Whose words?
This prototype is an evocation of one repeated choice a person makes when they speak through a brain-computer interface.
Decode a sentence, then decide whether to accept the machine version.
The choice is, “Do I accept this version of what I wanted to say or not?”
Take the job. You’re ready for this.
Take the job. You’ll do great, sweetheart.
It’s close. It isn’t quite what you meant.
Before you accept anything.
After accepting an AI-interpretation.
interaction design & prototyping
The trade-off you just felt is real.
The interaction is a Wizard of Oz prototype, a common design method where the ‘smart’ system is faked so you can test the experience before building the real thing. The responses are scripted. The energy meter and the word predictions represent a real struggle: the more an AI fills in what you’re trying to say, the faster you can “speak,” but the harder it can be to keep the words truly yours.
That slow loosening of authorship is what is called “drift.” The trade-off you just felt is real: studies have found people giving up some control of their exact words for speed, and AI assistance quietly changing what a person meant. The energy meter and the fatigue framing help you empathize with the people using these systems.
Each predicted word is an invisible assertion about the speaker’s identity.
Whose Words? focuses on the relationship between an individual who has lost their ability to verbally communicate and a speech brain-computer interface that attempts to anticipate and complete speech for the individual. The individual bypasses the system to gain an external vocal apparatus. However, each predicted word is also an invisible assertion about the speaker’s identity.
The work examines a recurring question. At what point does your speech get lost when saying it requires more of the machine to do the talking?
Rather than attempting to articulate the effects of that trade-off to you, the prototype provides a more direct understanding. You become the speaker, and the prototype displays the remaining capacity to speak using the construct of an energy meter that is drained each time you attempt to speak.
In each turn, you are presented with the same challenge as the people who use the brain-computer interface. You can accept the phrasing provided by the machine, which is quick, or expend the little energy you have to replace the phrasing with something less machine-like. Because the cost of fixing the phrasing is always greater than the cost of accepting the phrasing, you lose energy to the machine’s version rather than to gain anything from it.
The real speech-BCI science this piece evokes.
How fast and accurately attempted speech can be decoded.
- 2023Stanford / BrainGate · Nature 2023
Attempted speech at 62 wpm
An intracortical speech BCI decoded the attempted speech of a participant with ALS.
Willett et al., Nature 620, 1031–1036 (2023) - 2023UCSF / UC Berkeley · Nature 2023
78 wpm from the cortical surface
High-density surface recordings decoded text for a participant paralyzed by a brainstem stroke.
Metzger et al., Nature 620, 1037–1046 (2023) - 2024UC Davis / BrainGate · 2024
97–99% accuracy in real conversation
For a participant with ALS (Casey Harrell), a 256-electrode system was used in self-paced conversation.
Card et al., NEJM (2024) · UC Davis Health
- Each result involves a single participant in an investigational trial, not an approved product. Whether it generalizes is not yet known.
- Speeds vary by study, vocabulary size, and controlled tests vs. free conversation. Numbers are ranges, not constants. Casual speech is roughly 150–160 wpm.
- These systems decode attempted speech, signals from the motor cortex as a person tries to speak. They cannot “mind-read.”
The decision science behind the accept-or-correct mechanic.
- 1995Sniezek & Buckley · 1995
The Judge–Advisor System
The standard paradigm for studying how people take advice.
Sniezek & Buckley, OBHDP 62(2), 159–174 (1995) - 2001Sniezek & Van Swol · 2001
Familiarity can inflate trust
An ongoing relationship can raise trust independent of an advisor’s accuracy.
Sniezek & Van Swol, OBHDP 84(2), 288–307 (2001)
The authorship and agency question at the heart of the piece.
- 2023Sankaran et al. · 2023
Designing for user agency
Recommendations for promoting user agency in speech neuroprostheses.
Sankaran, Moses, Chiong & Chang, FHN 17 (2023) - 2024Freudenburg et al. · 2024
Preserving “speech ownership”
A 2024 editorial frames the open question of agency as neural devices produce speech.
Freudenburg, Berezutskaya & Herbert, FHN 18 (2024)
The thread
Cadence
Helping people communicate with an EEG brain-to-computer interface
BCI Emotion Capture
Brain-computer interfaces already turn attempted speech into text. This turns those same words into images, so a person who cannot speak can show what they feel rather than only spell it.
Where this stands
An independent design exploration grounded in the published research on authorship and agency cited above. The interaction is a Wizard of Oz prototype and the responses are scripted. Nothing here was built, and nothing was tested with people who rely on assistive communication. It is an argument about a tradeoff, made in the form of something you can use.