AI Can Read Your Mind: Here's Exactly How Far It's Gone

A team at the University of Texas at Austin trained a large language model on 16 hours of fMRI scans from three human subjects, then pointed it at a new story the subjects were silently reading. The model generated text. About half the time, it matched the meaning of what was going through their heads. Not the words: the meaning. Published in Nature Neuroscience, led by grad student Jerry Tang and Prof. Alexandra Huth, this wasn't a trick. It was proof of concept that the semantic maps inside GPT-style models and the semantic maps inside human brains are measurably, usably the same. The bottleneck is now engineering, not physics.

The Race That Most People Haven't Noticed

The UT Austin experiment works in three stages. Subjects spend 16 hours inside an fMRI machine while listening to narrative podcasts. The scanner records blood-oxygen-level-dependent (BOLD) signals in the brain, not individual neurons, but clusters of roughly 100,000 neurons at once. A language model, trained on the same audio, learns to map those fMRI patterns to semantic meaning. When a new, unseen story is played, the decoder reconstructs the approximate meaning of what the subject is hearing.

The accuracy numbers: 72–82% for perceived speech (something the subject is actually hearing), 41–72% for imagined speech (something they're picturing in their head), 21–45% for silent video clips, where the decoder is attempting to reconstruct what subjects are watching without any audio. Those numbers drop sharply once you remove the training requirement. The decoder works only for the specific subjects who spent 16 hours in the scanner. Apply it to a stranger and you get gibberish.

On the other side of the invasive/non-invasive divide, two separate teams published results in the same stretch of 2023 that are harder to dismiss. At Stanford, a patient called Pat, ALS patient, speech and facial movement compromised, had microelectrode arrays implanted in her speech motor cortex. The array reads from 100+ individual neurons. A language model with a 100,000-word vocabulary interprets those signals. Pat now communicates at 62 words per minute. Her previous best, with eye-tracking technology: roughly 10 words per minute.

At UCSF, a second team achieved 78 words per minute using a surface electrode array (ECoG) and a separate avatar speech synthesis layer that produces a synthesized voice in real time. The patient had a brain stem stroke. Both systems are active clinical trials, not consumer products, not FDA-cleared devices, but real people getting real utility from real brain implants attached to real language models.

That's a 5× improvement in two years. Moore's Law for paralyzed people. The AI component isn't incidental: the jump from 15 to 62–78 words per minute happened precisely because language models got good enough to fill in context, correct neural noise, and handle the ambiguity of reading electrical signals from a brain.

What the Surveillance Framing Gets Wrong

The fMRI decoder is not surveillance. Not yet, and not for obvious reasons. It requires the subject's willing participation for 16+ hours of labeled data collection inside an fMRI machine: a device that costs $3–7 million, requires a specially shielded room, and demands the subject stay nearly motionless. And even then, it works only for the specific person who donated the training data. The decoder trained on Pat's brain reads nothing from yours.

But Dr. Huth's lab noticed something the surveillance framing obscures entirely. To build the decoder, they needed to know what "understanding language" looks like in both an LLM and a brain. What they found: the context embeddings in GPT-style language models: the internal representations of how words relate to each other, predict brain activity patterns with striking accuracy. The brain and the transformer are not merely analogous. They're measurably, usably similar.

That finding cuts both directions. It suggests LLMs are doing something cognitively real, not just statistical pattern-matching, but something structurally closer to the kind of semantic encoding biological brains do. And it suggests that as models improve, their semantic maps may become better proxies for human neural patterns, eventually making decoders more general, not more individual-specific.

Jean Remi King (Meta FAIR) describes the current limitation as "interindividual variability of neural representations": each brain maps the same concept slightly differently, based on lifetime experience and development. That's the bottleneck. It's real. But it's an engineering problem, not a physics problem. And meanwhile, the law hasn't started the conversation. Chile passed the world's first constitutional neurorights amendment in 2021. Colorado enacted the first US neural data privacy law in 2024 (HB24-1058). The other 49 states have nothing. No EU equivalent. No UK equivalent.

The Consumer Path: Meta's MEG Decoder

Meta's MEG-based brain decoder is the non-surgery version of this technology. Magnetoencephalography captures magnetic fields produced by electrical activity in neurons, at millisecond resolution, roughly 1,000× faster than fMRI, without requiring the subject to lie motionless in a $5 million machine. MEG helmets exist that look something like bike helmets. The spatial resolution isn't as sharp as fMRI, but the temporal precision is dramatically better.

The architecture Jean Remi King's team built has three components: an image encoder that compresses visual information into embeddings; a brain encoder that learns to map MEG signals to those same embedding spaces; and an image decoder that reconstructs the original visual content from the brain-aligned embedding. The result: a system that can reconstruct, in real time, approximately what a subject is looking at. Not perfectly. Not reliably. But the direction is clear.

The invasive path (Neuralink, Synchron, BrainGate) involves implanted electrodes and surgery. The non-invasive path (MEG, fMRI, EEG) involves wearable or external sensors with lower resolution but no surgery. The speed benchmark below maps both paths against each other:

The Right You Don't Know You're About to Lose

The most useful framing for this technology didn't come from a research paper. It came from a video discussing brain-data privacy: "If your credit card gets swiped, you can cancel it. But you can't just change your neural data. That data is you."

A field called neuro law is forming to address this. Brain scans, in criminal proceedings, may already be more diagnostic than lie detectors. Researchers as far back as 2018 were demanding new human rights legislation covering when and how brains could be read, before the 2023 wave of published BCI results, and before LLMs got good enough to decode meaning rather than just categories.

Hans Berger recorded the first human EEG in 1924. For 100 years, we had no way to act meaningfully on brain data at scale. Now we do. The window between "feasible to decode" and "regulated" is open, and nobody knows how long it stays open.