Human Neurons Played Doom—but They Did Not Run the Game
As of August 2026, Cortical Labs’ Doom project is best understood as a hybrid biological-computing demonstration, not a computer game running inside a miniature human brain. Living neurons influenced movement and shooting through a closed feedback loop, while conventional software ran Doom, translated game events into electrical stimulation and converted recorded neural activity into commands.
The meaningful update is that the experiment is no longer only an isolated laboratory spectacle. Cortical Labs says its Cortical Cloud is now available, giving developers remote access to CL1 systems through browser-based tools and a Python SDK. That makes the Doom demonstration a public example of a programmable research platform, although it does not turn the still-experimental technology into a general-purpose replacement for silicon computers.
What the neurons actually controlled
The culture did not receive ordinary video frames, understand demons as visual objects or execute Doom’s program code. Software reduced selected parts of the game state—such as the position of an enemy or the player’s surroundings—to patterns of electrical stimulation delivered through electrodes beneath the cells.
The same interface recorded the neurons’ electrical spikes. A decoder associated patterns in that activity with a limited set of permitted actions, including turning, moving and firing. The visible player character therefore reflected output from a chain containing the game engine, an encoder, living cells, a decoder and conventional control software.
This distinction matters because saying that neurons “ran Doom” assigns the whole result to one component. The more accurate claim is that a culture of human-derived neurons participated in a real-time control loop connected to Doom. The achievement lies in maintaining that loop and observing adaptive behaviour, not in replacing the processor or graphics hardware responsible for the game itself.
Why Doom was a harder demonstration than Pong
Cortical Labs had already used a related closed-loop system for Pong, where the environment can be compressed into the position of a ball and the vertical movement of a paddle. Doom adds corridors, orientation, enemies and several possible actions, forcing the interface designers to decide which information the culture receives and how its activity affects play.
The early-2026 demonstration involved roughly 200,000 living human neurons on a silicon interface. Scientific American’s account of the experiment reports that independent researcher Sean Cole paired the cells with a standard learning algorithm and that the hybrid system outperformed the algorithm operating alone. That comparison indicates a contribution from the biological network, but it does not establish that the cells independently learned every part of the task.
The earlier Pong work provides the stronger scientific foundation. Published in 2022, it described neuronal cultures embedded in a simulated game world through electrical stimulation and recording, with changes in goal-directed activity appearing during play. Doom extends the same broad idea into a richer environment, but the publicly described Doom result remains a demonstration rather than a peer-reviewed replication of all the claims associated with the video.
Learning here does not mean human understanding
The cells can alter their firing patterns in response to structured input and feedback. In this context, learning means a measurable change in activity or performance across interaction with the environment; it does not mean that the culture recognizes Doom, understands a weapon or experiences winning and losing as a person would.
The neurons were derived from cells reprogrammed into stem cells rather than removed as an intact piece of a person’s brain. They form a cultured network on an electrode array, without the sensory organs, anatomy or organized brain regions that support ordinary human perception. Terms such as “brain cells” are biologically accurate, but “a brain on a chip” can create a misleading picture of the system’s organization and capabilities.
Performance also remained modest. The character could wander into walls, fire inaccurately and require several attempts to target an enemy. Later reporting noted both improvement in targeting and an important engineering constraint: the cultures were not yet producing consistent, programmable results, and their operating lifespan was about six months.
What CL1 contributes beyond the game
CL1 packages the elements needed to keep neural cultures functioning and communicate with them: a nutrient-supported environment, a silicon electrode interface, recording and stimulation electronics, and software for running closed-loop experiments. Cortical Labs describes the device as self-contained, with onboard life support and real-time interaction between neurons and software.
For researchers, the practical attraction is not superior Doom performance. A controllable environment can expose living neural networks to repeatable tasks while investigators record how activity changes, potentially adding functional behaviour to studies of disease models or drug responses. Remote access also lowers the infrastructure barrier for software researchers who do not operate a cell-culture laboratory.
The computing case is more speculative. Biological networks may prove useful where rapid adaptation, low-data learning or interaction with changing environments matters, but the Doom video supplies no general benchmark for energy consumption, reliability or performance against modern AI systems. It therefore cannot support claims that wetware is already faster, cheaper or more capable than conventional hardware.
The real milestone is the interface
Doom makes the work easy to see because every neural output becomes an action on screen. The deeper result is that a developer could connect a complex software environment to living neurons, stimulate them according to changing conditions and read their activity quickly enough to influence that environment in return.
That is a narrower claim than “human brain cells became gamers,” but it is also more consequential. The CL1 demonstration shows a programmable bridge between biological activity and ordinary software; it does not show consciousness, human-like understanding or a standalone biological PC. Its next test is whether that bridge can deliver repeatable scientific measurements and useful control tasks after the novelty of watching cells stumble through Doom has worn off.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.