Organoid intelligence (OI) is an emerging field that uses lab-grown, three-dimensional clusters of human brain cells, called brain organoids, as living computing hardware. Organoids sit on microelectrode arrays that stimulate them and record their electrical activity, while software encodes inputs and decodes responses, so the tissue can process information and learn.
Key takeaways
- Organoid intelligence was formally proposed as a field in 2023 by Smirnova, Hartung and colleagues in Frontiers in Science, building on decades of work with cultured neurons.
- An OI system has four core parts: brain organoids, a microelectrode array (MEA), a closed-loop control system, and machine learning for encoding and decoding.
- Landmark results so far are proofs of concept: DishBrain playing Pong (2022), Brainoware recognizing vowel sounds (2023), and the remotely accessible Neuroplatform (2024).
- The long-term motivation is energy and learning efficiency: a human brain runs on roughly 20 watts, while AI’s electricity demand is growing quickly.
- Today, organoids are slow, variable, short-lived and far smaller than a brain. They do not outperform conventional computers on any practical task.
- Ethics (consent, donor rights, the question of sentience) is treated as part of the research program from the start, not an afterthought.
What is organoid intelligence, exactly?
At its simplest, organoid intelligence is computing with living human neurons grown in three dimensions. The word “organoid” refers to a miniature, simplified, self-organizing tissue grown from stem cells that reproduces some of the structure and function of a real organ. A brain organoid is a millimeter-scale sphere of neural tissue that contains many of the cell types found in the developing human brain, including neurons and supporting glial cells.
These neurons do what neurons do everywhere: they generate electrical impulses called action potentials, connect to each other through synapses, and change those connections in response to activity. That last property, synaptic plasticity, is what makes the idea of computing with them interesting. A network that rewires itself based on experience can, in principle, learn.
OI researchers try to turn that capacity into something useful. They place an organoid on an electrode array, send it patterned electrical stimulation that represents some input (a sound, a game state, a sequence of numbers), record how it responds, and use software to interpret the response as an output. If feedback is delivered in the right way, the network’s behavior can change over time.
Why “intelligence” and not just “computing”?
The name signals an ambition rather than a current capability. The 2023 paper that defined the field framed OI as the pursuit of biological learning and memory in a dish, potentially leading to forms of computing that are more efficient than silicon for certain tasks. Nobody claims that today’s organoids are intelligent in the everyday sense. The term describes a research direction: harnessing the learning mechanisms of real neurons, studied with tools from neuroscience, bioengineering and computer science.
Related terms you will see
Several overlapping terms appear in the literature, and it helps to know how they relate:
- Biocomputing is the broad umbrella: any computation performed with biological materials, including DNA computing and engineered cells.
- Wetware computing emphasizes the use of living tissue as hardware (“wetware” instead of hardware or software).
- Synthetic biological intelligence (SBI) is the term preferred by Cortical Labs, the company behind DishBrain, as described in a 2023 review by Kagan and colleagues in Biotechnology Advances.
- Organoid intelligence is the specific branch that uses three-dimensional brain organoids, rather than flat (2D) neuron cultures, as the computing substrate.
Where did the term organoid intelligence come from?
The phrase was set out in a February 2023 article in Frontiers in Science by Lena Smirnova, Thomas Hartung and an international group of co-authors, led from Johns Hopkins University. Titled “Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish”, the paper proposed OI as a coordinated, multidisciplinary field and outlined what it would take to build it.
The authors argued that three developments made the idea newly practical: brain organoids that self-organize into more brain-like structures, better electrophysiology tools for interfacing with 3D tissue, and machine learning methods capable of making sense of complex neural data. They also proposed an “embedded ethics” approach, in which ethicists, the public and researchers work together throughout the program rather than reviewing it after the fact.
A companion paper in the same journal, the Baltimore declaration (Hartung et al., 2023), summarized the outcome of an early workshop on organoid intelligence and set out shared principles for exploring the field responsibly.
