Overview
The following essay, originally titled “Guaranteeing Agency in the Age of Neural Prosthetics,” was selected as one of three finalists in MIT’s 2026 Envisioning the Future of Computing Prize. Its author, Strahinja Janjušević, SM ’26, who studied computer science and cybersecurity at the US Naval Academy and just earned his master’s in the Technology and Policy Program of MIT’s Institute for Data, Systems, and Society, now runs Refractal, an AI cybersecurity company based in London.
His entry posits that AI could make it possible to restore physical agency to millions living with limb loss and paralysis—but warns that tethering the human body to proprietary technologies comes with lots of risk. He proposes establishing a new set of standards to protect patients.
The annual competition is cosponsored by the School of Humanities, Arts, and Social Sciences and the Social and Ethical Responsibilities of Computing (SERC) initiative at the Schwarzman College of Computing. You can also find the entries by the other finalists, grad students Rachel Sava and Cordiana Cozier, here.
The year is 2029. Julian, a 24-year-old software engineer, sits in a coffee shop in Cambridge, drinking coffee and coding. Four years ago, after a car accident, a spinal cord injury left him paralyzed from the neck down. Today, he is typing code at roughly 90 words per minute and holding a hot espresso. His movements look indistinguishable from anyone else’s. Julian is not biologically healed, though; he is “bridged.”
A high-bandwidth brain-computer interface implanted in his motor cortex streams raw neural spikes to an AI model running on an exoskeleton. The AI doesn’t just listen; it anticipates, smoothing his neurologically noisy intent into graceful motion. For Julian, the AI is part of him. It is his hands, his legs, his freedom.
As he gets ready to stand up, however, a notification appears on his retinal display.
“SERVER UNREACHABLE. LICENSE VALIDATION FAILED. MOTOR FUNCTIONS SUSPENDED.”
The miracle vanishes. Julian’s arms lock. The cup falls from his hand and shatters on the floor. He is paralyzed again. Not because of biology, but because the startup that owns his subscription just declared bankruptcy.
This scenario is the inevitable conclusion of our current trajectory in neurotechnology. As we witness the rapid strides occurring to establish direct communication pathways between the brain and external devices, the stakes are becoming tangible. We see this most clearly with Neuralink, the company pioneering implantable chips that allow users to control digital devices through neural activity alone.
While the capabilities of A I prosthetics are promising, we must recognize the danger in treating these devices like consumer electronics.
As proprietary AI models become capable of acting as the “operating system” for the human body, we face a profound ethical divergence. We are building a future where people with disabilities can walk, run, and work again, but only as tenants in their bodies. To realize the true potential of AI in health care, we must answer a question: When the software crashes, who owns the limb?
AI offers tangible restoration to those who have suffered catastrophic accidents. For decades, prosthetics were unintelligent mechanics or passive devices. To operate a robotic arm, users had to engage in myoelectric switching, a process of toggling muscles in their chest or back to trigger simple movements. It is so cognitively exhausting, unintuitive, and slow that 23% of people who use these myoelectric devices abandon them.
A shift to AI models I call generative motor policies would address this friction. Just as a large language model predicts the next word in a sentence from context, a generative motor policy would simulate the next millisecond of muscle contraction on the basis of seamless, thought-driven intent.
Today, the use of what’s known as transformer architecture, which can process an entire sequence in parallel rather than one step at a time, is advancing the development of more capable neural prosthetics. Unlike earlier neural networks, transformers can flexibly attend to many signals at once, including the dynamics of the cup, the geometry of the hand, and the noisy neural recording, weighing them together. When Julian thinks “grab,” the model doesn’t just decode the signal; it infers the trajectory most likely to fulfill that intent.
Details
This shift from command following to intent prediction is what makes the technology viable for widespread deployment within the decade.
The potential impact extends beyond spinal cord injuries to reach the approximately 1.6 million Americans living with limb loss (a figure projected to double by 2050)—including thousands of veterans from recent conflicts. Traditional prosthetics essentially act as fancy springs. They cannot run, jump, or stabilize on uneven terrain. AI changes this dynamic. A veteran with a smart amputation wouldn’t get a plastic foot but a robotic leg that would identify stairs or gravel and adjust its torque in milliseconds. This would make it possible to return to active duty, compete in sports, or simply take care of a child.
