200,000 human neurons learn Doom on CL1 chip

Researchers at Cortical Labs connected a culture of 200,000 human neurons to a CL1 silicon interface and trained the cells to navigate corridors and fire in the video game Doom. The experiment was conducted in a Melbourne laboratory and the results were published on June 11, 2026.

The neurons were grown from blood‑derived stem cells and attached to a multielectrode array on the CL1 chip. The system converted visual events from the game into patterned electrical stimulation across the electrode array and recorded neural spiking activity. Recorded spikes were mapped to discrete game commands such as move, turn and shoot, creating a real‑time feedback loop between the game and the neural culture.

Training began with simple, Pong‑style input–output pairings and progressed to Doom’s three‑dimensional navigation and targeting. The research team reported frequent misfires and overcorrections during play. Over repeated training sessions the culture’s outputs became more goal‑directed, and performance showed measurable improvement while remaining imprecise.

At the technical level, the CL1 functions as a bridge between the digital environment and living tissue. The chip delivers stimulation patterns tied to game states and reads multiunit activity to select actions in real time. Typical CL1 cultures currently maintain active spiking patterns for roughly six months before activity declines.

Energy efficiency is a stated reason for the project. The human brain operates at about 20 watts, and Cortical Labs has described that figure as motivation to explore whether living neural tissue can perform local learning or control tasks with lower energy use than large, conventional AI training systems. The company frames biological co‑processors as potential complements to, rather than replacements for, GPUs and data center models.

The team and other researchers propose several research and commercial uses for neuron‑chip systems, including patient‑specific drug screening on living neural tissue, disease modeling, and adaptive control systems for robotics. Current limitations cited by the researchers include short culture lifespan, variability in outputs across preparations, and the absence of standardized, programmable interfaces that would be required for scaling and deployment.

The researchers noted that regulatory agencies such as the U.S. Food and Drug Administration and the National Institutes of Health could establish standards for safety, reporting and the use of human tissue if neuron‑chip systems move into clinical or patient‑derived testing. The paper lists reproducibility and tighter control of culture conditions as technical priorities before broader adoption.

The Cortical Labs report provides a documented instance of living human neurons trained to act on a complex digital task under laboratory conditions. The authors identify extending culture longevity, improving consistency and developing standard interfaces as the next technical challenges to address before evaluating biological computing alongside conventional AI for specific real‑world applications.