The longer history behind OI
OI did not appear from nowhere. It rests on at least three older lines of work:
- Cultured neurons on electrodes. Researchers have grown neurons on planar electrode arrays for decades. In 2001, DeMarse and colleagues described a “neurally controlled animat”, in which a culture of rat neurons on an electrode array controlled a simulated body, published in Autonomous Robots.
- Stem cell reprogramming. In 2006, Takahashi and Yamanaka showed in Cell that adult cells could be reprogrammed into pluripotent stem cells using a small set of factors. These induced pluripotent stem cells (iPSCs) mean human neurons can be made from a skin or blood sample.
- Brain organoids. In 2013, Lancaster and colleagues reported in Nature that human pluripotent stem cells could self-organize into “cerebral organoids” with regions resembling parts of the developing brain. Our guide on how brain organoids are grown covers this step in detail.
OI ties these together and adds modern machine learning and engineering.
How does organoid intelligence differ from AI, neuromorphic computing and biocomputing?
People often confuse OI with neighboring fields. The key distinction is the substrate: what physically does the computing.
| Approach | Substrate | How it learns | Maturity | Main strength | Main weakness |
|---|---|---|---|---|---|
| Conventional AI (deep learning) | Silicon chips (CPUs, GPUs, accelerators) | Gradient descent on artificial weights | Mature, widely deployed | Speed, precision, scale, reliability | High energy use for training and inference |
| Neuromorphic computing | Silicon chips designed to mimic neurons and synapses | Spiking rules, on-chip plasticity, or offline training | Research and early products | Low power for event-driven workloads | Hard to program; smaller software ecosystem |
| Biocomputing (broad) | DNA, proteins, engineered cells, neurons | Depends on system | Mostly research | Massive parallelism, chemical sensing | Slow, hard to interface |
| Organoid intelligence | Living 3D human brain organoids | Real synaptic plasticity, guided by feedback | Early research | Native learning mechanisms, potential energy efficiency | Variability, short lifespans, low-bandwidth interfaces |
OI versus artificial intelligence
Modern AI runs artificial neural networks on digital hardware. Those networks are mathematical abstractions loosely inspired by the brain. In OI, the network is not a model of neurons; it is made of neurons. The two fields depend on each other, though. Every OI system uses conventional software to translate inputs into stimulation and to decode responses, and AI researchers look to biology for ideas about efficient learning.
OI versus neuromorphic computing
Neuromorphic engineering, a term associated with Carver Mead’s 1990 paper “Neuromorphic electronic systems” in Proceedings of the IEEE, builds silicon circuits that behave more like neurons. IBM’s TrueNorth chip, described by Merolla and colleagues in Science in 2014, implemented a million spiking neurons on one chip, running spiking neural networks that communicate with discrete events rather than continuous values. Neuromorphic chips borrow the brain’s design principles. OI uses the original material. Both pursue energy-efficient, event-driven computation, and our article on biocomputing versus silicon energy efficiency compares them in more depth.
OI versus 2D neuron cultures
DishBrain, the best-known demonstration of learning in a dish, used flat cultures of neurons rather than organoids. Organoids are thicker and more structured, with layered and region-like organization and denser 3D connectivity. Proponents of OI argue that this richer structure could support more complex computation. The trade-off is that 3D tissue is harder to keep alive and harder to record from, because flat electrodes only touch its surface.
OI is one branch of a broader field. For the full family tree, from DNA computing to engineered cells to living neurons, read biological computing: from DNA molecules to living neurons. For how living tissue and silicon divide the work inside a real system, see bio-silicon hybrid computing.
What are the components of an organoid intelligence system?
Every OI setup, whatever the lab, needs the same basic building blocks.
1. Brain organoids
The organoids are usually derived from human iPSCs. Through a series of steps using signaling molecules and a supportive 3D matrix, the stem cells are guided to become neural progenitors and then neurons and glia. Over weeks to months, the tissue develops spontaneous electrical activity. Trujillo and colleagues reported in Cell Stem Cell (2019) that cortical organoids developed increasingly complex oscillatory network activity over months, with features that resembled aspects of the preterm human brain’s electrical development.
