in six. Those figures are meaningfully lower than the “one in four” statistic that circulates widely in popular science coverage of this research, and it is worth correcting: the best-supported number from the primary literature is closer to one in six within that specific patient group, and closer to one in ten across the full study population.
That research shows something narrower and more useful than a headline figure, though: consciousness can exist without any behavioural sign a bedside clinician can detect, even in a person with an intact, fully developed human brain. Extending that finding to an isolated culture of a few hundred thousand neurons in a dish is a much larger and currently unsupported leap.
Some theoretical neuroscientists argue the leap may be smaller than intuition suggests. Mark Solms, Chair of Neuropsychology at the University of Cape Town, has proposed that consciousness is not primarily a cognitive phenomenon generated by the cortex, but an affective one generated deeper in the brainstem, tied directly to an organism’s basic drive to maintain internal balance. In his account, “the primary function of consciousness is not perceiving or remembering or comprehending, but feeling.” Solms draws on cases such as children born without a cerebral cortex who still show clear emotional responses, and he frames feeling itself as an old, homeostatic drive rather than a late evolutionary add-on. This is a minority position within neuroscience, not a settled consensus, and Solms himself frames it as a working theory rather than a proven mechanism.
Two additional facts complicate any simple test for feeling in cultured tissue. Insects lack a mammalian-style visual cortex entirely and still process visual information through structurally different optic lobes, which undercuts the idea that any one specific brain structure is a required ingredient for sensation. And the free energy principle that Cortical Labs invokes to explain why its neurons “avoid surprise” is itself a mathematical framework for describing self-organizing systems in general — applicable, in principle, to a thermostat as much as to a brain — which is exactly the kind of overreach the 2023 Balci rebuttal warned against when the same language gets applied to a dish of neurons without qualification.
Cortical Labs Is Not Alone: FinalSpark Rents Access to Human Brain Organoids by the Month
Cortical Labs is not alone in pursuing this. A Swiss company called FinalSpark, founded by Fred Jordan and Martin Kutter, takes a different biological approach: rather than growing neurons flat, it cultures them as three-dimensional organoids. Its Neuroplatform houses up to 16 organoids across four multi-electrode arrays, with each array’s eight electrodes both stimulating and recording from the tissue’s outer surface. Since 2024, the company has offered paid remote access — around $500 per user per month for shared use of four organoids — and reported that roughly nine academic institutions had adopted the platform. FinalSpark claims its system uses up to a million times less power than an equivalent digital process, though that figure comes from the company itself and has not been independently benchmarked against production AI hardware at matched task performance.
The organoid approach and Cortical Labs’ flat-culture approach represent a genuine trade-off, not just a branding difference. Three-dimensional organoids grow into something structurally closer to real brain tissue, but current electrode arrays can only read and stimulate their outer layer, leaving most of the tissue electrically invisible. Flat cultures give up that structural realism in exchange for full electrode contact with every neuron in the network — which is precisely why Cortical Labs has kept DishBrain and CL1 flat rather than moving to organoids, despite organoids being the more biologically faithful model. Reports on FinalSpark’s organoids have also noted shorter viability windows, with some coverage citing survival estimates around 100 days, compared with Cortical Labs’ stated six-month lifespan for CL1’s flat cultures.
The Limits of What’s Actually Been Demonstrated So Far
A few things are worth stating plainly, in the same register as the rest of this piece.
The Doom demonstration is a marketing and engineering showcase run by the company that sells the hardware, not an independently reviewed scientific finding on its own. The peer-reviewed science underneath it — the 2022 DishBrain paper — has been published, scrutinized, and formally challenged by 30 neuroscientists over its specific choice of language, though not over its core behavioural result.
Viability remains a real engineering constraint rather than a solved problem. Cortical Labs’ six-month figure for CL1 and the shorter lifespans reported for organoid-based systems both represent current limits on how long these systems can run before the tissue needs replacing, not evidence of long-term durability at scale.
No accepted method currently exists to detect sentience, suffering, or subjective experience in isolated neural tissue, in either direction. The 2023 ethics review of DishBrain found no clear evidence of artificial suffering, but explicitly did not rule the possibility out, and no technology developed since has closed that gap.
Sean Cole’s own documentation shows that a substantial share of the “learning” behind the Doom demo reflects a conventional reinforcement-learning agent working alongside the neurons, not the neurons acting alone — a distinction the company’s own developer was careful to document, even as his ablation tests confirmed the biological layer is doing real, necessary work.
Finally, the energy-efficiency claims driving commercial interest in this technology — Cortical Labs’ framing and FinalSpark’s stated million-fold figure alike originate from the companies building and selling the hardware. Outside specialists, including Michael Barros at the University of Essex, have pointed to genuine construction and scalability challenges that stand between today’s lab-scale demonstrations and any real competition with silicon at data-centre scale.
References:
- Kagan, B.J., Kitchen, A.C., Tran, N.T., et al. (2022). “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.” Neuron, 110(23), 3952–3969.e8.
- Balci, F., Ben Hamed, S., Boraud, T., et al. (2023). “A response to claims of emergent intelligence and sentience in a dish.” Neuron, 111(5), 604–605.
- Milford, S.R. (2023). “Playing Brains: The Ethical Challenges Posed by Silicon Sentience and Hybrid Intelligence in DishBrain.” Science and Engineering Ethics, 29.
- Takahashi, K., and Yamanaka, S. (2006). “Induction of Pluripotent Stem Cells from Mouse Embryonic and Adult Fibroblast Cultures by Defined Factors.” Cell, 126(4), 663–676.
- Owen, A.M., Coleman, M.R., Boly, M., et al. (2006). “Detecting Awareness in the Vegetative State.” Science, 313(5792), 1402.
- Monti, M.M., Vanhaudenhuyse, A., Coleman, M.R., et al. (2010). “Willful Modulation of Brain Activity in Disorders of Consciousness.” New England Journal of Medicine, 362, 579–589.
- Reger, B.D., Fleming, K.M., Sanguineti, V., Alford, S., and Mussa-Ivaldi, F.A. (2000). “Connecting brains to robots: an artificial body for studying the computational properties of neural tissues.”
- Solms, M. (2021). The Hidden Spring: A Journey to the Source of Consciousness. W.W. Norton & Company.
- Cortical Labs, CL1 product materials and Doom demonstration coverage — Tom’s Hardware, Decrypt, Popular Science, GameDaily, RD World Online, igor’sLab, Data Center Dynamics, BigGo Finance (2025–2026).
- FinalSpark, Neuroplatform materials and coverage — Tom’s Hardware, ScienceAlert, BioAlps, KNWA/KFTA (2024).