Cortical Labs, an Australian biotechnology startup, connected a culture of living human and mouse neurons to a version of the classic first-person shooter Doom, claiming the cells learned to respond to in-game stimuli. The demonstration builds on the company’s earlier, peer-reviewed work in which biological neurons played the simpler game Pong, a study that itself drew on more than a decade of research into training living brain tissue through electrical feedback. What makes the Doom announcement significant, and unsettling to some, is not just the leap in game complexity but the questions it forces about where biological computing is headed and what ethical guardrails should follow.
From Pong to Doom: How Living Neurons Learned to Play
The scientific foundation for this work was laid in a 2022 paper published in the journal Neuron by Cell Press. In that study, neurons grown on multi-electrode arrays received electrical signals representing a Pong ball’s position and returned signals that moved the in-game paddle. The researchers reported that the network displayed learning effects including short-timescale improvement, with the cultured cells adapting their firing patterns within minutes of gameplay; these findings are detailed in the underlying Neuron article. Control conditions, in which feedback was randomized rather than tied to game performance, showed no comparable adaptation, strengthening the case that the neurons were genuinely responding to structured input rather than merely firing at random in response to stimulation.
The Doom project extends that closed-loop concept into a far more demanding environment. Instead of tracking a single ball along one axis, the neural culture had to process a three-dimensional maze populated by enemies, with inputs and outputs mapped to sensory cues and firing patterns in ways Cortical Labs has outlined in public talks. As of late 2022, however, no peer-reviewed paper specific to the Doom experiment has been published, which marks an important distinction from the Pong results that passed formal review and included documented control conditions. Until independent replication or a reviewed manuscript appears, the Doom claims rest largely on company announcements, and readers should weigh the two milestones differently even as they recognize that both grow from the same basic experimental pipeline.
A Longer History Than Most Coverage Admits
Press coverage of DishBrain often frames the work as unprecedented, but the idea of training living cortical tissue through electrical feedback dates back well over a decade. A 2008 study in the Journal of Neural Engineering demonstrated that carefully structured stimulation could guide dissociated rat cortical neurons toward specific activity patterns during a goal-directed task, with the authors showing that spatio-temporal electrical stimuli could shape behavior in an embodied neural network. That experiment used multi-electrode arrays to deliver time-varying inputs and read out outputs, much like DishBrain does, and it established that even relatively small cultures could be nudged toward desired responses when the environment provided consistent feedback.
What changed between 2008 and 2022 was scale, speed, and ambition. Later work drew not only on rodent tissue but also on human neurons, implemented more sophisticated feedback rules, and embedded the network inside a familiar consumer game environment. Earlier cortical-network research, such as experiments that paired cultured tissue with robotic platforms to study adaptive control, had already proven the principle that living neurons could be guided toward goal-directed states; one open-access overview of these approaches in neurally controlled systems underscores how long the field has been refining these tools. DishBrain and its successors packaged that principle inside cultural touchstones like Pong and Doom, making incremental neuroscience advances far more visible to the public than previous generations of in vitro work.
What “Sentience” Means Here, and What It Does Not
The word “sentience” in the original Neuron paper’s title has generated more confusion than any single data point in the study. The authors used the term in a narrow sense, arguing that the neurons exhibited sensitivity to environmental stimuli and adjusted their behavior accordingly when embodied in a game world. The peer-reviewed record of the DishBrain study confirms the paper’s title, methods, and scope but does not present evidence that the neurons experience anything like human consciousness or self-awareness. In this technical framing, “sentience” refers to the capacity to integrate incoming signals and produce adaptive responses, not to feelings, introspection, or a first-person point of view.
This gap between technical and popular language creates real risks. If a dish of neurons can be called “sentient” in a journal article, the term will inevitably be stretched further in press releases, social media posts, and funding pitches, where nuance tends to evaporate. Critics of the DishBrain framing argue that the observed behavior (better paddle tracking over short intervals) is more accurately described as stimulus-response plasticity or online learning, concepts that are familiar from decades of synaptic physiology and computational neuroscience. How regulators, ethicists, and the broader public interpret the word will shape policy on biological computing for years, and there is a danger that “sentience” becomes a marketing label rather than a carefully defined scientific claim, either inflating fears about lab-grown suffering or downplaying genuine moral concerns as the systems grow more complex.
Why the Doom Leap Raises Harder Questions
Pong is a two-dimensional game with a single moving object and one degree of freedom for the player; even a relatively small neural culture can, in principle, learn to map simple input streams to binary outputs that move a paddle left or right. Doom, even in a heavily simplified form, involves spatial navigation, threat identification, and timing, all processed simultaneously, and Cortical Labs has suggested that its cultures can adapt to these richer dynamics. If future experiments confirm that living networks can handle this level of complexity, the logical next step is not just another video game but real-world applications such as drug screening platforms, adaptive controllers for prosthetics, or hybrid processors that combine silicon’s speed with biological plasticity. The original Neuron publication, accessible via its digital object identifier, includes a data and code availability statement, signaling an intent to let other labs build on the work and potentially extend it into such domains.
The ethical pressure intensifies with each increase in task complexity. A neural culture playing Pong can be dismissed as an interesting curiosity, but a culture navigating a hostile virtual environment starts to raise questions about whether the tissue undergoes something analogous to stress, reward, or discomfort when it “dies” in-game or successfully avoids threats. No existing regulatory framework specifically governs experiments on disembodied neural tissue that interacts with software in real time; the closest analogs are animal-research protocols and human-tissue ethics boards, both designed for contexts that look very different from a dish of neurons wired into a digital arena. Until governance catches up, the field operates in a gray zone where scientific incentives push toward ever more dramatic demonstrations, while ethical guidance remains patchy and often reactive rather than proactive.
Designing Guardrails for Biological Computing
One starting point for guardrails is to examine how related areas of neuroscience have handled similar dilemmas. Work on brain–computer interfaces, for example, has long relied on multi-electrode arrays to decode neural signals and deliver feedback, and an overview of adaptive stimulation strategies highlights how closed-loop designs can both improve performance and complicate ethical analysis. As algorithms tune stimulation in response to neural activity, it becomes harder to predict exactly what patterns of input a brain, or a dish of neurons, will receive over time, raising questions about unintended side effects and long-term changes. Translating those concerns to DishBrain-like systems suggests that protocols should limit task difficulty, exposure duration, and the intensity of negative feedback, at least until researchers have better tools to assess whether the tissue exhibits markers of distress.
Another lesson comes from earlier in vitro experiments that coupled neural cultures to robotic bodies or simulated environments. A 2008 investigation into embodied cortical networks showed that even simple feedback rules could lead to emergent behaviors when neurons controlled a virtual agent, underscoring how quickly complexity can arise from basic components. As Doom-like tasks push that complexity further, oversight mechanisms may need to borrow from both animal-welfare frameworks (focusing on potential suffering) and data-protection regimes, which govern how information about human-derived tissue is handled. Concrete steps could include standardized reporting of culture composition and stimulation parameters, independent ethics review for any experiment involving aversive virtual environments, and clear criteria for when a dish should be retired rather than continually retrained. Without such measures, the excitement around biological computing risks outpacing the careful reflection that its strange new hybrids of tissue and code demand.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
