A young adult man with closely cropped dark brown hair sits in a wheelchair in a dimly lit room, facing a computer screen on which a simple phrase is typed out: “How are you today?” Underneath the question on the screen, three dots appear—similar to the ones messaging apps use to signal that your conversation partner is composing a response—and flash with anticipation. One by one, new words materialize: “I am very good.”
At first glance, this brief exchange may seem ordinary, but the person behind it is far from typical. A small box, roughly an inch across, sits atop his scalp and is connected to a circular opening in his skull about the size of a nickel. A slender black cable extends from the box. The man—who goes by the nickname Pancho—has been paralyzed and unable to speak, yet a brain-computer interface (BCI) records signals from the region of the brain responsible for language production, the vocal motor cortex, and translates them into text on the computer screen.
In 2003, when he was twenty, Pancho suffered a severe brainstem stroke while recovering from a car accident and major surgery. After a week in a coma, he woke up paralyzed and without the ability to speak. The stroke severed connections between his brain and the muscles needed for voice, and even crying could not produce sound.
For years, Pancho’s principal means of verbal communication relied on a pointer attached to a baseball cap. He used the limited mobility in his neck to tilt his head forward and downward to tap keys on a screen, one by one. Spelling out words in this way allowed him to express himself, but at a painstaking pace of roughly five words per minute. “Doctors told me I had zero chance of getting better,” he wrote in a 2021 interview. Yet in recent times, a groundbreaking brain implant and a collaboration with researchers at the University of California, San Francisco have offered him what feels like “a second chance to talk again.”
The Chang lab, led by neurosurgeon and scientist Edward Chang, builds technologies intended to help paralyzed people regain communication by converting brain activity into language. The lab’s speech-prosthesis project integrates many elements of the neuroscience toolkit discussed in other chapters; for instance, like other BCIs, it relies on precise electrophysiological recordings taken from brain regions crucial for guiding behavior. Before engaging with Pancho, Chang and colleagues spent more than a decade exploring the fundamental neurobiology of speech in the human brain by using similar recordings in non‑paralyzed individuals. In brain-mapping studies echoing Wilder Penfield’s pioneering work, the team gradually uncovered core principles about how speech is encoded in neural activity. Pancho’s implant consists of a delicate array of electrodes spread across a flat, flexible rectangular chip resembling a business card, placed on the surface of his vocal motor cortex, the area that governs speech production. Fine wires exit Pancho’s skull through a small coin-sized opening. These electrode grids, hosting between 128 and just over 200 individual electrodes, capture the collective activity of numerous neurons and transmit the neural waveforms to a computer for analysis.
These are not the fever dreams of a tech CEO; they are the result of decades of ingenuity and perseverance in the lab and in the clinic.
The BCI operates on a principle known as persistent somatotopy, whereby a stroke or similar injury interrupts the brain’s outgoing signals to the body (for example, the mouth and tongue), yet the cortical neurons that generate those signals continue to fire. In such cases, BCIs can read the intended muscle movements from the brain. The idea is that, although paralysis blocks signals from reaching the muscles, those signals are still produced in the brain. For Pancho, Chang and his team did not know in advance what those signals would look like or how faithfully they would map to language after sixteen years of paralysis (Pancho was thirty-six when he began collaborating with Chang).
Fortunately, Pancho’s vocal motor cortex still exhibits a robust level of neural activity, though translating those impulses into spoken language remains a formidable challenge. This is where artificial intelligence steps in. An artificial neural network searches for patterns in brain activity associated with phonemes—the building blocks of speech that compose words—that the user reads on a screen. AI does this far more quickly and accurately than a human data analyst could. Whereas earlier models linked neural activity to a fixed list of words, the new AI architecture broadens patients’ vocabularies by focusing on phonemes. When phonemes are combined into different sequences, they form nearly 90 percent of the English language’s phonemic repertoire.
With time, an individual’s AI model becomes better at aligning their neural signals with the words they read and focus on articulating. Eventually, even spontaneous internal speech—the thoughts that arise without prompting—can be anticipated by the AI.