A typical organoid is a few millimeters across at most. For scale, Azevedo and colleagues (2009) estimated that the adult human brain contains about 86 billion neurons. Organoids contain many orders of magnitude fewer.
2. Microelectrode arrays
A microelectrode array (MEA) is a chip covered in tiny electrodes that can both record voltage changes near neurons and deliver small electrical pulses. High-density MEAs pack thousands of electrodes onto a small area, allowing researchers to track activity at fine spatial resolution. MEAs are the main interface for OI because they are non-invasive to the tissue and can run for long periods. See our explainer on microelectrode arrays for how they work.
Raw MEA signals need processing. Spike sorting algorithms detect individual action potentials and attempt to assign them to specific neurons. The broader discipline of measuring electrical activity in cells is electrophysiology.
3. The closed loop
A closed loop means the system responds to the neurons in real time. Activity is recorded, interpreted, and turned into new stimulation within milliseconds. Closed-loop stimulation is what allows feedback-driven learning: the culture’s output changes the input it receives next, just as an animal’s actions change what it senses.
Without a closed loop, you can observe an organoid but you cannot train it. Latency matters here. If feedback arrives too late, the neurons cannot associate it with the activity that caused it.
4. Machine learning for encoding and decoding
Neurons do not understand pixels or audio files. Software has to translate inputs into spatial and temporal patterns of stimulation (encoding) and translate recorded activity back into actions or predictions (decoding). This is the domain of neural encoding and decoding.
In some designs, almost all the “learning” happens in the decoder, with the tissue serving as a rich, nonlinear transformer of inputs. That approach is called reservoir computing. In other designs, the aim is for the tissue itself to change in response to feedback.
5. Life support and data infrastructure
Organoids need temperature control, gas exchange, sterile handling and regular replacement of culture medium. Microfluidics, systems that move small volumes of fluid through engineered channels, is one approach to automating this. Experiments also generate large volumes of time-series data, which is why standards such as Neurodata Without Borders (NWB) matter for storing and sharing recordings reproducibly.
How do organoids actually learn?
Learning in a dish relies on the same basic mechanisms that operate in the brain. Hebbian learning, often summarized as “cells that fire together wire together”, describes how connections strengthen when neurons are active together. Spike-timing-dependent plasticity refines this: whether a synapse strengthens or weakens depends on the precise order and timing of the spikes on each side of it.
The challenge is to give a culture feedback that steers these local rules toward a useful global behavior. DishBrain’s designers drew on the free energy principle, a theory proposed by Karl Friston (summarized in Nature Reviews Neuroscience, 2010) that self-organizing systems act to minimize surprise, or unpredictable input. In practice, this meant giving neurons predictable stimulation when they performed well and unpredictable, noisy stimulation when they failed. The idea is that the network reorganizes its activity to make its sensory world more predictable.
Spontaneous activity also matters. Organoids frequently show network bursts, synchronized episodes in which many neurons fire together. Bursts are a sign of a connected, maturing network, but excessive bursting can swamp the subtler patterns that carry information. Our deep dive on how neurons learn in a dish covers training protocols and their debates.
What are the landmark results in organoid intelligence?
A handful of studies define the field so far. Each is a proof of concept, not a product.
DishBrain: neurons playing Pong (2022)
Kagan and colleagues at Cortical Labs reported in Neuron in 2022 that cultures of human and mouse neurons grown on a high-density MEA could be “embodied” in a simulated game of Pong. Stimulation on specific electrodes encoded the ball’s position, activity in designated regions moved the paddle, and feedback followed the free energy logic described above. The cultures improved their performance, measured as longer rallies, within minutes of gameplay, while control conditions did not show the same improvement.
The paper’s use of the word “sentience” drew significant debate among neuroscientists and ethicists, many of whom argued that responsiveness to feedback is not the same as sentience in any morally relevant sense. DishBrain used 2D cultures, not 3D organoids, but it remains the most widely cited demonstration of goal-directed learning in vitro.
