In October 2022, Cortical Labs, a Melbourne-based biotech company, published a paper in the journal Neuron showing that 800,000 human and mouse neurons grown on a multi-electrode array had learned to play Pong in approximately five minutes. The neurons received electrical signals encoding the ball’s position, fired activity back to control the paddle, and improved their performance over time through a feedback loop designed around the brain’s natural drive to minimize prediction error. Nothing in computing or biology had demonstrated this combination of speed, efficiency, and biological learning in a single system before.
The neuroscience and AI communities reacted with a mix of fascination and philosophical unease, because DishBrain did not fit neatly into any existing category. It was not a brain. It was not a computer. It was not conscious in any meaningful sense the researchers would claim. Yet it played a game, learned from its mistakes, and did so using a fraction of the energy consumed by digital AI systems. The implications reach across medicine, computing, and the oldest questions about what it means to be sentient. Here is a clear account of what DishBrain actually was, what it did, and why it matters.
What the DishBrain Experiment Actually Was
DishBrain was a system developed by Brett Kagan and his team at Cortical Labs in which living neurons were grown directly onto a multi-electrode array (MEA), a flat chip embedded with 59 electrodes capable of both sending and receiving electrical signals to individual neurons. The neurons were a combination of human cortical neurons derived from induced pluripotent stem cells and mouse cortical neurons. Together they numbered approximately 800,000 cells, enough to form functional synaptic connections and exhibit spontaneous electrical activity within days of being placed on the chip.
The MEA served as the interface between the biological neurons and the Pong game simulation running on connected hardware. The chip was divided into zones: electrodes in the left half of the array corresponded to the ball appearing on the left side of the screen, electrodes on the right corresponded to the ball appearing on the right. When the simulated ball was in motion, the corresponding electrodes fired electrical pulses into the neuron culture. The neurons’ own electrical activity was then read back through the same electrodes and translated into paddle movement commands. If the neurons fired more strongly on the left, the paddle moved left; more activity on the right moved the paddle right.
No reinforcement learning algorithm, no loss function, no gradient descent was involved. The feedback mechanism was purely biological in its logic: when neurons successfully returned a volley, their electrical environment remained relatively stable. When they missed, the system delivered unpredictable, random electrical stimulation, creating the neural equivalent of “surprise.” The neurons, following their natural behavior of minimizing unexpected input, learned to keep the ball in play.
How the Neurons Were Taught: The Free Energy Principle in Action
The theoretical foundation behind DishBrain is the Free Energy Principle, developed by neuroscientist Karl Friston at University College London. In simplified terms, this principle holds that all biological neural systems are fundamentally driven to minimize the difference between their predictions about the world and the sensory information they actually receive. Neurons are not passive receivers of input; they are constant predictors, and they generate activity to bring the world into alignment with their expectations. When a prediction fails, that mismatch, or prediction error, drives learning and behavioral adjustment.
The Cortical Labs team operationalized this principle directly. A correct paddle return kept the electrical environment around the neurons stable and predictable. A missed return triggered a burst of random electrical noise delivered through the MEA, which from the neurons’ perspective was a strong, unpredictable perturbation. The neurons did not “know” they were playing a game. They simply experienced the structured stimulation of a game environment and adjusted their activity to reduce disruption. The game became the feedback system; the neurons became the adaptive controller.
This is mechanistically different from how digital AI learns. Traditional reinforcement learning uses mathematical reward signals passed to an algorithm that adjusts numerical weights across layers of computation. DishBrain used no such algorithm. The neurons self-organized their responses through biological synaptic plasticity, the same mechanism through which human brains form memories and acquire skills. Learning was not programmed; it emerged.
Human Neurons vs Mouse Neurons: What the Data Showed
One of the more striking findings in the published data was the performance difference between human and mouse neurons. Human cortical neurons derived from stem cells consistently outperformed mouse neurons at the task, reaching competent paddle control faster and sustaining performance more reliably. The human neurons learned to play in approximately five minutes. Mouse neurons took longer and showed more variable performance across experiments.
