Zhang, 76, has not moved his hands in years. Last spring, he picked up a pen — and wrote eight characters without lifting a finger.
The characters spelled out "浙江大学脑机接口" (Zhejiang University brain-computer interface). Zhang did not move a muscle to produce them. He imagined writing, and a robotic arm holding a pen traced each stroke on a board. Behind that quiet act is one of China's longest-running invasive brain-computer interface (脑机接口) cases, run by Zhang Jianmin (张建民) at the Second Affiliated Hospital of Zhejiang University School of Medicine.
A volunteer who refused to disappear
Zhang became the country's first volunteer for invasive BCI clinical translation back in 2020. A car accident had crushed his cervical spinal cord, leaving him with high cervical quadriplegia — what doctors call 高位截瘫. His mind stayed sharp; his body did not.
The team, working with the Zhejiang University Qiushi Institute for Advanced Study, implanted microelectrode arrays into his motor cortex (运动皮层) — the brain region that plans movement. A surgical robot placed the arrays with sub-millimeter precision, landing them in the fifth cortical layer where the relevant neurons sit. From there, the work was about translation: turning the faint electrical chatter of intent into commands a robotic hand could obey.
In the early years the wins were coarse but life-changing:
- He could drink water by willing a robotic arm to bring a cup to his mouth.
- He could eat — the team's famous demo was "thinking" a youtiao (fried dough stick) into reach.
- He could shake hands and, later, play digital mahjong by imagining clicks to "play" a tile against an AI.
For a man who told researchers that joining the study proved "I am still a useful person," the bar was not a publication. It was agency.
From big joints to fine strokes
Most invasive BCIs stop at what researchers call large-joint motion — grasp, reach, hold. Writing Chinese characters is harder by an order of magnitude. A character is not one motion; it is a sequence of strokes, each with its own shape, direction, and timing, built from a vocabulary of radicals and structures.
That is the breakthrough the team reported. Starting in the summer of 2022, Zhang Jianmin's group set out to decode the brain activity behind imagined handwriting. By 2023 they had a real-time system where Zhang's imagined strokes drove a robotic arm. In April 2024 they demonstrated it publicly, writing those eight characters.
The number that matters: in real-time tests, the system reached 91.3% accuracy across 100 common Chinese characters. Qi Yu (祁玉), a researcher at Zhejiang University's Frontier Science Center for Brain and Brain-Computer Fusion, noted that handwriting differs from large-joint motion because the brain's expression switches patterns depending on the local structure and movement state of each stroke. The team built a dynamic decoding model to capture that switching.
Zhu Junming (朱君明), who leads functional neurosurgery at the hospital, framed the stakes plainly: for patients who have lost movement and speech — ALS patients, for instance — being able to extract intention from brain signals and turn it into written text is a way to communicate at all.
Why Chinese characters are a useful test
There is a subtle point worth pulling out. Decoding English letters or a QWERTY key press is one thing; decoding the continuous, curved trajectory of a handwritten character is a harder signal-processing problem. If a BCI can recover that trajectory from motor-cortex activity, it is learning the fine structure of intent, not just a binary "move or don't."
That makes the Hangzhou work a useful marker for the field: it is evidence that invasive BCIs are moving from "open the hand" toward "express the thought."
Honest limitations
- A single volunteer. Everything here rests on one long-term case. That is invaluable depth, but it is not statistical evidence.
- Hospital-bound hardware. The invasive system ran inside the hospital research ward; Zhang spent years living there. There is no demonstration of a portable, at-home, long-term implant like some minimally invasive routes aim for.
- Accuracy is task-specific. 91.3% on 100 common characters is strong, but it is not open-ended, free handwriting across the full character set.
- No independent replication. Results come from the Zhejiang team and affiliated media; we have not seen a separate lab reproduce the decoder on another patient.
What readers can do now
- Read this as a fine-motor milestone, not a cure. Writing characters by thought shows the ceiling rising, especially for communication in locked-in patients.
- Watch for a second patient. The single-volunteer ceiling lifts only when a second, third, and tenth implant show the same decoding holds.
- Compare routes honestly. Invasive BCIs like this chase high bandwidth; minimally invasive and non-invasive approaches trade bandwidth for simpler, safer deployment. Different patients need different trade-offs.