An AI thinking cap that turns thoughts into text

If you wanted to turn your thoughts into text in the past, you basically had two bad options: undergo brain surgery to implant chips like Neuralink, or lie motionless inside a noisy multi-million dollar MRI scanner.

Now, researchers in Australia might have found a much less terrifying middle ground.

Scientists at the University of Technology Sydney have developed an AI-powered thinking cap. It reads EEG brainwave signals straight through the scalp and turns those silent thoughts into written sentences on a screen.

Teaming up deep learning with LLMs

Reading brainwaves from outside the skull is notoriously difficult because human bone and tissue muffle electrical signals into noisy static.

To solve this, lead UTS researcher Chin-Teng Lin and his team split the task in two. First, a deep learning model translates the raw brain signals into intended individual words. Then, a large language model steps in to assemble those words into coherent sentences, actively correcting decoding errors along the way.

It is essentially like having an AI autocorrect that understands grammar and context watching over an inaccurate transcriptionist.

Right now, the system hits an accuracy rate of around 75%. The UTS team says they are actively tweaking the system to push that figure closer to 90%.

Mohit Shivdasani, a bioelectronics researcher at the University of New South Wales, noted that the core goal here is addressing unmet medical needs and restoring lost functions, after which the sky is the limit.

Why it matters

For stroke victims and patients suffering from severe speech impairments, this could mean regaining a voice without going under a surgeon's knife.

While studies in 2023 proved that invasive implants could restore speech to paralyzed patients, surgical risks and hefty hospital bills prevent them from scaling. A non-invasive cap brings assistive tech much closer to actual home care.

From an AI perspective, using an LLM as an error-correction layer over raw physical sensors is a clever shortcut. Scalp EEG will never be as clean as an intracortical wire, so using language context to bridge the fidelity gap makes immense sense.

Still, we think that shortcut comes with questions. When an LLM fixes broken signals, how do you verify if the output is the user's authentic thought or just the model hallucinating the most probable sentence? For medical and everyday communication, proving intent without bias will be the real hurdle to clear before this hits clinic shelves.