Brainoware: reservoir computing with an organoid (2023)
Cai and colleagues at Indiana University Bloomington published “Brain organoid reservoir computing for artificial intelligence” in Nature Electronics in 2023. They placed a brain organoid on a high-density MEA and used it as the reservoir in a reservoir computing system they called Brainoware. Inputs were converted into stimulation sequences, the organoid’s responses were recorded, and a simple trained readout layer mapped those responses onto answers.
The system was tested on tasks including recognizing Japanese vowel sounds from different speakers and predicting a nonlinear chaotic equation. The authors reported that accuracy improved with training and that the organoid’s response properties changed over the training period, suggesting a contribution from plasticity in the tissue. Read more in our article on reservoir computing with brain organoids.
The Neuroplatform: remote access to organoids (2024)
Jordan and colleagues at FinalSpark, a Swiss company, described in Frontiers in Artificial Intelligence (2024) an “open and remotely accessible Neuroplatform” for wetware computing research. Forebrain organoids on MEAs, kept alive by automated life-support systems, could be stimulated and recorded by researchers over the internet. The paper reported long-running organoid recordings sustained over months.
The significance was infrastructural: it showed that experiments with living neural tissue could be run as a remote service, with standardized hardware and software, much as cloud computing made servers accessible without owning a data center.
Summary of key milestones
| Year | Milestone | Publication |
|---|---|---|
| 2001 | Neuron culture controls a simulated body (“animat”) | DeMarse et al., Autonomous Robots |
| 2006 | Induced pluripotent stem cells from adult mouse cells | Takahashi & Yamanaka, Cell |
| 2013 | Cerebral organoids from human stem cells | Lancaster et al., Nature |
| 2019 | Oscillatory network activity in maturing cortical organoids | Trujillo et al., Cell Stem Cell |
| 2022 | DishBrain: cultured neurons learn to play Pong | Kagan et al., Neuron |
| 2023 | “Organoid intelligence” defined as a field; Baltimore declaration | Smirnova et al.; Hartung et al., Frontiers in Science |
| 2023 | Brainoware: organoid reservoir computing | Cai et al., Nature Electronics |
| 2024 | Remotely accessible organoid Neuroplatform | Jordan et al., Frontiers in Artificial Intelligence |
Why does organoid intelligence matter?
There are three main motivations, and they are worth separating because they operate on different timescales.
Energy and learning efficiency
The human brain operates on roughly 20 watts, a figure discussed by Balasubramanian in a 2021 PNAS perspective titled “Brain power”. Meanwhile, the International Energy Agency’s 2025 report Energy and AI estimated that data centers used around 415 terawatt-hours of electricity in 2024, about 1.5% of global electricity consumption, with AI a growing share of that demand.
Brains also learn from comparatively little data. A child does not need millions of labeled examples to recognize a new animal. If even a fraction of that efficiency could be captured in engineered biological systems, the payoff would be large. The honest caveat is that no organoid system today comes close to replacing a GPU, and life support carries its own energy costs. We examine this argument carefully, with numbers, in biocomputing vs silicon.
A new window on the human brain
OI platforms are also tools for neuroscience and medicine. Organoids can be grown from the cells of people with neurological or developmental conditions, and learning experiments might reveal differences in how their networks process information. Lancaster’s 2013 paper, for example, used patient-derived organoids to model a form of microcephaly. Measuring learning and memory in a dish could add a functional dimension to such disease models and to drug and toxicity testing.
Reducing animal research
Smirnova and colleagues highlighted that human-derived organoid models could complement, and in some cases reduce, the use of animals in research, particularly in neurotoxicology. Human tissue may also capture features of human biology that animal models miss.
What are the current limits of organoid intelligence?
It is important to be clear-eyed. OI is at a very early stage, and the obstacles are substantial.
- Scale. Organoids contain many orders of magnitude fewer neurons than a human brain. Growing larger organoids runs into the next problem.
