UCSF neuroprosthesis lets Pancho produce sentences at 15.2 wpm

AI system

System: UCSF speech neuroprosthesis (BRAVO trial)
Model: Deep-learning models detect speech and classify words. A language model and a Viterbi decoder assemble the output.
Release: Exact release not established.
Application developer: UCSF Chang Lab
Role: The models detected attempted speech, classified words, and assembled sentences.
Configuration: The system used a 128-electrode array with 4 mm spacing and a vocabulary of 50 words.
Inputs and outputs: The system detected attempted speech in cortical recordings and classified words from a 50-word vocabulary.
Autonomy: Not assessed
Workflow: Preserved speech-motor cortical pathways more than 15 years after loss of speech. High-density ECoG array over sensorimotor cortex. Deep-learning models trained on 22 hours of attempted-speech data plus language-model decoding.
Human oversight: Not assessed
Relation to AI: Deep-learning models detected attempted speech in 98% of attempts and classified words. A language model with Viterbi decoding assembled sentences. The output depended entirely on these models. Funding included a research contract under Facebook's Sponsored Academic Research Agreement. The paper states that no study hardware or data were transferred to any sponsor and that author disclosure forms are available at NEJM.org.
AI contribution: Deep-learning models detected attempted speech in 98% of attempts and classified words. A language model with Viterbi decoding assembled sentences. The output depended entirely on these models. Funding included a research contract under Facebook's Sponsored Academic Research Agreement. The paper states that no study hardware or data were transferred to any sponsor and that author disclosure forms are available at NEJM.org.

Benefits

Outcome identifiers: Not assessed
Benefit: sentences at a median 15.2 words per minute with a median word error rate of 25.6%
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: sentences at a median 15.2 words per minute with a median word error rate of 25.6%
Comparison: Not assessed
Results: sentences at a median 15.2 words per minute with a median word error rate of 25.6%
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: UCSF Chang Lab. Operator: University of California, San Francisco.
Jurisdictions: Benefit: San Francisco, 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: Participant can attempt speech and retains intact cognition. Implant remains stable with adequate signal quality. Accumulated training data maintained over time.
Replication attempts: 2 follow-up entries. Benefit recurred: 2.
Later findings: Not assessed
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Evidence

Source numbers: AIBD-0005-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: first decoding of his own full words, three times faster than his pointer typing. But limited to a 50-word vocabulary in research sessions.
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. Post-stroke anarthria.

Purpose

Built to decode attempted speech into words and sentences for a person with anarthria. Restoring word-level communication was the intended benefit and is what occurred.

Dates

Date basis unassessed: 2019 (precision unassessed). Updated: 21 October 2026.

Stage

Not assessed

Version

SHA-256: fb4b93baaa4c…

References

Other identifiers: AI Incident Database: 732.
Category mappings: Not assessed
Brief: Not assessed
Revisions: 5 revisions. Latest: 21 October 2026.

Sources

Peer-reviewed study Developer

Moses DA, Metzger SL, Liu JR, et al. (2021). Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria. N Engl J Med.
News report Independent author

Advisory Board (2021). For the first time since 2003, this paralyzed man can 'speak'. Advisory Board Daily Briefing.
Other Independent author

Reis e Sousa C, Poirier E (2021). Scientists Detect Brain Signals to Help a Paralyzed Man Speak. Harvard Science in the News.