Stanford neuroprosthesis decodes Pat Bennett's speech at 62 wpm

AI system

System: Stanford intracortical speech neuroprosthesis (BrainGate2 trial)
Model: A five-layer recurrent neural network with gated recurrent units, combined with a trigram language model.
Release: Exact release not established.
Application developer: Stanford Neural Prosthetics Translational Laboratory (Willett, Henderson, Shenoy) with the BrainGate consortium
Role: The decoder and language model converted neural activity into text.
Configuration: Four arrays, each with 64 electrodes, recorded activity in cortical areas 6v and 44. The language model used a vocabulary of 125,000 words.
Inputs and outputs: The system converted cortical recordings into words.
Autonomy: Not assessed
Workflow: Preserved articulatory code for phonemes in ventral premotor cortex years after loss of speech. Four intracortical arrays in speech-related cortex. RNN decoder plus large-vocabulary language model trained on the participant's attempted speech. Regulated trial setting (BrainGate2) with research support.
Human oversight: Not assessed
Relation to AI: An RNN decoder plus an n-gram language model is the mechanism translating neural activity into text. The Stanford neurosurgeon describes the language model as “a sophisticated auto-correct”. The benefit is attributable to the decoding system.
AI contribution: An RNN decoder plus an n-gram language model is the mechanism translating neural activity into text. The Stanford neurosurgeon describes the language model as “a sophisticated auto-correct”. The benefit is attributable to the decoding system.

Benefits

Outcome identifiers: Not assessed
Benefit: produce text at 62 wpm, 23.8% word error rate
Category: Improved Accessibility
Status: Not assessed
Beneficiary groups: People.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Measurement

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: produce text at 62 wpm, 23.8% word error rate
Comparison: the previous state-of-the-art BCI (a handwriting BCI)
Results: Reference: 18 words per minute for a handwriting BCI. Result: 62 words per minute.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Money

Kind of value: Not assessed
Amount: Not assessed
Currency and prices: Not assessed
Period and unit: Not assessed
Realization: Not assessed
Recipients and payers: Not assessed
Gross or net: Not assessed
Included costs: Not assessed
Comparison: Not assessed
Method: Not assessed
Return and payback: Not assessed

People and places

Participants: Developer: Stanford Neural Prosthetics Translational Laboratory (Willett, Henderson, Shenoy) with the BrainGate consortium. Operator: Stanford Medicine.
Jurisdictions: Benefit: Stanford, United States. Developer: United States. Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Follow-up

Conditions for replication: Array signal quality remains adequate. Participant can attempt speech and tolerate recording sessions. Decoder retraining available across sessions.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: No dated development for this participant after the Nature report was found in the cited sources.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Evidence

Source numbers: AIBD-0004-source-0001–0003
Citations: Numbered references below.
Language: Not assessed
Source relationships: Developer, Independent
Original evidence: Not assessed
Access and versions: Not assessed
Supported findings: Not assessed
Evidence links: Not assessed
Evidence strength: Not assessed
Missing evidence: Not assessed

Grades

Assessed benefits: People.
Depth: People: 2.
Reasons: People: Material benefit to one identifiable person: large-vocabulary communication at conversational pace in research sessions. The team itself called it a proof of concept rather than an everyday device.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Entry details

Kind

Not assessed

Related entries

Both decode speech for people with impaired speech; Whisper fabricated words and violent phrases when transcribing impaired speech.

Field

Healthcare / assistive communication. ALS with loss of intelligible speech.

Purpose

Built to restore communication by decoding attempted speech from intracortical recordings into text; large-vocabulary text communication was the intended benefit and is what occurred.

Dates

Date basis unassessed: 2022 (precision unassessed). Updated: 21 October 2026.

Stage

Not assessed

Version

SHA-256: 9faa389e1dc7…

References

Other identifiers: AI Incident Database: 732.
Category mappings: Not assessed
Brief: Not assessed
Revisions: 6 revisions. Latest: 21 October 2026.

Sources

Peer-reviewed study Developer

Willett FR, Kunz EM, Fan C, et al. (2023). A high-performance speech neuroprosthesis. Nature.
News report Independent author

Hamilton J (2023). These experimental brain implants can restore speech to paralyzed patients. NPR (via BPR).
Press release Developer

Goldman B (2023). Brain implants, software guide speech-disabled person's intended words to computer screen. Stanford Medicine News (developer institution release).