Thought-to-Text Status

"Mind typing / brain-to-text" is one of the earliest practical goals of BCI: letting the user type with nothing but thought. In 2024–2026 this goal reached milestones of 60+ WPM invasively and 30+ WPM non-invasively, but real-time, consumer-grade accuracy remains distant. This article compares the current state of the art.

1. Evaluation Metrics

Main Metrics

WPM (Words Per Minute): text-input speed
CPM (Characters Per Minute)
WER (Word Error Rate): accuracy
Latency: thought → display
Calibration time: training before first use

Baseline References

Input method	Speed
Proficient keyboard	60–80 WPM
Smartphone	30–40 WPM
Siri voice	100+ WPM (theoretical)
Eye-gaze typing (tracker)	10–15 WPM
Early BrainGate	6 WPM (2006)
Willett 2021 handwriting BCI	90 CPM (~18 WPM)
Willett 2023 speech BCI	62 WPM
Metzger 2023 avatar	78 WPM

2. Invasive Status (2024–2025)

Top Performance

Willett 2023: 62 WPM, WER 9.1% (50-word vocabulary)
UC Davis Card 2024: 62+ WPM with LLM
Metzger 2023: 78 WPM speech + avatar

Limitations

Requires Utah Array or ECoG
Craniotomy
Only used for severe disability

Trends

Neuralink's 1024 channels → theoretical ceiling of 150+ WPM
Higher speeds expected to be reported clinically in 2026–2027
Approaching "natural conversation"

3. Non-invasive Status

EEG Brain-to-Text

MindBig Data P300 typing: ~5 WPM
DeWave 2023: category-level, not word-level
EEGPT 2024: general pretraining, weak at fine detail

MEG Brain-to-Text

Meta Défossez 2023: recognizes 10 words from MEG signals
Progress but far from practical
MEG equipment is not portable ($2M+)

fMRI Brain-to-Semantics

Tang 2023: semantic level, not word level
Akin to "meaning translation"
Not "typing"

Non-invasive Performance Bottlenecks

Poor signal
High noise
Consumer grade may reach 2020-era invasive levels in 10 years

4. Invasive vs Non-invasive Comparison

Dimension	Invasive (Willett/Neuralink)	Non-invasive (Meta/consumer)
Speed	60+ WPM	< 10 WPM
Accuracy	WER < 10%	WER 30–50%
Latency	< 500 ms	Seconds
Invasiveness	High (craniotomy)	Zero
Channel count	96–1024	4–128
Users	Severe disability	Anyone
Usable in 2026	Yes (research)	Partially

5. Middle Ground: Minimally Invasive

Stentrode

16 intravascular channels
~10–15 WPM (estimated)
Already useful for ALS users
See Synchron_Stentrode

Precision Layer 7

1024 surface channels
Theoretically close to invasive grade
Lower surgical risk
May be the intermediate stop on the consumer path

6. LLM Acceleration

Rescoring

See LLM post-processing fusion.

GPT-4 semantically corrects BCI output
Sometimes boosts effective fluency from 7 WPM to 30 WPM equivalent

Autocomplete

Type "I want" → LLM predicts "a coffee"
User confirms rather than types
Greatly accelerates real communication

Dialog Management

LLM manages the whole conversation
BCI only needs low-bandwidth confirm/select
Consumer-grade BCI becomes viable

7. Consumer-grade Feasibility Analysis

Status Quo (2026)

Non-invasive EEG: not accurate enough, not fast enough
Stentrode: too invasive
Apple AirPods EEG: too early

Problems

Consumers won't wear an EEG helmet
Won't undergo surgery to type
Existing touchscreens and voice are already good enough

Possible Breakthroughs

Dry-electrode headgear (integrated with AR glasses)
Wrist EMG (Meta CTRL-Labs style)
Emotion/intent level rather than word level
"Assistive input" rather than "keyboard replacement"

8. Mind Typing in AR / VR

Vision Pro (2024)

Eye tracking + gestures + voice
No EEG
Gaze ≈ approximate mind typing

Future

Periocular EEG
Gesture EMG
Multimodality = each modality weak, combined strong

Meta Orion (2024)

Wrist EMG band + AR glasses
Gesture-neutral (no large movements needed)
Replacement for keyboard input on AR

9. Who Breaks Through to Consumer Grade First?

Candidates

Apple: AirPods EEG + Vision Pro integration
Meta: CTRL-Labs EMG + Orion
Snap: NextMind integrated into Spectacles
Google: no consumer-BCI initiative publicly disclosed as of 2026-04
Samsung: Galaxy ecosystem

Projections

2027–2028: AR glasses + multimodal input, approaching practical assistive input
2030: mind typing as one of AR's primary inputs
2035+: on par with the keyboard

10. Limits: Why "Full Keyboard Replacement" Is Hard

1. Bandwidth Limits

Cortical surface signal is not dense enough
Whole-word speed is capped

2. User Fatigue

Sustained "thinking" is more tiring than fingers
Neurofeedback demands focus

3. Precision

Natural thought is non-linear
Typing demands symbolic thinking

4. Privacy

The boundary of "think-to-type" is fuzzy
May leak things not intended for transmission

11. Disabled Users vs Healthy Users

Disabled Users

Genuine must-have: no other option
Tolerate low speeds
Early beneficiaries

Healthy Users

Superior alternatives (keyboard, voice)
Only have an edge in AR scenarios
Slower adoption

Strategic divergence: invasive targets medical; non-invasive targets consumer AR.

12. Logic Chain

Mind typing reached 62 WPM invasively and < 10 WPM non-invasively in 2024–2026.
Willett 2023 and Metzger 2023 are invasive milestones; Meta Défossez is a non-invasive probe.
LLM acceleration brings low-bandwidth BCIs close to practical (autocomplete + rescoring).
Consumer-grade isn't practical today, but AR glasses + EMG + eye tracking form a multimodal path.
Apple, Meta, Snap are competing in consumer BCI.
2027–2030 consumer AR BCI expected to mature.
Disabled vs healthy: invasive medical vs non-invasive consumer is the differentiation.

References

Willett et al. (2023). A high-performance speech neuroprosthesis. Nature.
Métzger et al. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature.
Willett et al. (2021). High-performance brain-to-text communication via handwriting. Nature.
Défossez et al. (2023). Decoding speech perception from non-invasive brain recordings. Nat Machine Intelligence.
Card et al. (2024). An accurate and rapidly calibrating speech neuroprosthesis. NEJM.