The researchers were careful not to overinterpret this result as evidence of superior “intelligence” in human neurons in any meaningful sense. The difference more likely reflects the distinct electrophysiological properties and synaptic density characteristics of human cortical neurons compared to the mouse equivalent. Human cortical neurons have longer dendritic arbors, different ion channel distributions, and different timescales of synaptic integration that may simply be better suited to the specific timescales and feedback frequencies used in this experimental design.
What the result does confirm is that human neurons from stem cells can form functional, adaptive circuits on a chip and respond to structured electrical input in ways that influence external outputs. This matters enormously for drug testing and disease modeling, independent of any gaming or computing application.
Is DishBrain Conscious? The Scientific and Philosophical Answer
DishBrain is not conscious by any definition that neuroscientists or philosophers of mind have proposed. Consciousness in the scientific literature requires, at minimum, integrated information processing across a unified system, global workspace activity, or the presence of something like a unified model of self and world. The DishBrain neuron culture has none of these. It is a flat 2D array of cells with no sensory organs, no embodied interaction with a physical world, no sleep-wake cycles, no homeostatic regulatory systems, and no architecture that would support the type of recursive self-modeling associated with even basic animal awareness.
Brett Kagan explicitly described the neurons as exhibiting “sentience” in the paper’s title, which generated predictable media confusion. In the context the paper used, sentience referred narrowly to goal-directed adaptive behavior, the biological property of adjusting actions in response to feedback to achieve a locally stable state. It did not claim subjective experience, awareness, or anything resembling human consciousness. The word choice was provocative but technically defensible within a very specific philosophical framing.
Questions about consciousness and substrate are ancient. Discussions of whether a sufficiently complex simulation could host genuine experience connect to simulation theory and to what philosophers call the “hard problem of consciousness,” the question of why physical processes give rise to subjective experience at all. DishBrain does not resolve that question. But it does complicate simplistic answers, because it demonstrates that biological substrate and adaptive behavior can be partially separated from the full organismal context in which we usually observe them.
How DishBrain Compares to Traditional AI Learning
The comparison between DishBrain and digital AI is instructive precisely because DishBrain performed so poorly by conventional AI metrics and yet so impressively by biological ones. AlphaGo from DeepMind mastered Atari games after processing millions of simulated game frames using deep reinforcement learning running on clusters consuming megawatts of electricity. DishBrain neurons, using biological synaptic plasticity and approximately 0.1 watts of metabolic energy, reached functional competence in five minutes.
That energy comparison is not trivial. The entire field of AI is facing an energy scaling problem as models grow larger and more computationally intensive. Biological neurons perform computation at extraordinary energy efficiency because they use analog, spike-based processing rather than digital floating-point arithmetic. A human brain runs on roughly 20 watts. A large language model inference run can consume kilowatts per query at scale. If even a fraction of the brain’s computational efficiency could be harvested in biological or neuromorphic hardware, the implications for sustainable computing are transformative.
This does not mean DishBrain is about to replace silicon. It means that the gap between biological and artificial information processing is smaller at the substrate level than the current AI paradigm suggests, and that hybrid biological-digital systems may offer paths that neither pure biology nor pure silicon provides alone.
What DishBrain Means for Medicine and Neuroscience
The most immediate practical applications of DishBrain-style research are medical, not computational. Neurons grown on MEAs that can be tested for functional responses to electrical stimulation provide a platform for testing neurological drugs without full animal models. Current preclinical neurology research relies heavily on rodent models that differ substantially from human brain tissue in their pharmacological responses. Human-derived neurons on chips could provide a faster, cheaper, and more predictive testing environment for drugs targeting epilepsy, Parkinson’s disease, Alzheimer’s disease, and traumatic brain injury.
Horn et al. (2022) found in a Nutrients review that creatine supplementation after traumatic brain injury reduced post-concussion symptoms. The kind of biological computing substrate demonstrated in DishBrain could eventually allow researchers to model TBI damage at the cellular level, test neuroprotective interventions in human-derived tissue, and screen therapeutic candidates before ever entering an animal or human trial. This is not speculation; it is the stated direction of the broader field of organs-on-chips, in which DishBrain represents a high-profile and sophisticated entry.
