The bionic hand that works 83 percent of the time
A neural interface let a man feel his wife's hand again. The same system lets go of what he is holding roughly a dozen times a day. Both of those facts come out of a single number.
At sixteen, Avi Davidson fell thirty-five feet in an electrical accident. He survived with a spinal cord injury and the loss of his left hand. Years later, working as a therapist in Florida, he was referred by a biomedical engineer at the University of South Florida to a laboratory in Utah that had been looking for someone exactly like him.
The Utah NeuroRobotics Lab at the University of Utah is a joint venture between the College of Engineering and the College of Medicine, led by Jacob George. In early 2026, surgeons implanted small interfaces directly onto the nerves and muscles remaining in Davidson's left arm. Those interfaces connect to a LUKE arm, a commercially produced bionic limb. After months of training with the lab and an occupational therapist, he took it home in July.
What he wanted from it was not what an engineer might guess.
This will actually literally be a part of me that I will feel, and I can hold her hand. It brings tremendous joy to feel her hand in mine as something I thought I'd never experience.
A loop, not a limb
Most people picture a powered prosthesis as a one-way device: you think, it moves. This one runs in both directions, and the return direction is the harder half.
Going out, the path is intent, nerve, electrode, decoder, motor. Davidson imagines a movement; the implanted interfaces record the electrical activity that intention produces; software works out which movement he meant; the hand executes it. Coming back, the path reverses. Sensors in the fingertips detect contact and trigger stimulation through the same nerves, so touching something registers as touch rather than as a picture of touching something.
That return path is why the goal he named first was holding a hand rather than picking up an object. Without feedback, a user has to watch the prosthesis to know what it is doing. Vision is slow, it occupies attention, and it fails the moment the object is behind something. Sensory feedback closes the loop.
It is not closed evenly. Davidson reports his thumb as accurate and his middle finger as “zappy” — pins and needles rather than contact. The sensation is real, and it is not yet uniform.
The decoder is guessing
The software does not read his mind. It classifies. It looks at a pattern of neural activity and decides which of a small number of trained movements that pattern most resembles — rest, a power grasp, a pinch, an open hand. It is a statistical model making a best guess, hundreds of times a day.
Which means it is sometimes wrong. Engineers record exactly how wrong in a table called a confusion matrix: every intended movement in a row, every executed movement in a column, and the count of how often each pairing happened. The diagonal is the successes. Everything off the diagonal is an error.
Why 83 percent is the wrong number to quote
A note on what follows: the trial has not published a public error table. The figures below are a realistic teaching model built to show how this arithmetic works, not measurements from Davidson's device. The distinction matters, and it is one the classroom packet makes to students explicitly.
Suppose a decoder is tested on 200 attempted movements and gets 166 of them right. That is 83 percent accuracy. In a design review, 83 percent sounds like a working system.
Now ask a different question: which errors actually cost the user something? Reading a pinch as rest is an annoyance — nothing happens, he tries again. Reading a holding grasp as open is a dropped cup. Those are not the same event, but overall accuracy weights them identically.
Take the release errors alone. If two percent of power grasps and six percent of pinches get read as an open hand, and a person makes something like three hundred holding grasps in a day, the arithmetic gives roughly twelve unintended releases per day. Eighty-four a week. An object hitting the floor every waking hour or two.
Eighty-three percent accurate and drops something twelve times a day are the same sentence about the same device. One of them would pass a review. The other tells you what it is like to live with. A percentage is not an experience until you multiply it by how often the thing happens.
The weight you cannot argue with
The other limitation Davidson describes is blunt: the hand is heavy, and the weight makes fine control harder. That sounds like a comfort complaint. It is a lever problem.
A hand's weight acts at its centre of mass, roughly a third of a metre from the elbow. The biceps tendon inserts about five centimetres from that same joint — a very short lever arm on the other side. So the muscle is not simply carrying the extra mass; it is fighting a moment that has been multiplied by the ratio of those two distances.
Run the standard textbook numbers and a hand under a kilogram heavier than a biological one adds on the order of fifty-seven newtons of continuous tension at the tendon — close to six kilograms of extra pull, held for as long as the arm is up. That is the difference between “a bit heavy” and “I cannot use it for long.”
And there is no free fix. Cut the mass and something goes with it: battery life, grip strength, durable gearing, or the fingertip sensors that made the feedback loop possible in the first place.
What happens when the money stops
The trial was planned to run for a year, with Davidson keeping the hand afterwards. Then the federal research funding supporting it was reduced. Eight additional amputees who were to receive the same system will not now be enrolled, and Davidson may be required to return a device that is surgically wired into his nerves.
This has a name in research ethics: post-trial access — what a research team owes a participant once the study ends. It is genuinely unsettled. The arguments on each side are real ones. The device was built with public money and is, formally, a research instrument; without the team maintaining it, an unsupported implant may not stay safe. Against that: he accepted surgical risk on a stated understanding, and what was implanted is now integrated with his nervous system, which makes “returning equipment” a strange description of what would happen.
The engineering questions in this story all have answers. This one does not, and students notice the difference immediately.
Bionic Hands & Neural Interfaces
The full classroom version of this story, built for grades 10–12, dual enrollment, adult education, and introductory college courses. Print-ready PDF, no prep required.
- Signal-loop diagram students label themselves, with a word bank and two deliberate distractors
- Confusion matrix audit — accuracy, precision and recall worked from a two-way frequency table
- Biomechanics extension: the moment about the elbow, then the inverse design problem
- Structured deliberation on post-trial access, with a four-point writing rubric
- Teacher guide with worked answer key, minute-by-minute timing, and standards alignment
Left: the classroom packet on Teachers Pay Teachers. Right: a beginner electronics kit with the servos, sensors and microcontroller behind the signal loop in Part 1 — an Amazon affiliate link. As an Amazon Associate this site earns from qualifying purchases, at no additional cost to you.
Sources
“Bionic hands aren't just for Luke Skywalker,” KSL.com — reporting on Avi Davidson, the Utah NeuroRobotics Lab, the LUKE arm trial, the sensory feedback results and the funding situation. Utah NeuroRobotics Lab, University of Utah — research programme. The error table and mass figures used above are instructional models, as noted in the text, and are not attributable to the research team.