Brain-Computer Interfaces and AI in 2026: What Brainwave Decoding Can Actually Do

A brain-computer interface (BCI) records electrical activity from the brain and turns it into a command, a cursor movement, a word or a sound. The recording either comes from electrodes placed inside the skull or from sensors on the scalp, the second being electroencephalography, or EEG. AI does the translation. In practice that means a small neural network trained on one person's signals, often followed by a language model that picks the most likely words. How well that works depends far more on where the sensor sits than on the size of the model, and that single fact explains most of what is real and most of what is hype in 2026.

• 97.5%word accuracy an implanted speech BCI sustained over 8.4 months of daily conversation in one man with ALS³
• 65%average character error rate for Meta's best decoder using scalp EEG while people typed⁵
• $89.1M2025 revenue of Ceribell, whose AI detects seizures in hospital EEG⁷

What implants can do with AI in 2026

The strongest evidence comes from a handful of academic studies in people with ALS or brainstem stroke who can no longer speak. Electrodes in or on the speech areas of the cortex record the attempt to speak, a recurrent neural network turns those signals into sounds of speech every few tens of milliseconds, and a language model assembles words. The best results are close to everyday use. Natural conversation runs at about 160 words per minute, and the fastest decoders now reach around half of that.¹

Cursor control has moved further. Neuralink reported on October 1, 2026 that one participant set a record of 11.32 bits per second on its target-clicking test.⁴ The company puts able-bodied people using a mouse at around 8 to 10 bits per second, reports finger-mapped typing at up to 40 words per minute, and counted 21 participants worldwide in January 2026.¹² These are company figures on a company test, published on its own site, and they have not been peer reviewed.

The AI layer is where the cost of running these systems lies. Neural signals drift from day to day, so decoders need regular recalibration. Neuralink says its users averaged 55 minutes a week of calibration just to keep control, and that pretraining an encoder on more than 50,000 hours of unlabeled brain data cut that burden for some users from 10 minutes a day to 10 minutes a week. In the same post it says decoders built from models pooled across participants "perform no better than their single-participant counterparts" in live use.⁴ A decoder that works out of the box for a new person does not exist yet, for implants or for anything else.

The Kunz study also matters for privacy. The team showed that inner speech could be decoded while people silently recalled a sequence or counted, and built safeguards, including a password the user has to think before decoding starts, to stop a speech BCI from decoding private inner speech by accident.¹¹ Any product built on implant data has to design for that risk.

Where the implant companies stand

The field moved from university labs into company trials in 2025 and 2026. As of October 2026 no permanently implanted BCI has US marketing authorization, and the first approval for commercial use anywhere came from China.

Approval is only the first wall. The US Government Accountability Office found in December 2024 that BCIs "are not yet on the market", that experts find it "challenging to interact with CMS" about coverage, and that some trial participants have had devices removed because no funding or medical support continued after the trial.¹⁹ Cigna's 2026 coverage policy lists implanted BCIs for imagined speech as not medically necessary.²⁰ China is treating the field as industrial policy, with a seven-ministry plan that targets breakthroughs in key BCI technologies by 2027 and a secure industrial system by 2030.²¹

Long-term safety data are reassuring and thin. Across 14 people in the BrainGate trial between 2004 and 2021, implants accumulated 12,203 days with 68 device-related adverse events, 6 of them serious.²² Neuralink reported that in its first participant "a number of threads retracted from the brain" in the weeks after surgery, reducing performance until it changed its recording algorithm.²³

What headsets and scanners can decode

Scalp EEG, magnetoencephalography (MEG) and functional MRI are where the mind-reading headlines come from, and they hold up least well under careful testing. The skull blurs and weakens the signal, and muscles, eye movements and walking add noise many times larger than the brain activity a model is trying to read.

