# Is EEG-to-Text Actually Usable Outside the Lab?

A 128-channel high-density [Electroencephalography (EEG)](https://bciintel.com/glossary/eeg) benchmark released today by researchers from Harbin Institute of Technology, Nanjing University, and Zhongguancun Laboratory identifies the core reason EEG-to-text decoding has stalled at the academic proof-of-concept stage: virtually every published model depends on **teacher-forcing** during evaluation, a methodological crutch that makes decoding look functional when it is not. Without teacher-forcing — meaning without feeding the model the correct previous tokens at each step, as would be required in any real deployment — current EEG-to-text (EEG2Text) systems fail to generate meaningful output. The new corpus, called **COFETT (Corpus OF Eeg-To-Text)**, is designed specifically to expose and correct this flaw. According to the paper (arXiv:2607.18749), COFETT achieves state-of-the-art ability to distinguish among model performances and enables robust, teacher-forcing-free evaluation. The work directly bears on whether non-invasive neural communication devices could ever reach patients with [Amyotrophic Lateral Sclerosis (ALS)](https://bciintel.com/glossary/als) or severe paralysis who cannot tolerate surgery — the population currently served only by invasive [ECoG](https://bciintel.com/glossary/ecog)-based systems.

---

## The Teacher-Forcing Problem: Why It Matters

Teacher-forcing is a training and evaluation shortcut borrowed from sequence-to-sequence natural language models. During decoding, instead of using the model's own previously predicted token as input to generate the next one, teacher-forcing feeds in the *correct* ground-truth token. This works fine for training language models on text corpora, where the correct sequence is known. Applied to EEG2Text evaluation, however, it creates a fundamental validity problem: the model is never actually tested on its ability to generate free-form output from brain signals alone.

The authors argue this explains a pattern that has confused the field for years — EEG2Text papers reporting apparently impressive metrics that do not replicate in any practical setting. The dependency on teacher-forcing has, per the paper, "fueled numerous debates about whether EEG2Text is a meaningful direction, by extension, and whether EEG truly contains decodable linguistic information."

This is not a minor technical quibble. For a [communication BCI](https://bciintel.com/glossary/communication-bci) to be clinically useful, it must generate text autoregressively from neural signals in real time, with no access to ground-truth tokens. A model that only works with teacher-forcing is, by definition, not a communication device — it is a classification system that happens to output words.

---

## What COFETT Does Differently

The researchers built COFETT using a **128-channel high-density EEG cap** and grounded its design in neuropsychology — specifically targeting the EEG instability problem they identify as a flaw in existing benchmarks. EEG signals are notoriously non-stationary: signal characteristics shift across sessions, electrode placements, and cognitive states. Existing benchmarks, the authors contend, have failed to account for this instability adequately, which confounds model inference and makes cross-session generalization nearly impossible to assess honestly.

COFETT's neuropsychology-informed paradigm addresses this directly. The paper reports that in head-to-head comparisons with multiple existing benchmarks, COFETT better distinguishes among model performances — meaning it has higher discriminative validity, the ability to tell apart a genuinely better model from a worse one. Critically, it supports teacher-forcing-free evaluation, the condition that actually approximates real-world use.

The dataset is open-sourced, which is significant. Proprietary or restricted datasets have been a bottleneck in this subfield, limiting independent replication of published results.

---

## Skeptical Analysis: What the Paper Does Not Resolve

The paper provides "key evidence for the feasibility of teacher-forcing-free EEG2Text decoding" — that phrasing deserves scrutiny. Feasibility evidence from a single new benchmark developed by the same team that designed it is promising, not definitive. Several open questions remain:

**Decoding accuracy at usable rates.** The paper does not report bits-per-second throughput for their best-performing teacher-forcing-free model. The field's benchmark for clinical utility in communication BCIs is roughly 40 bits per minute for basic communication, with leading intracortical systems now exceeding that substantially. Non-invasive EEG has fundamental spatial resolution limits — skull and scalp act as low-pass filters, blurring the cortical signals that carry fine-grained linguistic information. Whether 128 channels at scalp level can carry enough decodable linguistic content for practical communication remains unproven.

**Generalization across participants and sessions.** The abstract does not detail the number of participants, session counts, or cross-subject decoding performance. EEG non-stationarity is addressed in benchmark design, but whether trained models generalize without per-session fine-tuning — a requirement that would be impractical in clinical use — is unclear from the available text.

**Ecological validity of stimuli.** "Neuropsychology-informed paradigm" suggests structured tasks, likely involving reading or imagined speech. Whether this transfers to spontaneous communication intent is a separate question.