AI-based prosthetics would also have a positive fiscal impact. In 2022, the lifetime cost of care for a young person who sustained high tetraplegia (an injury to the C1–C4 vertebrae) at age 25 exceeded $5.8 million. AI prosthetics have the potential not only to reduce the cost of care but also to unlock the productivity of a workforce that has been systematically sidelined.
But while the capabilities of this technology are promising, we must recognize the danger in treating the devices like consumer electronics. You cannot swap out a brain implant like a SIM card. We have already seen a preview of the dystopian possibilities. In 2020, Second Sight Medical Products, a pioneer in bionic eyes, suffered a financial collapse, leaving hundreds of patients with Argus II retinal implants stranded. Their bionic eyes still worked, but the software ecosystem vanished. When parts broke, there were no repairs. When the software glitched, there was no patch. They were left with useless metal in their skulls.
Extrapolate this situation to 2030, and the risks compound. We can imagine tens of thousands of people who rely on proprietary AI models to walk. We face the risk of ransomware, where a hacker seizes control of the server and demands payment to unlock 50,000 pairs of legs. We face the risk of subscription traps, where a company introduces surge pricing for mobility. Most likely, we face the risk of bankruptcy: If a startup fails and the servers go dark, the user will be blocked.
We must adopt a regulatory framework that treats the code running a human body differently from the code running a thermostat.
Open-source software standards could help keep devices going, but critics may worry that this poses a security risk. If the code for walking is public, can’t a malicious actor hack a user’s legs? But “security through obscurity,” the belief that hiding code makes it safe, is misguided. In 2017, the FDA issued a voluntary firmware recall covering approximately 465,000 Abbott pacemakers after researchers demonstrated vulnerabilities. Because the code was closed, the gaps remained hidden in firmware for years. The obscurity protected the vulnerability, not the patients.
To ensure that the net impact of neural prosthetics is positive, we cannot rely on the benevolence of tech companies to secure human agency. We must adopt a regulatory framework that treats the code running a human body differently from the code running a thermostat. I propose what I call the Neuro-Escrow Standard (NES).
This policy safeguard has two pillars. The first is the right to repair. The NES would mandate that any company seeking FDA approval for a neural interface deposit its source code, model weights, and code-signing keys into a federal escrow trust that remains sealed as long as the company supports the device. However, a “trigger event”—such as bankruptcy or server shutdown—would automatically put the code in the public domain. This means open-source developers can ensure that no patient is ever marooned by a balance sheet.
The consequences of not having this mechanism are already measurable. When Medtronic could not patch multiple models of its insulin pumps in 2019, the FDA stated plainly that the manufacturer was “unable to adequately update” the devices, forcing roughly 4,000 US patients to transfer to alternative pumps.
Open-source medical communities have demonstrated that they are capable of stepping in, creating a kind of global technological “immune system.” We see this in the OpenAPS movement, where patient-hackers built an open-source artificial pancreas system for automated insulin delivery. In a 2022 randomized controlled trial published in the New England Journal of Medicine, the system was proved safe and effective. Tidepool Loop, another patient-built open-source system, secured a key FDA clearance in 2023.
The second pillar of NES would focus on sovereignty, rejecting cloud-dependent architecture for vital functions. The minimum viable function (MVF)—the subset of operations necessary for basic embodied agency, specifically the ability to walk, grasp, and speak—should be required to run via edge computing, meaning it must be performed on or adjacent to the device itself. Cloud connectivity can be used for optimization or learning, but it cannot be essential for operation. If the internet cuts out or the server fails, the legs must still walk.
Ultimately, this local-first architecture ensures that agency is never rented from a data center, so that the user’s MVF can never be compromised by a balance sheet loss or a server failure. This technical decoupling is necessary to ensure that cybernetic restoration is as reliable as biological health.
For people with disabilities, the transition from biological bodies to cybernetic ones is inevitable. The technology is already here. The policy is what’s falling behind. If we do nothing, we risk creating a future where people are liberated from their physical limitations only to be shackled by new ones. We risk a world where a declined subscription payment causes a literal loss of footing.
With safeguards, we can build a future where neural prosthetics are not corporate appliances we use but open-source instruments we control. By guaranteeing that the code governing them is durable and accessible, we ensure that AI does not simply manage disability but truly erases it. We unlock a future where Julian does not drop his cup, because his hands, finally and permanently, belong to him.
Source
Originally published at www.technologyreview.com.