Scientists use GPT to read minds non-invasively

Until now, helping paralyzed people "speak" usually meant a neurosurgeon had to cut open their skull and plant electrodes directly on the brain.

Researchers at the University of Texas at Austin just found a way around the operating table. In a new study published in Nature Neuroscience, the team demonstrated an AI language decoder that non-invasively translates human brain activity into text using an fMRI machine and a transformer model similar to the technology behind ChatGPT.

Reading the gist, not the muscle

Existing brain-computer interfaces typically focus on motor control—recording neural impulses sent to the lips, tongue, or vocal cords as someone tries to speak.

UT Austin took a completely different path. Their system monitors changes in blood flow through the brain to capture the abstract meaning of ideas. As senior author Alexander Huth explained, the decoder operates at the level of concepts, semantics, and meaning rather than low-level motor function.

To build it, three volunteers each spent 16 hours inside an fMRI machine listening to narrative podcasts. The team paired these brain scans with a GPT-1 model trained on Reddit posts and autobiographical stories, teaching the AI to associate specific brain patterns with words and phrases.

When tested on new stories the model had never seen, the decoder could reconstruct what participants were hearing in real time. It didn't copy the audio word-for-word, but captured the essence. When a participant heard "I don't have my driver's license yet," the decoder produced: "She has not even started to learn to drive yet."

Decoding silent films and pure thoughts

The team also pushed the decoder beyond external audio. When volunteers watched silent films or simply imagined a story in their heads, the AI could still decipher the general storyline.

In one test, a participant imagined the phrase: "went on a dirt road through a field of wheat and over a stream and by some log buildings." The decoder translated that internal visualization into: "he had to walk across a bridge to the other side and a very large building in the distance."

Audio decoding proved more reliable than pure imagination, but the fact that unspoken thoughts could be translated into coherent sentences without any physical movement is a massive breakthrough.

Why it matters

The immediate promise of this technology is giving a voice back to stroke victims and patients suffering from ALS without invasive, high-risk neurosurgery. We think using transformer models to read thoughts at the semantic level is a brilliant pivot—it bypasses the physical limits of damaged motor pathways entirely.

However, you will not see this in clinics tomorrow. An fMRI machine is massive, wildly expensive, and requires a patient to lie motionless for hours. The researchers hope future setups might work with wearable near-infrared spectroscopy (fNIRS) caps, but whether those sensors can deliver the resolution GPT needs remains an open question.

Then comes the obvious ethical elephant in the room: mental privacy. Today, the system only works if a user voluntarily sits still and cooperates for 16 hours of calibration. But the authors themselves warned that future iterations could bypass these constraints, raising real risks of surveillance by employers or governments. In our view, this study should serve as a wake-up call for regulators—the boundary around our private thoughts is no longer entirely impenetrable.