Kurzweil Scorecard: Thinking a Hand into Motion

In 2005, Ray Kurzweil sketched a specific picture of how paralysis would end. A sensor in the motor cortex would read the intention to move. That signal would be wired to something that pushed the body back into motion — either the person’s own muscles, or a nano-scale replacement for them.

Twenty-one years later, the picture is half right in a way that matters a great deal, and half wrong in a way that’s telling about how hard technology is to forecast.

The half that’s right: a man with quadriplegia reached for a coffee mug using only his thoughts, in 2017. Another, paralyzed from the waist down, walked uphill on uneven terrain in 2023. By the end of 2025, a 45-year-old with ALS had communicated nearly two million words by thinking them. All three used systems descended from the same piece of hardware — a grid of hair-thin silicon electrodes pressed into the surface of the motor cortex, designed at the University of Utah.

The half that’s wrong: Kurzweil expected the electrodes would drive artificial muscles — nanoelectromechanical systems that would contract and extend in place of damaged biological tissue. That path turned out to be a dead end. The field went somewhere else entirely.

The predictions

This batch groups three of Kurzweil’s claims from the chapter “on the Human Body” in The Singularity Is Near:

First, an observational claim: Kurzweil wrote that University of Utah researchers had found via MRI that long-term quadriplegic patients retained brain activity patterns for intended limb movement very close to those of nondisabled persons (ch. “on the Human Body”). This was the foundation. If the motor cortex still knew how to move a hand, you could in principle read the signal.

Second, a prediction scheduled for the 2010s: “Sensors placed in the brains of paralyzed people will recognize intended-movement patterns and stimulate the appropriate sequence of muscle actions” (ch. “on the Human Body”). The sensor would decode intent; something else would deliver the motion.

Third, a claim about the “something else”: designs already existed for nanoelectromechanical systems that can expand and contract to replace damaged muscles and be activated by real or artificial nerves (ch. “on the Human Body”). This was Kurzweil’s bet on the actuator side of the loop.

Where we actually are

The decoder worked. The foundational claim held up. In 2006, Leigh Hochberg and colleagues published in Nature that a 96-electrode Utah array implanted in a C5/C6 quadriplegic patient, three years post-injury, could still read intent to move the hand with enough fidelity to drive a cursor and a robotic arm. Motor cortex kept its language even when it had no audience to speak to. Every subsequent intracortical BCI result rests on that finding.

Twenty years later, a July 2025 medRxiv preprint reports on the long-term performance of these same arrays across 14 BrainGate clinical trial participants. Over 2,319 recording sessions, the arrays successfully recorded neural spiking waveforms on 35.6% of electrodes, with only a 7 percent decline over the full enrollment period. The implant’s basic premise — that you can listen to the motor cortex — is now an engineering reality, not a hypothesis.

The “stimulate appropriate muscle actions” part landed on schedule — barely. In March 2017, Bolu Ajiboye and colleagues at Case Western Reserve and the Cleveland VA published in The Lancet a proof-of-concept where Bill Kochevar, an eight-year-post-injury tetraplegic, used his BrainGate implant to command a functional electrical stimulation system wired to the muscles of his own arm. He fed himself mashed potatoes. He drank from a mug. He scratched his nose. Twelve years after Kurzweil set the deadline, Kurzweil’s sentence described, almost word-for-word, what Kochevar did in a Cleveland hospital room.

Since then the approach has generalized in several directions. Grégoire Courtine’s team at Lausanne (EPFL and CHUV) published in Nature in June 2023 a “digital bridge”: two 64-electrode ECoG arrays over sensorimotor cortex talking wirelessly to epidural stimulation electrodes over the lumbosacral spinal cord. Gert-Jan Oskam, paralyzed from the waist down for over a decade, used it to stand, walk, and climb stairs, with performance stable for over a year of home use. The sensor reads the intention; the stimulator wakes up the nerves below the injury. The loop Kurzweil imagined, closed at spinal-cord scale.

The “artificial muscles” part missed the target entirely. Kurzweil’s 2005 claim that NEMS designs already existed for muscle replacement was, strictly, true — carbon nanotube yarn actuators have been a real research subject for two decades. A 2022 Nano Research paper reports CNT/polyaniline yarn muscles generating 17% contractile stroke and 8 MPa of isometric stress at under 2 volts in biocompatible solutions. Torsional CNT muscles appeared in Science in 2011. There is a steady cadence of improvements.

