Updated November 1, 2026: I rewrote this post to add news since March. New sections cover the Doom demo, learning in organoids, and FDA and NIH moves away from animal testing. I also corrected the fruit fly comparison.
Organoid intelligence: computing with living brain cells
Early in 2026, about 200,000 living human neurons played Doom. They sat on an electrode chip inside Cortical Labs’ CL1, a $35,000 box sold as the first commercial biological computer. They played badly. Even so, they could move, turn and fire inside a 3D game. That demo is the loudest result so far in organoid intelligence, the effort to compute with living human brain tissue.
The idea has real science behind it. It also has more hype than any field I follow. Having read the papers, I find the biology fascinating. The computing claims run years ahead of the data. This post tries to separate the two.
Why build a computer out of neurons?
The pitch starts with power. A single Nvidia H100, the workhorse chip of AI data centers, draws up to 700 watts. Training a frontier model takes many thousands of them. By contrast, your brain runs on about 20 watts, roughly a dim light bulb.
Brains also learn differently. A neural network in software adjusts numbers while the chip stays the same. A biological network rebuilds itself. Neurons strengthen some synapses, prune others and grow new connections in response to what they experience. So the hope is that living tissue could learn from far less data and far less energy.
That hope is reasonable. However, nobody has shown it at any useful scale yet. Keep that in mind through everything that follows.
What is a brain organoid?
A brain organoid is a ball of human neural tissue grown from stem cells. It is not a tiny brain. It has no blood supply, no senses and no body. Still, it organizes itself into layers and cell types that resemble parts of the developing human cortex.
The recipe starts with ordinary cells, usually from skin or blood. Scientists reprogram them into induced pluripotent stem cells, the trick that won Shinya Yamanaka a Nobel Prize in 2012. These cells can become almost any cell type.
Next, the stem cells clump into a small ball. A switch in growth factors pushes them toward neural tissue. Researchers then embed the clusters in a gel such as Matrigel. Spinning bioreactors keep oxygen and nutrients flowing to them. Over months, the cells mature into neurons and glia and form layers that look like an early cortex.
How big is an organoid?
Small. Lena Smirnova’s group at Johns Hopkins describes a typical brain organoid as about 50,000 cells. It is no bigger than a needle tip. For scale, the FlyWire project counted 139,255 neurons in one adult fruit fly brain. A human brain has about 86 billion.
Size runs into a hard limit. Without blood vessels, nutrients reach the center only by diffusion. Past a certain size, the core starves and dies. Several labs have tried to fix this by fusing brain organoids with separately grown blood vessel organoids. A 2022 eLife study produced vessel networks and even blood brain barrier-like structures this way. Nobody has yet shown vessels that carry a steady flow through a large organoid.
Plugging living tissue into a computer
To compute with neurons, you need a way to talk to them. Most groups grow the tissue on a microelectrode array, a chip studded with tiny electrodes. Each electrode can record the spikes of nearby neurons. It can also deliver small pulses of current back into the tissue.
That two-way traffic is the whole game. Software turns information, such as the position of a ball, into a pattern of pulses. The neurons respond. The software reads the response, decides whether it was good, and sends feedback. Over many rounds, the network’s activity shifts. Researchers call this a closed loop.
Pong, then Doom
The best-known closed loop is DishBrain. Brett Kagan and colleagues at Cortical Labs described it in a 2022 paper in Neuron. About 800,000 cultured neurons got better at a simple version of Pong. The cells did not see a screen. They received electrical patterns that stood for the ball and paddle. A hit earned a predictable pulse. A miss brought random noise.
The paper’s title said the neurons “exhibit sentience.” In 2023, 31 neuroscientists led by Fuat Balci answered in Neuron that the claim went far beyond the data. I think they were right. Getting better at Pong through feedback shows adaptation. It says nothing about feeling.
In early 2026 the company pushed the setup further. An independent developer, Sean Cole, used the CL1’s new Python interface to hook its neurons up to Doom. It took him about a week. The cells beat a player firing at random but stayed at what the company called beginner level. Alon Loeffler, an application scientist at the company, was candid. “The system doesn’t actually know it’s playing Doom,” he told Decrypt.