Meta's Brain2Qwerty is the best non-invasive language result. In its 2025 preprint, 35 volunteers typed memorized sentences while their brain activity was recorded, and the model averaged a 32% character error rate with MEG against 67% with EEG.²⁴ The peer-reviewed 2026 version, trained on about 22,000 sentences from nine people, reported 29% with MEG and 65% with EEG. Press coverage quotes the authors that it cannot yet support real-time communication and that locked-in people who cannot move are unlikely to benefit, because the model reads the motor act of typing.⁵ MEG also needs a magnetically shielded room, so the headset form factor sits on the weaker side of that comparison.

Functional MRI decoders recover meaning, and only from willing people. A University of Texas decoder trained on hours of a person's brain scans produced text that closely matched the meaning of a story about half the time, worked only with cooperative participants, and gave unusable output for people it had not been trained on.²⁵ A 2025 follow-up cut the training needed for a new person to about an hour.²⁶ Image reconstruction from brain scans looks more impressive than it is: a Kyoto group argues the realistic pictures are mostly classification into trained categories plus images invented by the diffusion model.²⁷

EEG-to-text is the weakest link. A 2024 audit found that earlier models were scored with the correct previous word fed in at every step, which inflated results, and that models given pure noise instead of EEG scored about the same.²⁸ The models had learned the training sentences, with the brain data adding little.

Why EEG accuracy claims shrink under testing

Much of the high accuracy in EEG papers comes from letting the same person, or the same recording, appear in both training and test data. On the DEAP emotion dataset, six deep learning models scored at least 35.71 points higher on valence and 25.00 points higher on arousal under that kind of leakage than in a clean setup.²⁹ On the SEED emotion dataset, a standard model reached 0.81 accuracy on people it had seen and 0.53 on people it had not, while fitting its training subjects at 0.999.⁶ Between 10% and 30% of users cannot produce the signals a motor-imagery BCI needs at all.³⁰

EEG foundation models, pretrained on up to 61,415 hours of recordings, have not changed this yet. A 2026 benchmark found that with frozen features only five of ten such models beat a small task-specific network, and most of their gains came only after full fine-tuning on labeled data.³¹ A separate 2026 audit found that one leading model's embeddings identified which dataset a recording came from with an AUROC of 1.000, and that on an external dementia cohort, classical hand-built features scored 0.734 against 0.699, 0.669 and 0.568 for three foundation models.³²

Where the money is

The products that earn regulatory trust and revenue score a defined clinical signal against a reference that doctors already accept. Ceribell sells a rapid-response EEG system with an AI seizure detection algorithm for hospitals. It reported 2025 revenue of $89.1 million, up 36%, an 88% gross margin and 647 active accounts, and guided to $111 million to $115 million for 2026. Its algorithm is cleared for patients from pre-term neonates to adults.⁷ Persyst's seizure and spike detection software was cleared in December 2022,³⁴ and autoSCORE, which reviews EEG recordings, received a further clearance in April 2025.³⁵ Beacon Biosignals' Dreem 3S, an at-home headband with six EEG electrodes, was cleared in September 2023,³⁶ and in December 2024 the FDA authorized a predetermined change control plan that lets Beacon improve its sleep-staging algorithm without a new submission each time.³⁷ That precedent matters for any team shipping a medical model that has to keep learning.

Outside the clinic the evidence thins out. The largest meta-analysis of neurofeedback for ADHD, covering 38 randomized trials and 2,472 participants, found no significant improvement on probably blinded ratings of symptoms.³⁸ In a large test of brain and body measures for TV advertising funded by the Advertising Research Foundation, traditional survey measures were by far the best predictors of real-world ad response, and once they were included only fMRI added significant predictive power.³⁹

At work, the one lasting EEG product we found is SmartCap, a fatigue monitor worn in caps and hard hats in mining, now part of Hitachi Construction Machinery's Wenco. Its 2026 page calls the approach "scientifically validated" and cites no study or outcome data.⁴⁰ Driver monitoring at scale went to cameras: Seeing Machines alone counted 8,216,143 cars on the road with its camera-based system in June 2026, ahead of the EU's General Safety Regulation mandate.⁴¹ In schools, a Chinese trial of attention-tracking headbands was suspended in 2019 after 88% of 3,362 people in an online poll called the monitoring unnecessary or unacceptable.⁴²