- No blood supply. Without vessels, oxygen and nutrients must diffuse in from the surface. Larger organoids often develop a dead (necrotic) core. Organoid vascularization is an active research area.
- Variability. Organoids grown with the same protocol can differ in size, cell composition and activity. Reproducible experiments require careful quality control and large sample sizes.
- Lifespan and maturation. Organoids mature slowly and model early developmental stages more closely than the adult brain. Long-term culture is labor-intensive.
- Interface bandwidth. Flat MEAs touch only the bottom surface of a 3D organoid. Recording from and stimulating the interior requires new 3D or flexible electrode designs.
- Speed and reliability. Neurons signal on millisecond timescales, far slower than transistors switch, and their responses are noisy. Learning effects reported so far are modest and task-specific.
- Standards and reproducibility. Labs use different hardware, protocols, and analysis pipelines, making results hard to compare. Shared data formats and detailed provenance records help.
None of these limits is necessarily permanent, but each requires years of engineering and biology. The future of biocomputing article discusses the roadmap for addressing them.
Is organoid intelligence ethical?
Because OI involves human neural tissue that may become more complex over time, ethics is central to the field. The 2023 OI paper made “embedded ethics” a pillar of its proposal, and the Baltimore declaration emphasized responsible exploration.
Consent and donor rights
Every iPSC line comes from a person. Donors need to understand, and agree to, the possibility that their cells could be used to build computing systems, not only for disease research. Questions also arise about commercial use, data privacy (cells carry the donor’s genome), and whether donors can withdraw consent. Hyun, Scharf-Deering and Lunshof reviewed these issues in Brain Research (2020).
Could organoids become conscious?
This is the question people ask most. There is no scientific evidence that current organoids are conscious or sentient. A 2021 report by the U.S. National Academies of Sciences, Engineering, and Medicine judged that present-day neural organoids are very far from the complexity associated with consciousness, while recommending continued attention as the technology advances. The difficulty is that there is no agreed scientific test for consciousness, which is why many researchers favor monitoring, transparency and precautionary limits.
Responsible research practices
Common recommendations include clear consent processes, oversight committees with outside members, public engagement, and careful language that avoids overstating what organoids can do. For a full treatment, see our article on the ethics of organoid intelligence and our commitments on the ethics section of our homepage.
What is the outlook for organoid intelligence?
Progress in OI is likely to come in layers rather than a single breakthrough.
Near term: better tools and reproducible experiments
The immediate work is in infrastructure. That means more reliable organoids, higher-density and 3D interfaces, automated life support, standardized closed-loop software, and open, well-annotated datasets. Remote platforms like the Neuroplatform suggest that shared access to living tissue could speed up research, allowing groups without wet labs to run experiments.
Medium term: more complex tissue
Researchers are working on vascularized organoids and on assembloids, which combine organoids representing different brain regions so that they connect to each other. Such structures could provide more realistic circuits, longer lifespans, and possibly richer computational behavior.
Long term: hybrid biological and digital systems
The most plausible future is not organoids replacing chips, but hybrid systems in which biological tissue handles specific tasks it may be well suited to, such as adaptive learning from sparse data, alongside conventional hardware for everything else. Whether that vision becomes practical depends on solving the limits above, and on public trust built through transparent, ethical practice.
How you can follow or join the field
Organoid intelligence sits at the junction of stem cell biology, electrophysiology, machine learning and ethics. If you are coming from software, the most accessible entry points are data analysis (spike sorting, decoding), closed-loop control, and experiment design. Making these experiments easier to program, run and reproduce is the problem we are working on at Organoid Programming, through a protocol SDK, a closed-loop runtime and open data tooling described on our platform page. If you want to try it as it develops, you can request early access.
Organoid intelligence is still young. What it has already shown is that living human neurons, given the right interface and feedback, can be coupled to digital systems in ways that change their behavior. What comes next depends on careful science, honest reporting of limits, and thoughtful attention to the ethical questions that come with computing with human cells.
Frequently asked questions
What is organoid intelligence in simple terms?