Over the last two decades, paralyzed individuals have been able to communicate using BCIs or similar devices, yet the process tended to be slow. Earlier approaches often captured signals from motor cortex areas that controlled cursor movements on a screen, allowing a user to navigate a keyboard-like array and select letters one by one. Other devices tracked eye movements in patients with locked-in syndrome, enabling letter-by-letter communication as they stared at characters. In both cases, the output hovered around ten words per minute—still faster than Pancho’s tally but far below natural speech, which typically runs around 120 to 150 words per minute. The new Chang lab BCI now achieves roughly 80 words per minute in productive use.
Speech encompasses more than phonemes and vocabulary; it is intimately tied to facial expressions and vocal tone. In a world where much communication is done by text, the nonverbal cues carried by facial movements are essential for conveying emotion. When a patient cannot move facial muscles, that expressive layer is lost. Acknowledging this, Chang’s team engineered their BCI and AI models to interpret activity in the vocal-motor regions in order to restore a voice that sounds personal and to reproduce facial expressions that accompany the utterances. Although reviving a patient’s actual facial movements is not yet feasible, these technologies enable them to convey speech and corresponding expressions via a humanoid avatar displayed on a screen.
The patient at the heart of this research is a woman named Ann, who suffered a devastating stroke while playing volleyball with friends on a sunny day when she was thirty.
Rather than only appearing as text, Ann’s predicted words are voiced by a computer-generated avatar of a woman whose voice closely resembles the one Ann used to have. The process of creating a synthetic voice from prior recordings is known as voice banking. People with neurodegenerative conditions like ALS (amyotrophic lateral sclerosis) can build a library of voice samples before their condition progresses, and those samples can later be used to produce more natural-sounding speech from a BCI like Ann’s. In Ann’s case, the avatar’s voice was generated using recordings taken before her stroke, including her wedding speech captured on video.
The Chang lab’s AI model also enables Ann to convey nonverbal cues by animating the avatar’s facial movements. This latest capability grew out of another state-of-the-art feature of her speech BCI: the human vocal motor cortex contains subregions that drive movements in distinct parts of the vocal tract—such as the tongue, jaw, and lips. Ann’s BCI translates neural activity into the avatar’s speech by focusing not on single words or even phonemes, but on the actions of these vocal muscles. By dissecting speech into the muscular components that underlie individual phonemes, the team could recreate both vocal output and the corresponding facial expressions that accompany it, effectively delivering emotion through the avatar’s animation along with the speech.
In their 2023 Nature paper describing Ann’s BCI, the researchers noted that the AI model made errors in roughly a quarter of cases, though the overall sentences were often still understandable even when several words were off. As technology progresses, the team hopes to reduce errors and enable real-time communication between paralyzed patients and the people around them. A notable advance is the development of a bilingual speech prosthesis that can seamlessly translate neural signals into English or Spanish. To achieve this, scientists used transfer learning—the same approach employed by Mackenzie Mathis to teach DeepLabCut to recognize and track body parts in lab videos. Since different languages may have distinct vocabularies but share phonemes, the AI model applied neural data learned from one language to facilitate learning in the other.
When we imagine brain implants and AI avatars communicating with humans, it’s easy to picture dystopian, ominous, or purely commercial uses: an AI with a lifelike avatar designed to mislead us or a brain implant that merely steers our neural signals to cue the next song on a commute playlist. Even among neuroscientists working in the lab, such applications can sound alluring. Yet this is not what neural-circuit research is fundamentally about. Pancho and Ann, by contrast, stand as concrete examples of how this technology—and the neural-circuit science behind it—are being employed to aid patients. They are not the fever dreams of a tech CEO; they are the product of decades of dedication in the lab and clinic, plus substantial public funding and support.
Funding agencies like the NIH frequently publicize their “moonshots,” such as the Human Genome Project or the BRAIN Initiative, with splashy press events and online campaigns. These grand projects emphasize long-term, nonlinear progress in science, often without a foreseeable cure or a single concrete endpoint. But the motivations driving individual scientists to pursue basic research are often more personal. Across neuroscience labs in the United States and around the world, researchers are chasing a different kind of moonshot—the hopes of patients and their families—the small but meaningful milestones that could improve a child’s life with cerebral palsy or extend the birthday of a grandparent facing Alzheimer’s.
Though science is imperfect and fallible by definition, I think people would feel more trust if they understood a bit more about how it is done, who does it, and why.
