Brain-computer interfaces
Implants decode movement and speech intentions from the brain, giving paralyzed people back communication.
Open in the interactive tree →Since 2023 deep-learning decoders have transformed speech BCIs: two studies reached 62 and 78 words per minute (Stanford and UCSF), several times faster than before and closer to the roughly 160 words per minute of natural conversation. Companies such as Neuralink, Synchron, Paradromics and Precision are moving implants toward commercial devices. Most users have ALS, stroke or spinal-cord injury.
As of October 2026
Neuralink reported 21 participants implanted worldwide on 28 January 2026, up from 12 in September 2025 (company figures reported by Reuters). Synchron's less invasive Stentrode completed a six-patient US feasibility study, and in November 2025 the company raised 200 million US dollars to fund a pivotal trial planned for 2026. Stanford researchers decoded imagined 'inner speech' with up to 74 percent accuracy (Cell, August 2025), raising privacy questions, and 2024 Caltech work showed results varying sharply between two participants (79 percent versus 23 percent on a set of six words and two pseudowords). As of mid-2026 no long-term implanted BCI had received full FDA market approval.
Open steps
- Electrodes that last decades Low AI leverageScar tissue and failing insulation make recordings fade; in one monkey study failed arrays lasted a mean of 332 days. Materials must last for decades.
- Decoders that need no daily retraining High AI leverageBrain signals drift, so decoders are retrained often; stable decoders that work across days and people would make implants practical at home.
- Wireless, low-power implants Medium AI leverageToday's systems often need connectors or external hardware; fully implanted devices must decode on-chip within tight heat and power limits.
- Decoding only what the user intends Medium AI leverageInner-speech decoding raises privacy risk; devices must reliably separate intended from private thoughts and keep neural data secure.
Where AI could help
Medium AI leverage. AI decoders turn neural signals into speech and control, but implant lifetime, surgery and approval limit rollout.
- Decoding attempted speech into text or voice in real time
- Calibration-free decoders that adapt to signal drift
- Foundation models pooled across patients and sessions
- Spotting implant degradation early
Shown so far
- In June 2025 Nature published a UC Davis speech neuroprosthesis that turns brain activity into synthesized voice with about a 25 ms delay; listeners understood about 60% of words, in one participant with ALS. source
Prerequisites
- Penfield's brain map1937Decoding movement and speech uses cortex areas first mapped by Penfield
- Neuroprostheses and DBS1987
- First implanted BCIs2004
- Deep Learning2012