Big technology companies are entering at the input layer. Apple announced in May 2025 that iOS, iPadOS and visionOS would support a protocol for brain-computer interfaces in Switch Control,⁴³ and Synchron showed a person with ALS controlling an iPad through it that August.¹⁵ Meta's Neural Band, which ships with its display glasses, reads muscle signals at the wrist (EMG), and Meta got it to work across people without calibration by collecting training data from thousands of participants, a scale no EEG dataset approaches. It handwrites at 20.9 words per minute.⁴⁴

Market size figures vary by a factor of a thousand because firms mix revenue with theoretical addressable markets. A revenue-based estimate from Precedence Research puts the BCI market at $2.94 billion in 2025 and $3.33 billion in 2026.⁴⁷ The largest rounds went to implant makers, and one tracker counted $1.63 billion raised across 26 disclosed neurotech deals in the twelve months to July 2026.⁴⁸ Investors are paying for a future medical-device market; today's revenue comes mostly from non-invasive and clinical products.

The law on neural data

A product that reads brain signals meets three bodies of law at once: privacy law for the data, AI law for what the model infers, and medical device law for what the product claims. All three moved in 2025 and 2026, and none of them uses the same definition of neural data.

Outside workplaces and schools, emotion recognition is classed as high-risk under the EU AI Act, and the Digital Omnibus package moved those obligations from August 2, 2026 to December 2, 2027.⁵⁶ European data protection authorities say neurodata "often" count as special categories of personal data, whose processing is prohibited in principle with limited exceptions.⁵⁷ UNESCO adopted the first global Recommendation on the Ethics of Neurotechnology in November 2025.⁵⁸ India's Digital Personal Data Protection Act has no separate sensitive-data tier, so neural data falls under its general consent rules, with penalties of up to ₹250 crore for failing to keep personal data secure.⁵⁹

In the US, medical claims decide which regime applies. Implanted BCIs follow the FDA guidance of May 2021,⁶⁰ connected devices must meet the cybersecurity requirements in the FDA's June 2025 guidance,⁶¹ and the January 2026 revision of the general wellness policy keeps low-risk wellness products outside device review.⁶² The same EEG headband is a wellness product when it claims to help you focus and a medical device when it claims to detect ADHD or epilepsy.

How neurotech companies handle the data

The Neurorights Foundation reviewed the policies of 30 consumer neurotechnology companies in 2024. It found that 29 of them appear to have access to users' neural data with no meaningful limits, and that 29 can and may transfer it to third parties. Only six mentioned encryption, and only five committed to notifying customers of a breach.⁶³

The attacks published so far are research demonstrations, and they show what a careless product would expose. In 2012, researchers used a consumer EEG headset and the brain's response to familiar stimuli to cut the uncertainty of private information, such as PIN digits and known faces, by 15% to 40% compared with random guessing.⁶⁴ For implants, Yale researchers recommend non-surgical ways to update and recover devices, strong authentication for software changes, and encryption of data moving to and from the brain.⁶⁵ For a software team the practical list is short: process on the device where possible, encrypt everything that leaves it, store raw signals only with a stated purpose, and never let a model infer more than the user agreed to.

How to decide whether to build, buy or stay out

Brain signals and AI are producing real results, at the two ends of the field. Inside the skull, small personal decoders with a language model behind them give people with paralysis back speech and control of a computer. In hospitals, models that score seizures and sleep against an accepted clinical reference have earned clearances and revenue. The middle ground of headsets that claim to read focus, mood or thoughts has the weakest evidence and the fastest-growing legal exposure, and companies are better served by the ends.

This is how we approach signal-based AI in our AI development work: start from a labeled outcome the business already trusts, test on people the model has never seen before anyone quotes a number, keep raw data on the device where we can, and check the claims against the privacy, AI and device rules of every market the product will reach.