---

## Industry and Clinical Translation Implications

The practical stakes here are substantial. Every currently approved or investigational high-performance neural speech decoder relies on invasive recording: [BrainGate Consortium](https://bciintel.com/companies/braingate) trials use intracortical Utah arrays, [Synchron](https://bciintel.com/companies/synchron)'s Stentrode is endovascular ECoG, and [Precision Neuroscience](https://bciintel.com/companies/precision-neuroscience)'s Layer 7 Cortical Interface is subdural. These approaches achieve the signal fidelity necessary for high-throughput decoding, but they require neurosurgery — a barrier that excludes large populations of ALS patients and those with progressive neurological disease who cannot safely undergo craniotomy or endovascular procedures.

A validated, non-invasive EEG-based communication system would dramatically expand patient access. The catch is that the spatial and temporal resolution gap between scalp EEG and intracortical or ECoG recording is not a benchmark design problem — it is physics. COFETT addresses whether we are measuring EEG2Text models correctly; it does not resolve whether EEG contains enough information to support high-throughput communication decoding in the first place. The authors' claim that EEG "truly contains decodable linguistic information" is a position the field has debated seriously, and COFETT's evidence will need independent replication.

For the non-invasive BCI commercial space — companies building consumer or clinical EEG systems — this work signals that the evaluation methodology used in academic EEG2Text research has been systematically flawed. That is useful to know. It means prior negative results may reflect bad benchmarks rather than fundamental signal limitations, which is modestly encouraging. But it also means the field is further from clinical translation than published metrics have suggested.

---

## Key Takeaways

- **Teacher-forcing dependency is the central flaw in EEG2Text research**: models evaluated with teacher-forcing are not tested under real-world conditions and cannot generate meaningful output without it.
- **COFETT**, built on a 128-channel high-density EEG cap, is the first benchmark specifically designed to enable teacher-forcing-free evaluation and account for EEG signal instability.
- **The benchmark is open-sourced**, enabling independent replication — a meaningful step for a subfield where proprietary data has limited progress.
- **Clinical translation barriers remain**: scalp EEG signal resolution, cross-session generalization, and bits-per-second throughput at usable rates are not resolved by improved benchmark design alone.
- **Patient access angle**: non-invasive EEG-based communication BCIs would serve patients who cannot tolerate surgery, a large and underserved population currently excluded from high-performance neural communication systems.
- **Independent replication needed**: a team-designed benchmark demonstrating feasibility for the team's preferred approach requires external validation before conclusions about EEG linguistic decodability can be drawn.

---

## Frequently Asked Questions

**What is teacher-forcing and why does it break EEG-to-text models?**
Teacher-forcing is an evaluation method that feeds a model the correct previous word at each decoding step rather than the model's own previous output. In EEG2Text, this means the model never has to decode a full sentence from scratch — it gets hints at every step. In real use, no correct tokens are available, so models dependent on teacher-forcing fail to produce meaningful text.

**What is COFETT?**
COFETT (Corpus OF Eeg-To-Text) is a new open-source EEG2Text benchmark developed by researchers from Harbin Institute of Technology, Nanjing University, and Zhongguancun Laboratory. It uses a 128-channel high-density EEG cap and is designed around a neuropsychology-informed paradigm to enable teacher-forcing-free model evaluation and better account for EEG signal instability.

**Can EEG actually decode language, or is the signal too noisy?**
This is the central unresolved question in the field. Scalp EEG is spatially blurred by skull and scalp tissue, limiting resolution compared to ECoG or intracortical recording. The COFETT paper argues existing negative results partly reflect flawed benchmarks rather than fundamental signal limitations, but this claim requires independent replication. The physics of EEG signal propagation remain a hard constraint.

**How does EEG compare to invasive BCIs for communication?**
Invasive systems — intracortical arrays used in BrainGate trials, Synchron's endovascular Stentrode, Precision Neuroscience's subdural arrays — achieve substantially higher signal fidelity and decoding throughput. They require surgery, which limits the patient population. EEG is non-invasive and scalable but has lower signal resolution. A validated EEG communication BCI would serve patients who cannot undergo surgery, which is why the field keeps pursuing it despite the signal challenges.

**When might EEG-to-text be clinically viable?**
No timeline can be responsibly stated based on current evidence. COFETT establishes better evaluation methodology and provides feasibility evidence for teacher-forcing-free decoding, but clinical viability requires demonstrated accuracy at communicatively useful rates across sessions and participants, followed by regulatory evaluation. The field is at early feasibility stage for non-invasive neural text decoding.