The Word He Cannot Find

Your dad had a stroke in March. He knows who you are, what day it is, and he knows exactly what he wants from the kitchen. The word will not come out.

So you run through the list. Water, tea, his glasses, the last snack? He shakes his head at each one, and you both pretend it is a game.

Now picture a different visit. He lies in a hospital scanner, thinking about what he wants, while you watch a screen next door. A sentence types itself out. The wording is a little off, but the meaning is close enough. Something warm to drink, perhaps some tea. He didn't even have to move his lips to ask.

So far, this method in the University of Texas' new patent has only been tested on healthy volunteers, after sixteen hours each in the scanner.

The University of Texas System, based in Austin, wants to turn brain activity recorded from outside the skull into running text. The inventors are Alexander Huth, a neuroscientist who moved to UC Berkeley in July 2025, and Jerry Tang, who built the decoder as a PhD student and is still at UT.

Start with the autocomplete on your phone. It guesses your next word from the words you have already typed. This system makes the same kind of guess, then checks each guess against your brain.

First comes training. You lie in an fMRI scanner, which tracks blood oxygen as a stand-in for neural activity, and listen to about sixteen hours of spoken stories. Software learns how roughly 10,000 points in your cortex respond to the meaning of each phrase. The filing calls this the encoding model.

Blood is slow, though. A burst of neural activity shows up as a signal that rises and falls over about ten seconds, and people speak more than two words a second, so every brain image is a blur of twenty-odd words.

So the system guesses. A language model proposes likely next words for each sentence it is holding. If one candidate is "I saw a big", the options might be "dog" and "truck". The encoding model predicts what your brain would look like had you heard each version, and the system compares those predictions with the real scan.

It keeps the 200 best matches, drops the rest, and repeats word by word.

What comes out is the gist. One volunteer heard "I don't have my driver's license yet" and the decoder wrote "she has not even started to learn to drive yet". Across a test story, 72 to 82% of moments decoded better than chance.

The model tracks meaning, so it works when you're not even thinking of speaking but imaging. Volunteers imagined telling one of five stories and the decoder identified the right one every time. They watched silent films and it described the scenes.

One of the claims then reaches past the hospital. It covers fNIRS, a cap that measures the same blood signal with light. The inventors reduced their scans to that resolution and found that about half of the moments were still decoded.

If it works as described, a person could get a sentence out after a long training session in a scanner, in place of an operation.

Brain-to-text already works if you accept surgery. Implanted electrodes have reached 92% sentence-level accuracy, as The Register notes. They also need a neurosurgeon, and the filing points out that scarring can degrade the signal over time, which means going back in.

From outside the skull, earlier decoders could identify single words or pick one answer from a short list.

That leaves a large group with no good option. More than 2 million Americans live with aphasia and about 180,000 acquire it each year, mostly after a stroke, according to the National Aphasia Association.

The accurate option needs a neurosurgeon, and the safe option cannot hold a sentence.

The field splits by where the sensor sits. Implants are accurate and need surgery. Everything outside the skull is safer and far weaker. This patent sits in the second group and claims software, so its closest rivals are other decoders.

Meta is the nearest. Brain2Qwerty v2, published in June 2026, decoded typed sentences at 61% word accuracy across nine volunteers, The Register reported. It reads the motor signals of typing, so the person has to type, and it relies on MEG, a magnetic scanner that stays in the lab. UT decodes meaning, which is harder and needs no movement.

MindPortal is the closest technical cousin. Its MindSpeech model decoded imagined sentences from a portable fNIRS headset in four people, per Unite.AI. The scores sat a few points above shuffled data.

Sabi, a California startup backed by Khosla Ventures, says its beanie of 70,000 to 100,000 EEG sensors will type thoughts at about 30 words a minute and ship by the end of 2026. Inside BCI notes that no peer-reviewed data backs that yet.

The implant makers are moving toward decoding too. Synchron is opening a "cognitive AI" division to build thought-decoding models, Fierce Biotech reported.

Each company here is building a sensor and expecting the decoding to follow. UT holds claims on one decoding loop and owns no sensor. Whether Meta or Sabi would ever need that loop is unproven, because Meta's decoder takes a different route and maps brain signals straight to characters.

The direct market resists measurement. Grand View Research puts non-invasive brain-computer interfaces at US$397.6 million in 2025, rising to US$773.8 million by 2033. MarketsandMarkets estimates the entire category, including implants, at US$262 million in 2024. Precedence Research has it at US$2.62 billion for the same year.

The clearer market is the one this would serve first. Communication aids for people who cannot speak, known as AAC devices, were worth US$2.32 billion in 2025 and should reach US$3.97 billion by 2030, according to The Business Research Company. Those products are tablets and eye trackers, and they need a user who can still pick the words.

Morgan Stanley sized implantable brain interfaces at US$400 billion in the US across roughly 10 million candidate patients, as Forbes and Yahoo Finance reported. The same report expects about US$1 billion in annual revenue by 2041. Most of the distance between those figures is the surgery.

A non-surgical decoder aims at that distance and carries its own cost. Research time on a 3T scanner runs US$800 to US$1,350 an hour at the University of Pennsylvania, so sixteen hours of training comes to US$12,800 to US$21,600 before the first sentence. That is a long onboarding flow. The inventors have since cut training to about an hour of silent videos, per UT Austin.

