Brain Implants Reach the Market Before the Rules Do

Brain-computer interfaces crossed from investigational device to registered product this year, with China approving the first invasive implant and US speech neuroprostheses proving they hold up at home for years. What has not kept pace is everything around the hardware: who pays, who owns the signal, and whether consumer earbuds collecting the same data face any rules at all.

Key points

• China approved the first invasive BCI for commercial use in March 2026 and gave it an insurance code within days; no US device has a premarket approval yet.¹
• AI-decoded speech now works outside the lab: one ALS participant has used an implant near-daily for two years at 99% word accuracy and kept a full-time job.²
• US reimbursement is the binding constraint: a former CMS chief medical officer says no Medicare coverage framework exists for the implants nearing approval.³
• The same EEG sensor ships as an FDA-cleared medical device and an uncleared consumer earbud, and consumer EEG recordings identify individuals at better than 90% accuracy.⁴⁵
• Five US states and UNESCO now name neural data as a protected category, but the strongest enforcement provisions were stripped before passage and the global standard is non-binding.⁶⁷

What's changing

Implants cross from trial to product

China's NMPA approved Neuracle's eight-electrode NEO implant in March 2026, the first invasive BCI cleared for use outside a trial anywhere, on the same day the 15th Five-Year Plan named BCI a priority industry.¹ Behind it sits a state-coordinated pipeline of 134 registered trials and five approved products, 79% of them non-invasive rehabilitation systems for stroke.⁸ The US path runs through the FDA instead: Synchron raised $200M to fund the first pivotal trial designed around a premarket approval filing,⁹ Paradromics placed its first permanent 421-electrode speech implant under an IDE,¹⁰ and Neuralink says it will move to high-volume production and near-automated surgery this year with roughly two dozen participants across four countries.¹¹ The direction is clear and the evidence is strong. The quality of what gets approved is less clear: NEO is an eight-channel device with modest functional gains, and Neuralink's production targets come from a CEO post, not peer review.¹¹¹

AI decoding makes restored speech durable

The strongest single result of the year is longevity. Casey Harrell, implanted in 2023 with 256 electrodes, used a UC Davis speech neuroprosthesis for more than 3,800 hours at home over nearly two years, producing close to two million words at 56 words per minute and 99% accuracy without researcher setup.² Paradromics reports its first participant answering open-ended questions within three months of implantation.¹⁰ Non-invasive decoding trails far behind but is moving the same way: the 2026 PNPL benchmark targets word classification from MEG with ten minutes of per-user calibration, built explicitly so new paralysed users need minutes rather than hours.¹² We rate the implanted-speech evidence as emerging rather than established; it rests on one long-term participant and one company press cycle. The dependence on language-model decoders also opens a question nobody has answered: whose words are produced when the model fills in an uncertain phoneme.²

The sensor leaves the clinic

Naox received FDA clearance for a single-channel in-ear EEG in January and previewed a consumer earbud, explicitly not cleared, for late 2026, alongside NextSense and Neurable.⁴ Merge Labs, with OpenAI as the largest investor in a $252M seed, is building a non-implanted ultrasound interface it says will serve patients first and the general public second; OpenAI framed BCIs as a way for anyone to interact with AI.¹³ The medical and consumer versions of these devices collect comparable signals under different law: HIPAA for one, FTC Section 5 for the other.⁴ That split matters more now that a 14-channel consumer headset can identify individuals from short recordings regardless of mood, which makes de-identification promises in consumer neurotech privacy policies hard to keep.⁵

Rules lag the hardware

Lee Fleisher, CMS chief medical officer until 2023, warns that implanted BCIs are approaching FDA approval with no Medicare coverage precedent, because trials measure cursor control rather than the health outcomes Section 1862 requires.³ On data, Vermont became the fifth state with a neural-data statute in May, after dropping the consent requirement, private right of action and per-violation penalties from the original bill.⁶ UNESCO adopted the first global neurotechnology ethics recommendation in November 2025, but it is non-binding and, as a Stanford Law analysis argues, sidesteps ownership entirely: consent fails because users cannot know what their neural data reveals, and property framing invites its sale as a condition of employment or insurance.⁷ The default owner of neural data today is whoever wrote the device's terms of service.

What to watch

• Whether Synchron's pivotal trial reports function-based outcomes that CMS can score as medically necessary; if it does not, approval without coverage becomes the likely US outcome.⁹³
• Peer-reviewed safety and durability data from Paradromics and Neuralink, especially device longevity past two years, which the Chinese landscape study already flags as a barrier to routine use.⁸¹⁰¹¹
• Whether any state restores a consent requirement and private right of action for neural data before Vermont's 2028 consent rule takes effect, and whether a federal bill appears.⁶
• Shipment of Naox Wave and the first Merge Labs product roadmap: the moment a consumer device reads brain signals at scale, mental privacy stops being a hypothetical.⁴¹³