One detail often gets lost. DishBrain and the CL1 use flat layers of neurons, not organoids. They belong to the same field, but the 3D work is a separate line of research.
Brainoware: an organoid that recognized voices
The strongest organoid result so far came from Feng Guo’s lab at Indiana University. In Brainoware, published in Nature Electronics in 2023, the team treated a brain organoid as a reservoir. In reservoir computing, a messy physical system transforms the input. A simple software layer then reads out the answer.
They fed the organoid 240 audio clips of Japanese vowels from eight speakers, converted into electrical pulses. The task was to tell which speaker was talking. On the first day the system was right about 51% of the time. After two days of training, it reached about 78%. When the team blocked synaptic plasticity with drugs, the improvement disappeared.
That last control is what makes the paper worth reading. The organoid was changing itself in response to input. Still, 78% on a small task is far below ordinary software, which handles this easily.
New evidence: the building blocks of learning
In August 2025, Dowlette-Mary Alam El Din, Lena Smirnova and colleagues at Johns Hopkins published the most careful test yet. Writing in Communications Biology, they asked whether neural organoids have the basic cellular machinery for learning and memory.
They found much of it. The organoids showed synaptic plasticity specific to the input they received, over both short and long time scales. They switched on immediate early genes, the genes neurons activate after strong stimulation and need for lasting memories. Their networks also ran near a state called criticality, which brains seem to use to process information efficiently.
The authors described this as foundational work, and that framing is fair. Having the parts for learning is different from learning. Still, this is the kind of paper the field needs more of, slow and specific and honest about its limits.
Silicon and wetware side by side
It helps to stop thinking of neurons as a replacement for chips. Silicon is fast, precise and repeatable. Living tissue is slow, noisy and adaptive. The realistic future, if there is one, pairs them. Silicon would handle math, storage and bulk computation. Tissue would handle tasks where learning from feedback counts more than raw speed.
The power numbers deserve a second look. The brain’s 20 watts covers the brain alone. A CL1 has to keep its neurons alive with pumps, heaters and gas control. Reports put the whole unit’s draw at 850 to 1,000 watts. For now, then, a biocomputer uses more power than a single H100 while doing far less.
Access has become cheap, though. Cortical Labs rents time on a live culture for about $300 a week. The Swiss company FinalSpark rents remote access to its organoids for about $500 a month to universities. Its organoids survive about 100 days. That low price means many more labs can try ideas, which is good. Even so, what the money buys is a research tool, not a faster computer.
Organoid intelligence in medicine and the end of animal testing
The nearest payoff is not computing at all. It is medicine. Researchers can turn a patient’s skin cells into stem cells. From those, they grow brain tissue that carries the patient’s genome. They can then watch how the disease unfolds in that tissue and test drugs on it.
The results so far are real but narrow. In 2016, Li-Huei Tsai’s lab at MIT grew organoids from people with inherited Alzheimer’s disease. As the tissue aged, it built up amyloid plaques and tangled tau, the two hallmarks of the disease. Drugs that block amyloid production reduced both. That makes organoids good for studying mechanism and screening compounds. They are still a long way from predicting how a given patient will respond to a given drug.
The bigger shift is in regulation. Nine in ten drugs that enter human trials fail. A 2022 analysis by Duxin Sun’s group blamed 40 to 50% of failures on lack of efficacy. About 30% came from toxicity. Animal models share the blame, because a drug can work in a mouse and do nothing in a person.
Regulators have started to act. The FDA Modernization Act 2.0 of 2022 ended the legal requirement for animal tests before human trials. Then in April 2025, the FDA laid out a roadmap to phase out animal testing for monoclonal antibodies. It named organoids and organ chips as replacements. In July 2025, NIH said it would no longer seek proposals built only on animal models.
I expect organoids to earn their keep here first. Better drug testing is a modest goal, but a reachable one.
The hype problem
In November 2025, organoid researchers met at Asilomar in California, the site of the famous 1975 meeting on recombinant DNA. According to STAT, many of them worried about the biocomputing companies. They feared loose talk of “organoid intelligence” could bring a public backlash or sweeping laws. Medical research would suffer too.
