# Non-Invasive Brain-Computer Interface: What Does 94% Mean?

[Neural Tech](https://mindoxai.com/category/blog/neural-tech/)
**
Published: November 12, 2026
**
8 min read
**
[Pravesh Garcia](https://mindoxai.com/author/pravesh/)

![Non-Invasive Brain-Computer Interface: What Does 94% Mean?](https://d1p58i0iwq4isz.cloudfront.net/mindoxai/wp-content/uploads/2026/11/11064930/non-invasive-brain-computer-interface-what-does-94-mean.webp)

Would you wear a device that listens to your brain for 14 hours straight? That’s the quiet question behind this week’s headline about a non-invasive brain-computer interface that reportedly passed 94% accuracy in “mind typing.” No surgery. No implant. Just a cap.
The number may well be real. What it measures is much narrower than the headlines suggest. And the most interesting part of the story isn’t the accuracy at all.
It’s the comfort.
Before anyone calls this a mind-reading cap, the claim needs some unpacking.

## What the researchers reported, and what’s still unverified

According to [Dataconomy](https://dataconomy.com/2026/11/09/chinese-bci-electrode-hits-94-mind-typing-accuracy-in-day/), Chinese researchers built “gradient” electrodes and fitted them into a commercial EEG cap. The design targets an old tension between electrodes that are uniformly soft and ones that are rigid. In day-long use, the team’s non-invasive brain-computer interface scored over 94% accuracy in visual mind-typing tests.
The headline figures:

**14 hours** of testing while participants rested
- **Over 94% accuracy** on the visual typing task
- **8 hours** of stable recordings while one subject went about daily activities

A second outlet, kantan.news, adds detail. It describes the electrodes as gradient hydrogel, and says the mobile test had the subject walk, eat, read and sleep. It also names a journal (npj Flexible Electronics) and three institutions: Guangxi University, Guangxi Medical University and Tsinghua University.

Now the caveats. There are several.

Both articles appear to rewrite a single South China Morning Post report. That makes this one original account, not three independent confirmations. We couldn’t find the paper itself. The journal and university names come only from kantan.news, a small aggregator, and Dataconomy mentions neither.

The gaps go further:
- No outlet says how many people took the 94% test.
- None names a comparison electrode beyond the commercial cap.
- None gives an information transfer rate, which would show how fast the system works, not just how often it’s right.
- Nobody has replicated the result.

So treat it as promising and provisional. One report says it happened. That’s where the evidence sits today.

## Why a comfortable non-invasive brain-computer interface matters

Most coverage of brain interfaces chases decoding power. More words per minute, lower error rates, bigger vocabularies. This story points at something duller and possibly more important.

Can you keep the sensor on?

The electrode is where skin meets machine. The team frames the problem as a trade-off: comfort on one side, recording stability on the other. Rigid and uniformly soft electrodes each give something up, and the gradient design tries to sit between them.

That matters because a BCI you can only stand for a short session is a lab demo. A non-invasive brain-computer interface you can wear through lunch, a book and a nap starts to look like a tool. If the 14-hour and 8-hour figures hold up, they’re the real news here.

Scalp recordings have always been the noisy, accessible end of brain sensing. Our breakdown of [EEG vs fMRI vs ECoG](https://mindoxai.com/blog/neural-tech/eeg-vs-fmri-vs-ecog/) explains why reading the brain through bone and skin costs you signal.

## What “mind typing” means in this test

This is where the headline oversells.

During the 14-hour rest test, participants wrote number sequences from 1 to 9, plus dates. They did it by focusing on flashing characters on a virtual keyboard, according to kantan.news. The system works out which character the person is looking at. Nobody is decoding sentences someone silently composes.

That places the study among visual-selection spellers, often called SSVEP-style systems, though the coverage doesn’t name the exact paradigm. In setups like this, the software infers your target from how your brain responds to the flashing.

Two things follow.

**First, picking from a few targets is a far easier problem than open-ended typing.** With roughly ten options, random guessing lands about one time in ten. So 94% sits well above chance. It also sits a long way from reading thoughts.

**Second, the typing accuracy comes from the rest test.** The 8-hour mobile trial measured signal stability. None of the reports say anyone typed while walking or eating.

I’d call this a comfort result with a typing demo attached. That’s still worth something. It just isn’t telepathy.

## Non-invasive BCI vs implants: where each one wins

The honest comparison is with implanted BCIs, and the gap is wide.

In June 2026, a UC Davis team in the BrainGate2 trial published results in Nature Medicine from Casey Harrell, a 47-year-old man with ALS ([EurekAlert](https://www.eurekalert.org/news-releases/1132151)). He has four arrays carrying 256 electrodes in his cortex. Over nearly two years, he used the system at home for more than 3,800 hours.

His numbers:
- An average of **56 words per minute**
- **Over 99% word accuracy** in controlled testing, with a 125,000-word vocabulary
- **92% of sentences** rated accurate or mostly correct, by his own judgment

That is open-vocabulary communication, not picking digits off a grid. Any non-invasive brain-computer interface starts behind here, because it listens from outside the skull.

Implants pay for that edge with brain surgery. The UC Davis result also comes from one participant, and the device remains investigational.

Even inside the skull, limits remain. In a 2025 Cell paper, Stanford researchers decoded imagined sentences from four people with paralysis. They reached up to 74% accuracy, drawing on a vocabulary of up to 125,000 words. The team says the system can’t yet decode free-form inner speech without substantial errors.