Organoid intelligence (OI) is a research field that tries to use small, lab-grown clusters of human brain cells, called brain organoids, as living computing hardware. The organoids sit on electrode arrays that can stimulate them and record their electrical activity, and software translates between digital data and neural signals so the tissue can process information and, ideally, learn.
Who coined the term organoid intelligence?
The term was set out in a 2023 paper in Frontiers in Science by Lena Smirnova, Thomas Hartung and colleagues at Johns Hopkins University and collaborating institutions. Titled ‘Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish’, it proposed OI as a multidisciplinary field and described a research program combining organoids, interfaces, machine learning and embedded ethics.
Is organoid intelligence the same as artificial intelligence?
No. Artificial intelligence usually means software running on silicon chips, such as neural networks trained on GPUs. Organoid intelligence uses real, living human neurons as the substrate that does the computing. The two are complementary: OI systems rely heavily on conventional machine learning to encode inputs into stimulation patterns and to decode the neurons’ responses into useful outputs.
Can organoid intelligence beat computers today?
No. Current organoid systems perform simple tasks, such as reacting to a simulated Pong game or classifying a small set of speech sounds, and they are slow, variable and hard to keep alive. The interest lies in potential long-term advantages in energy use and learning efficiency, not in present-day performance. Conventional chips remain far better for almost every practical workload.
Are brain organoids conscious?
There is no evidence that current brain organoids are conscious. They are tiny, lack sensory input and blood vessels, and are far simpler than a human brain. A 2021 U.S. National Academies report judged present-day organoids to be very far from the complexity associated with consciousness. Researchers still recommend monitoring the question as organoids grow larger and more complex.
What was DishBrain?
DishBrain was a 2022 experiment by Brett Kagan and colleagues at Cortical Labs, published in Neuron. Cultured human and mouse neurons grown on a high-density microelectrode array were connected to a simulated game of Pong. Electrical feedback told the cells where the ball was and whether they missed, and the cultures improved their rally performance within minutes.
What do you need to build an organoid intelligence system?
A working OI setup combines four parts: brain organoids grown from stem cells, a microelectrode array to stimulate and record them, a closed-loop control system that delivers feedback in real time, and machine learning models that encode inputs and decode outputs. Around these sit life support (incubation, media exchange) and data infrastructure for storing and sharing recordings.
What are the biggest limitations of organoid intelligence?
The main limits are scale (organoids contain far fewer neurons than a brain), lifespan and maturation (no blood vessels, so cores can die), variability between batches, low-bandwidth interfaces that touch only the tissue surface, and slow, noisy learning. There are also unresolved ethical questions about consent, donor rights and possible future sentience.
Sources
- Smirnova et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science.
- Hartung et al. (2023). The Baltimore declaration toward the exploration of organoid intelligence. Frontiers in Science.
- Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron.
- Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.
- Jordan et al. (2024). Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence.
- Kagan et al. (2023). The technology, opportunities, and challenges of Synthetic Biological Intelligence. Biotechnology Advances.
- Lancaster et al. (2013). Cerebral organoids model human brain development and microcephaly. Nature.
- Takahashi & Yamanaka (2006). Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. Cell.
- Trujillo et al. (2019). Complex oscillatory waves emerging from cortical organoids model early human brain network development. Cell Stem Cell.
- DeMarse et al. (2001). The neurally controlled animat: biological brains acting with simulated bodies. Autonomous Robots.
- Mead (1990). Neuromorphic electronic systems. Proceedings of the IEEE.
- Merolla et al. (2014). A million spiking-neuron integrated circuit with a scalable communication network and interface. Science.
- Azevedo et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology.
- Balasubramanian (2021). Brain power. PNAS.
- IEA (2025). Energy and AI.
- Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
- National Academies of Sciences, Engineering, and Medicine (2021). The Emerging Field of Human Neural Organoids, Transplants, and Chimeras.
- Hyun, Scharf-Deering & Lunshof (2020). Ethical issues related to brain organoid research. Brain Research.