The meeting’s organizer, Sergiu Pasca of Stanford, urged the field to use terms “that neither hype nor misrepresent the work.” Madeline Lancaster, who made the first brain organoids, has said the science for organoid computing is not there yet. Tony Zador of Cold Spring Harbor Laboratory goes further. He calls the push for organoid intelligence on par with silicon AI a dead end.
I side with the cautious camp on timing. The cells are interesting. Calling them “intelligent” today mostly helps fundraising.
Consciousness, consent and ownership
The ethical questions are real, even if they are early. A brain organoid is human tissue that fires in organized patterns. If organoids grow larger, gain blood supply and receive sensory input, at what point could they feel anything? Nobody knows, partly because we cannot agree on what consciousness is. I covered that debate in my post on the neural correlates of consciousness.
My own view is that today’s organoids are nowhere near that line. Even so, the field should set rules before it gets close, not after.
Ownership is the other hard question. If a biocomputer helps find a new drug, who owns the result? The donor whose cells carry the genome? The lab that grew the tissue? The company that rents the hardware? Current law was not written with self-replicating human tissue in mind. The Baltimore Declaration, signed by researchers in 2023, calls for ethicists to work inside these projects from the start. Johns Hopkins has since done that on a new NSF-funded project to train organoids on simple games.
Where organoid intelligence goes from here
Here is where I come down. The biology of organoids is some of the most interesting work in neuroscience. It has already taught us about human brain development and disease. The computing story is much weaker. Four years after DishBrain, living neurons still do simple tasks worse than cheap software. And their life support burns more power than the chips they were meant to beat.
That could change. Vascularized organoids, better interfaces and the Johns Hopkins learning work all point somewhere. I would watch the medical uses closely and treat claims about living AI with patience and suspicion.
Your thoughts?
Would you trust a drug test run on organoids grown from your own cells more than one run on mice? And where would you draw the line on how complex a lab-grown brain should be allowed to get? Tell me in the comments.
Bleeding Edge Biology recommends: books, papers and talks on organoid intelligence
Books
- What Is Intelligence? by Blaise Agüera y Arcas
Agüera y Arcas treats intelligence as a physical process that shows up throughout biology, in cells and in software. - Neuroethics by Walter Glannon
A clear introduction to the moral questions raised by intervening in brains. - Human Brain Organoids
The most technical book on this list. It covers how organoids are made, what they model well and where they fall short.
Papers
- Smirnova et al. (2023), “Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish”
The paper that named the field and laid out its goals. - Alam El Din et al. (2025), “Human neural organoid microphysiological systems show the building blocks necessary for basic learning and memory”
The careful 2025 follow-up described above. - Kagan et al. (2022), “In vitro neurons learn and exhibit sentience when embodied in a simulated game world”
The DishBrain paper. Read the methods before the title. - Cai et al. (2023), “Brain organoid reservoir computing for artificial intelligence”
Brainoware, with the plasticity control that makes it convincing.
Talks and videos
- “Growing mini brains to discover what makes us human” by Madeline Lancaster
Lancaster explains how she grew the first cerebral organoids and what they reveal about human brain development. - “What are mini brains?” by Madeline Lancaster for TED-Ed
A short animated primer. - Brain Scientist On AI: You Aren’t Ready For What Happens Next
Alysson Muotri of UC San Diego on what organoids reveal about brain development and the AI link. - Living Human Brain Cells Play DOOM on a CL1
The Cortical Labs demo itself. - The Big Question: Dr. Sergiu Pașca
Pasca on the science and ethics of his assembloids, organoids from different brain regions fused together. - Growing Brain Cells in the Lab, from Healthy Minds with Dr. Jeffrey Borenstein
An overview of organoids in disease research and drug testing. - “Organoid Intelligence” by Thomas Hartung at XPANSE 2024
Hartung, the field’s leading advocate, makes the case for biocomputing.
Articles and websites
- Brain organoid pioneers fear inflated claims about biocomputing could backfire (STAT, 2025)
The report from Asilomar. - ISSCR Guidelines
The main international rules for stem cell research, including organoids. - Johns Hopkins Center for Alternatives to Animal Testing
Home of Hartung and Smirnova’s organoid intelligence work. - Cortical Labs
The company behind DishBrain and the CL1.