So don’t line up 94% against 99% and call it a near tie. One is selection among a handful of flashing targets. The other is free-vocabulary speech. We dug into the wider trade-offs, including surgical risk and signal quality, in [Invasive vs Non-Invasive BCIs: The Real Trade-offs](https://mindoxai.com/blog/neural-tech/invasive-vs-non-invasive-bcis/).

## How it compares with Meta’s Brain2Qwerty

For genuine typing, Meta’s research is the closest test of what a non-invasive brain-computer interface can decode. In the [Brain2Qwerty study](https://arxiv.org/abs/2502.17480), 35 healthy volunteers typed while the system tried to decode their keystrokes from brain activity.

The results split sharply by sensor:
- With **MEG** (magnetoencephalography), the average character error rate was 32%.
- With **EEG**, it was 67%.
- The best participants got down to 19%.

A 67% character error rate means EEG missed about two of every three characters. That’s the same sensor family the Chinese team used.

A second version, covered by The Batch in July 2026, cut word error rate to 39% from 43%. It used MEG only, with nine subjects and 90 hours of data. The authors say their error rates still don’t match the single-digit rates of implants.

There’s one more catch. Brain2Qwerty participants were physically typing, so the decoder likely draws on movement-related activity. It isn’t reading silent thought either.

Put the two side by side and the picture clears up. The Chinese electrode doesn’t solve EEG’s decoding problem. It may solve EEG’s wearing problem. Those are different bottlenecks, and both need fixing before a cap replaces a keyboard.

## Who reads a brain signal you wear all day?

Here’s the human stake the accuracy figure hides. A non-invasive brain-computer interface built for 14 hours of wear isn’t a lab session anymore. It’s ambient.

The coverage already floats sleep monitoring and seizure detection as possible clinical uses, though nobody has tested those yet. Both imply recording for long stretches, including while you sleep.

Implant researchers have started designing for this. “This is the first time we’ve managed to understand what brain activity looks like when you just think about speaking,” said Erin Kunz, lead author of the Stanford inner-speech study. Her team also built a password keyword that gates decoding. The system recognized it with over 98% accuracy.

A visual speller is far less intimate than inner speech. Still, an all-day EEG stream likely captures much more than the digits you meant to type. Who stores it? Who can ask for it? We looked at what the state could legally demand in [Neuroprivacy: What Governments Can Really Subpoena](https://mindoxai.com/blog/neural-tech/neuroprivacy-what-governments-can-really-subpoena/). Those questions get sharper once the hardware is comfortable enough to forget.

## What to watch next

A few things would turn this from a promising headline into solid evidence:
- **The paper itself.** Until it surfaces, the journal and institutions stay unconfirmed.
- **Participant numbers.** A 94% average across many people means more than one person’s best day.
- **The baseline.** How did standard electrodes do on the same test over the same 14 hours?
- **Typing on the move.** Accuracy during the 8-hour mobile session, not just signal stability.
- **Speed.** An information transfer rate, so the result can sit next to other spellers. Our look at [Neuralink’s 11.32 bits per second](https://mindoxai.com/blog/neural-tech/neuralink-bits-per-second/) shows why a speed figure tells its own story.

My read: this is the right kind of progress, aimed at an unglamorous problem. Going by Meta’s EEG numbers, scalp caps won’t match implants on raw decoding any time soon. But plenty of people will never accept brain surgery. For them, a cap they can forget about may matter more than a few points of accuracy.

Which brings us back to the opening question, with a twist. If a non-invasive brain-computer interface gets comfortable enough to wear all day, how much of your brain are you comfortable sharing?

Author & Contributor

### Pravesh Garcia

Writes explainers and analysis for Mindox AI on AGI, artificial intelligence, neural technology and cognitive augmentation. Articles are based on primary sources such as research papers, company releases and regulatory documents, and follow the [Editorial Policy](https://mindoxai.com/editorial-policy/).

[Follow](https://mindoxai.com/author/pravesh/)

## Frequently Asked Questions

**Q:** Can you type with your mind without surgery?

**A:** In a limited way, yes. EEG caps can let you pick characters by focusing on flashing targets on a screen, and a new Chinese report claims over 94% accuracy on that kind of task. Decoding free typing from EEG is far harder: in Meta's Brain2Qwerty study, EEG got a 67% character error rate.

**Q:** How accurate is non-invasive BCI typing?

**A:** It depends on the task. The new gradient-electrode cap reportedly scored over 94% when users selected flashing digits at rest, though that comes from a single unverified report. When Meta tried to decode what people actually typed, the average character error rate was 67% with EEG and 32% with MEG.

**Q:** What is a non-invasive brain-computer interface?

**A:** It's a system that reads brain activity from outside the skull, for example with electrodes in a cap on the scalp, and turns those signals into commands or text. No surgery is involved, but the signal is weaker and noisier than an implant's.

**Q:** Can EEG read your thoughts?

**A:** Not in the open-ended sense. The 94% result came from users focusing on flashing characters on a virtual keyboard, not from decoding sentences they silently composed. Even implanted inner-speech decoders at Stanford reached only up to 74% accuracy, with four participants.

**Q:** Is a non-invasive BCI as good as a brain implant like Neuralink?

**A:** Not for communication yet. A UC Davis implant user reached over 99% word accuracy in controlled tests and an average of 56 words per minute. That required brain surgery, though, and comes from one participant in an investigational trial.

**Q:** Why are EEG electrodes uncomfortable to wear for long?

**A:** Designers face a trade-off between comfort and recording stability. Uniformly soft electrodes and rigid ones each give up something, which is the problem the new gradient electrode design is meant to address.