UT's leverage is a licence. Its tech transfer office lists the method for speech restoration and for consumer brain interfaces. Its weakness is that the lead inventor now works in California, and know-how tends to travel with the person.

The people who need this most are a small medical market, and the companies with money are building consumer headsets. A university can license to either, so which one gets the exclusive?

Investors are paying for sensors, and mostly for the ones that go inside the skull.

Neuralink raised a US$650 million Series E in June 2025. Synchron, whose implant reaches the brain through the jugular vein, raised US$200 million in November 2025, led by Double Point Ventures, for US$345 million in total, per Fierce Biotech. Precision Neuroscience, which lays a thin electrode film on the brain's surface, closed a US$250 million Series D on 25 September 2026, led by Pershing Square and the Ackman Oxman Institute, MassDevice reported.

That is US$1.1 billion across three rounds.

Outside the skull, the cheques are fewer and tend to follow hardware or a famous name. Merge Labs, co-founded by Sam Altman to build ultrasound-based interfaces, raised a US$250 million seed at an US$850 million valuation in January 2026, with OpenAI writing the largest cheque, per TechCrunch. Becker's put the round at US$252 million.

Hemispheric, a Tel Aviv company training an AI model on EEG recordings to help diagnose brain disorders, left stealth in July 2026 with US$52 million, per Business Wire. Sabi has not disclosed its round. Kernel, which makes an fNIRS helmet of the kind claim 3 describes, raised US$53 million in 2020, TechCrunch reported, and sells it for research and clinic use.

There are a few potential issues here, across three scenarios.

The first is the vulnerable patient, who cannot speak proficiently for themself. The decoder writes fluent sentences, and part of every sentence comes from the language model. A reader cannot tell which part came from the brain. The filing's own table puts the word error rate at 92 to 94%, against 96% for text generated with no brain data at all.

If this system could be used by a carer who reads "she wants to go home" and acts on it, the person who cannot speak is the one least able to say the machine got it wrong. Their days could be shaped in serious ways by whatever comes out. Who needs to be able to affirm the accuracy level for this software, applied to patients producing sentences that can't be strongly verified, and to what standard should it be held to?

The second is consent. A volunteer can agree to be scanned, but they cannot choose which thoughts turn up while the scanner is running. Can anyone meaningfully consent to handing over information they cannot predict or inspect first?

In the EU, probably not. GDPR requires consent to be specific and tied to a stated purpose, and "whatever you find" is neither.

The US gives more leeway. Colorado, California, Montana and Connecticut have neural data statutes with different definitions, per Inside BCI, and the federal MIND Act has not advanced, as Davis Wright Tremaine sets out. However, the brain data is a whole other problem, and a Neurorights Foundation review found 29 of 30 consumer neurotech companies placed no meaningful limits on their own access to users' brain data, KFF Health News reported.

The third is national security. Governments have wanted a way into an unwilling mind for decades, and one of the earliest patents in this field belongs to the US Air Force.

Human rights law stands in the way. Freedom of thought under the International Covenant on Civil and Political Rights is absolute, with no exception for emergencies, and the Geneva Conventions bar any coercion to extract information from prisoners of war. The UN Human Rights Committee explains that the right to freedom of thought means no one can be compelled to reveal their thoughts, and running a decoder on an unwilling subject would do exactly that.

For now, the technology also stands in the way of reliably being used in contexts to reveal any national secrets stored in foreign enemies' brains. Decoders trained on other people's brains performed barely above chance, and volunteers could spoil the output by silently naming animals. Huth told UC Berkeley this year that people can shut the decoder out, "but it's pretty effortful".

For a family like the one at the top, the near-term change is small. A hospital with a research scanner might one day offer a session where a relative who cannot speak gets a rough sentence onto a screen. Jerry Tang is working with Maya Henry at UT's Dell Medical School to test the decoder with aphasia patients, UT Austin says.

The paperwork is further along than the science. The application was allowed in June 2026 without a rejection, and UT paid the issue fee on 18 September, according to Patently-O. A grant normally follows within weeks. After that, the signals to track are a named licensee and a result on a real fNIRS headset.

This week's patent is US 2025/0068841 A1, titled "Decoding Language From Non-Invasive Brain Recordings", published by the Board of Regents, The University of Texas System.

Read the filing, privacy section included, then reply and tell us who should hold the licence. We are on Instagram and LinkedIn.

•  The paper behind the patent with Nature Neuroscience: Read the 2023 study the claims are built on, including the resistance experiments that the filing uses as its privacy defence.

•  The one-hour version with Current Biology: See how the same inventors moved a decoder between people using silent films, and judge how much of the original privacy argument still holds.

•  Who holds the claims on mind reading with Patently-O: Explore the patent landscape around this application, from an early Air Force patent to the Siemens "universal decoder" and a Berkeley patent that lapsed.

•  Meta's typing decoder, in code with Meta AI Research: Compare a rival architecture that skips the propose-and-check loop and decodes keystrokes from MEG.

•  Neural data law, state by state with Davis Wright Tremaine: Learn where US neural data rules stand and why the definitions disagree on what counts as a brain signal.

•  Can anyone own a thought? with Stanford Law School: Zoom out to the 2025 UNESCO recommendation and the limits of property law when the asset is decoded speech.