None of this has been implanted in a human to replace damaged muscle. Not one. The clinical path for restoring movement after paralysis has gone a different direction: stimulate the biological muscle that’s still there. That’s what Ajiboye did in 2017. That’s what Courtine’s team does with spinal stimulation. That’s what the broader field, including every BrainGate spinoff in our literature scan (roughly eight-fold growth in BCI-paralysis papers between 2007 and 2024 — from 8 annual publications to 63), assumes. The muscle stays. The nerves stay. The injury site gets bypassed, not replaced.

There’s a reasonable explanation: biological muscle, actuated by electrical stimulation, is already a highly evolved actuator with the right mass, metabolism, and sensory innervation. Replacing it with a synthetic alternative would mean reinventing something the body already owns. The clinical calculus favored working with what’s there.

The commercial wave showed up late and is still arriving. Synchron’s Stentrode, delivered endovascularly through the jugular vein instead of through a craniotomy, completed its COMMAND early feasibility study in October 2024: six patients with severe bilateral upper-limb paralysis, all met the primary endpoint of no device-related serious adverse events at 12 months. A Canadian extension trial (FOCUS-CAN) starts in February 2026. Neuralink reported 21 trial participants worldwide by early 2026, up from 12 the previous September, with over 2,000 cumulative implant-days. A Mass General Brigham / Brown team reported in March 2026 that a BrainGate participant typed at 22 words per minute with a 1.6% error rate — the fastest hand-motor BCI result published.

The patent record tracks this acceleration quietly. US 12,390,144, granted August 2025, claims a microelectrode array architecture with reference-centered electrode groups for neural signal feedback. US 12,376,776, a week earlier, describes a monolithic polycrystalline silicon carbide neural interface — a material choice aimed at long-term biocompatibility, directly targeting the slow electrode degradation the 2025 BrainGate analysis documented. US 12,426,823 (September 2025) is an implantable flexible electrode “device and kit” — note the word kit. This is hardware moving from laboratory to supply chain.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern in this batch is sharper than in most. Kurzweil was right about the information half of the problem and wrong about the actuation half.

He nailed the neural side. He understood, two decades ago, that the motor cortex of a long-paralyzed person is functionally intact — that the signal is still there, waiting to be read. He got the rough timeframe right for turning that insight into working systems. The 2017 Ajiboye result is essentially the sentence from the book compiled into a clinical paper. A reader in 2005 who took Kurzweil seriously would not have been surprised by anything that has happened in BCI for paralysis.

He missed on the actuator. NEMS artificial muscles are a plausible idea that stayed plausible — active research continues; incremental improvements continue — but the clinical field voted with its feet. Restoring movement turned out to be an electrical problem, not a mechanical one. The nerves and muscles don’t need replacement. They need a detour around the damaged spinal cord. It’s a more conservative engineering solution than Kurzweil sketched, and it got to patients faster.

There’s a generalizable lesson here, maybe the most important one this batch offers. Forecasters who fixate on a specific technology path often miss the version of the future where the same outcome arrives through older, simpler components. Artificial muscles are a beautiful idea. Electrical stimulation of existing muscles is a 1960s idea. The 1960s idea won because it didn’t have to solve surgery, biocompatibility, and power delivery all at once. Kurzweil’s direction was right. His mechanism was 2005-vintage science fiction fighting a cheaper, older, winning alternative.

The part of this story Kurzweil could not have anticipated from 2005 is the speed at which the field has begun to consolidate commercially in 2025–2026. Three companies — Neuralink, Synchron, and the BrainGate consortium — are now running overlapping trials with dozens of total participants. One of the patents granted last month is an implant “kit.” That word signals that the research era is closing.

Method note

Evidence assembled from three lines of inquiry. First, a full-text search of the US patent corpus (granted and published applications, 2005–2025) for neural interface, intracortical, and artificial-muscle terminology, with publication-date histograms to establish the trend shape. Second, a citation-weighted search of 357 million indexed scientific works for the high-impact landmark papers and their 2025 follow-ups. Third, live web research for company trial counts, clinical-trial registrations, and the peer-reviewed journal publications behind the news. Numbers in the post are from these sources, checked this session. Verdicts are ours; the sentences Kurzweil wrote in 2005 are his.