Understanding the impact of environmental factors on human biology, from chemical exposures to genetic variants to nutritional deficiencies, is also highly relevant. Studies of extreme human biological adaptation, like the Bajau people’s genetic adaptations for underwater diving, demonstrate just how plastic human biology is across generations. MEA-based neuron cultures could eventually allow researchers to study whether specific genetic variants change neuronal electrical behavior in ways that explain cognitive differences between populations.
Organoid Intelligence: What Comes After DishBrain
In 2023, a team at Johns Hopkins University published a roadmap for what they termed “organoid intelligence” (OI), a field that proposes using three-dimensional brain organoids, miniature self-organizing neural structures grown from human stem cells in laboratory conditions, as biological computing substrates. This is related to but distinct from DishBrain’s 2D MEA approach. Brain organoids develop rudimentary cortical layering, synaptic architecture, and spontaneous network activity that more closely resembles an embryonic brain than a flat neuron culture does.
Cortical Labs received $600,000 AUD from the Australian government in 2023 to continue DishBrain development under its DishBrain 2.0 program, pursuing applications in biological computing and neuroscience research. The field is moving quickly from proof-of-concept toward applied research with defined goals in drug discovery, computational neuroscience, and potentially novel forms of biocomputing that leverage the energy efficiency of biological neural networks.
The societal dimension of this trajectory is genuinely complex. As populations age and birth rates fall globally, the demand for faster, cheaper drug discovery and better treatments for neurological disease accelerates. Examining population collapse trends alongside the timeline for neurological disease burden suggests these tools are arriving at a moment when medical systems will need them most. Organoid intelligence research could meaningfully shorten the drug development pipeline for neurological conditions that currently have no effective treatment.
Can brain cells in a dish actually think?
Brain cells in a dish exhibit adaptive, goal-directed behavior in response to structured electrical feedback, as DishBrain demonstrated, but they do not think in any meaningful cognitive sense. Thinking requires integration across large-scale neural networks, embodied sensorimotor feedback, memory systems, and likely other properties that a flat 2D neuron culture lacks entirely. DishBrain demonstrated learning and adaptation, not cognition or awareness.
How does DishBrain differ from artificial intelligence?
DishBrain uses living biological neurons that learn through synaptic plasticity, consuming approximately 0.1 watts. Artificial intelligence uses mathematical algorithms running on silicon hardware consuming megawatts at scale. DishBrain learned Pong in five minutes without a reward function or gradient descent. Traditional AI requires millions of training examples and vast computational resources. The two systems are not comparable in architecture, scale, or capability.
Is DishBrain research ethical?
The neurons used in DishBrain were derived from stem cells, not extracted from living humans or fetuses. The research operated under standard institutional ethics oversight. The more active ethical debate concerns future research directions: if brain organoids become sufficiently complex to show signs of distress or awareness, what obligations do researchers have? Current organoids and MEA cultures are far below any threshold that neuroscientists associate with awareness or suffering.
Who funded and published the DishBrain research?
The original DishBrain research was conducted at Cortical Labs in Melbourne, Australia, led by Brett Kagan as chief scientific officer. The paper “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world” was published in the journal Neuron in October 2022. Funding came from a mix of private investment in Cortical Labs and Australian government research grants, with DishBrain 2.0 receiving $600,000 AUD in 2023.
What is organoid intelligence and how does it relate to DishBrain?
Organoid intelligence refers to the use of three-dimensional human brain organoids as biological computing substrates, a concept formalized by a Johns Hopkins University team in a 2023 roadmap paper. DishBrain used a 2D neuron culture on a flat chip. Organoids are more architecturally complex, developing cortical layers and richer synaptic networks. Both approaches aim to leverage biological neural plasticity for computing and medical research, but they are at different stages of development and use different biological substrates.
DishBrain represents the earliest credible demonstration that biological neurons can be placed in an engineered feedback environment and show adaptive learning without any programming. If you are following the boundary between biology and computing, this is where the most unexpected developments are happening. The next five years of organoid intelligence research will likely produce results that challenge current assumptions about both artificial intelligence and neuroscience in ways that the DishBrain paper only hinted at.