When asked about her experience with the Chang lab’s BCI and avatar, Ann recalled becoming tearful upon hearing her own voice again after eighteen years, and she noted it was the first time she could “speak” to her husband in a very long time. Her dream is to one day return to a career in counseling using this technology.
*
Like Ann’s and Pancho’s BCIs, other pioneering treatment approaches drawing on the rapidly expanding neural-circuit toolkit may not come in the form of a pill. They could rely on an implanted electrode, a designed viral vector, or a noninvasive method such as a precisely timed sensory stimulus—such as a specific rhythm of light or sound. A prominent example from recent years stems from a 2016 Nature study conducted in the lab of MIT neuroscientist Li-Huei Tsai in collaboration with Ed Boyden and Emery Brown. The work proposed a potential new strategy to mitigate the severe consequences of Alzheimer’s disease.
One of the brain’s intrinsic patterns is rhythmic firing at about 40 hertz, known as the gamma frequency. Gamma activity correlates with focus and active engagement in tasks. Earlier studies suggested that gamma activity is reduced in people with Alzheimer’s, prompting Tsai and colleagues to explore what might happen if this brain rhythm could be artificially restored. They used a genetically engineered mouse model that mirrors early-stage Alzheimer’s disease. These mice exhibit hallmarks similar to human disease, including protein aggregates and cognitive decline. In the 2016 study, they activated neurons to emit gamma-frequency signals not through electrodes, but via noninvasive sensory stimulation—flashing lights, sound pulses, or skin vibrations at 40 hertz.
To their surprise, stimulating the gamma rhythm for about an hour led to a marked decrease in brain debris in the mice. Subsequent research revealed a mechanism: the 40-hertz gamma rhythm supports recruitment of microglia—the brain’s cleanup cells—helping clear waste products. In other words, the same neural rhythm that engages large networks also summons the brain’s housekeeping staff to tidy up. Gamma activity also appears well-suited to activating inhibitory neurons, which are thought to be dysfunctional in Alzheimer’s and play a crucial role in balancing neural circuit activity. Interestingly, the Alzheimer’s mice showed improved recognition of familiar objects and important spatial locations after gamma stimulation, suggesting debris clearance may drive cognitive benefits.
Although most experiments have used mice—scientists jokingly note that mice have never had it better—researchers have worked to translate these findings to humans and strengthen the evidence base with clinical trials. They tested the 40-hertz stimulation approach, called GENUS (Gamma ENtrainment Using Sensory stimuli), in human participants in a study reported in late 2022.
Early results were encouraging. In many participants with Alzheimer’s, atrophy progresses rapidly over a few months. Those who used GENUS showed less brain atrophy than the control group. They also demonstrated improved sleep and stronger functional connectivity between brain regions associated with learning and memory. While there was not a clear difference in cognitive performance versus controls within the short study window, the researchers noted the trial was cut short by COVID-19 disruptions, and they anticipate assessing longer-term benefits in future work. The GENUS procedure did not produce troubling side effects and uses a device that delivers light and sound at 40 hertz, a simple and affordable at-home option for patients.
Reflecting on the evolution of neuroscience since I began graduate study in 2009, I’m charged by the speed at which progress can occur when resources, collaboration, and persistence align. Back then, recording spikes from a single neuron in an awake animal was challenging, and identifying neuron types was not routine. We tended to analyze only a few basic features from a handful of cells. It feels like the Stone Age compared with today’s capabilities.
Yet, given the rapid strides described here, I’m deeply concerned that by 2025 the basic research funding environment for brain science in the United States stands at an unprecedented crossroads. Since the bipartisan NIH era that supported landmark achievements like the Human Genome Project, public enthusiasm for science has waned. The NIH’s reputation has suffered in the wake of the COVID-19 pandemic, and today expertise can be outpaced by political ideology, research institutions find themselves needing to align with a hostile government, and the United States risks losing its edge in attracting top scientists from abroad. Even if the public remains broadly supportive of biomedical research, there’s a level of skepticism and cynicism toward science that is perhaps greater than at any time in our lifetimes. Though science remains imperfect and fallible by definition, I believe people would trust it more if they understood a little better how it is done, who does it, and why.
_______________________________
Adapted from Circuit Breakers: How Neuroscientists Get Inside Your Head by Anna Chambers. Copyright © 2026. Available from Abrams Books.
