Here is the markdown body only for the annual review:
BCI Annual Review — 2025
1 January–31 December 2025
Introduction
2025 was the year invasive brain–computer interfaces moved from research and early trials into regulated devices, first-in-human milestones, and mainstream visibility. A second commercial cortical BCI path opened when the FDA cleared Precision Neuroscience’s minimally invasive implant in April, and the first U.S. clinical test of BCI for speech and language decoding ran at WVU’s Rockefeller Neuroscience Institute. Paradromics completed its first-in-human Connexus procedure in June; Neuralink closed a $650M Series E, expanded to the Miami Project, and by September reported 12 human implants and a planned speech-impairment trial. By year-end, Neuralink had submitted its first peer-reviewed human data and reported a 10,000-person waitlist, Synchron’s Stentrode and Ceribell’s point-of-care EEG had been named to Time’s Best Inventions of 2025, and the University of Michigan Health had opened one of the first dedicated BCI clinics in the U.S. The narrative was no longer “if” but “how fast” and “along which paths”—with Precision, Paradromics, Synchron, Blackrock Neurotech, and Neuralink all advancing human trials or commercial use.
Speech and communication BCIs defined the year’s application story. A streaming brain-to-voice neuroprosthesis in Nature (March), motor cortex encoding of handwriting via stable neural states, and Rush University’s “Finding a New Voice” work set the stage. Stanfordand BrainGate-led real-time inner-speech decoding (August) drew broad coverage and pushed mental privacy and consent into public debate. First-in-human results for Precision’s minimally invasive BCI, Neuralink’s speech trial plans, and Science Corp’s patient registry for sight-restoring BCI filled out the autumn. November brought real-time decoding of full-spectrum Mandarin from neural signals in Science, Paradromics’ FDA approval for the Connect-One speech BCI trial, and progress toward decoding individual words from non-invasive recordings. The pipeline from overt to mimed to imagined speech, and from English to tonal languages, was clearly in motion.
Sensory restoration and closed-loop systems ran in parallel. January’s Nature paper on a closed-loop sensory neuroprosthesis improving stumble recovery in lower-limb loss, bionic hands with tactile shape and movement feedback, and the LUKE Arm’s first real-world trial established that closing the loop with sensation was a priority. Movement-responsive deep brain stimulation with a remotely optimized decoder (June), USC’s wireless AI-driven implant for chronic pain, and an EEG-based system for real-time robotic hand control at individual finger level showed decoding and stimulation converging. A wireless subdural BCI with 65,536 electrodes and 1,024 channels in Nature (December) and an ultra-thin implant (BISC) with tens of thousands of electrodes and AI decoding of movement, perception, and intent set new hardware benchmarks and pointed to epilepsy, paralysis, and vision as near-term targets.
Regulatory and policy attention caught up with the technology. The FDA cleared Precision’s implant and, later, ONWARD’s ARC-EX for home use; Paradromics received an FDA IDE for the Connect-One speech trial; and the first U.S. tDCS-based depression treatment was approved in December. Ethics and governance advanced in step: a Frontiers paper on bridging ethics and regulatory oversight in implantable BCI human subjects research (July), the Barcelona AIMD workshop on active implantable devices, the ICRC’s question of whether BCIs can comply with international humanitarian law (August), and a U.S. Congressional neural data bill (September). October saw U.S. senators propose the MIND Act to study national standards for protecting consumer neural data, the World Economic Forum’s technology-neutral neuro-privacy and “neurosecurity stack” framing, and commentators highlighting the lack of international model law for neurotechnology. UNESCO adopted the first global standard on the ethics of neurotechnology in November, and Nature asked whether devices that infer preconscious states need new safeguards. By year-end, clinical pathways, evidence expectations, and data protection were being debated in the open.
Timelines
January–March. The year opened with sensory neuroprosthetics and real-world validation: a closed-loop sensory neuroprosthesis for stumble recovery in Nature, bionic hands that let users feel shape and movement on the prosthetic “skin,” the LUKE Arm’s first day-to-day trial at the University of Utah, and Blackrock Neurotech arrays in a BCI enabling finger-level control from decoded motor intent. Clinical Neurophysiology carried a wave of work on implant durability (silicon IC degradation, PDMS barriers), iEEG hippocampal memory neurofeedback, multi-branch CNNs plus intracranial high-frequency oscillations for seizure prediction, and an international update on TMS–EEG. February centered on invasive BCI and next-generation interfaces: intracortical design in chronic stroke, multiscale spiking decoders, Stim-BERT for reconstructing iEEG during stimulation, tensor methods for neural dynamics, and a Nature review of soft, biohybrid, and “living” neural interfaces. March was dominated by communication and motor restoration: a streaming brain-to-voice neuroprosthesis and motor-cortex encoding of handwriting in Nature, Rush’s “Finding a New Voice” speech work, a paralyzed patient controlling a robotic arm with imagined actions, spinal cord stimulation plus rehabilitation robotics for cycling and walking, and a proof-of-concept of real-time invasive recording during downhill skiing in Parkinson’s. New America’s analysis of neurotech growth and brain-data privacy and security risks put policy on the map.
April–June. April delivered two headline regulatory and clinical steps: Precision Neuroscience received FDA clearance for its minimally invasive brain implant, establishing a second commercial cortical BCI path alongside Neuralink, and WVU Rockefeller Neuroscience Institute reported the first successful U.S. test of BCI to decode speech and language. The month also featured multimodal EEG+fNIRS for semantic decoding, NLP-revealed neural dynamics of conversation, transparent PEDOT:PSS interfaces for electrophysiology and two-photon imaging, clinical translation of ultrasoft Fleuron probes, Paradromics’ new chief medical officer, a sham-controlled RCT of home-based tACS in Alzheimer’s, and VA-backed individualized neurofeedback for concussion. May was defined by BCIs reaching mainstream platforms and non-invasive expansion: Synchron announced the first native BCI integration with Apple’s iPhone, iPad, and Vision Pro; a University of Michigan team reported the first in-human recording with a new wireless BCI; Coinbase co-founder Fred Ehrsam launched the non-invasive BCI startup Nudge; and Cognixion and Blackrock Neurotech expanded access to multi-modal non-invasive research with Axon-R. Methods advances included alignment of latent dynamics to stabilize BCIs over time, quantification of crosstalk in brain recordings, open iEEG datasets for epilepsy and conscious perception, and the Neuroelectrophysiology Analysis Ontology (NEAO). June put the invasive BCI race in full view: Paradromics completed its first-in-human Connexus BCI procedure, Neuralink closed a $650M Series E and added a paralyzed veteran implant at the Miami Project, and Blackrock Neurotech (Tether-backed) was framed as a more advanced alternative to Neuralink in public statements. Research highlights included ECoG syntactic role encoding during sentence production, movement-responsive DBS with a remotely optimized neural decoder in Nature, USC/UCLA’s wireless AI-driven implant for chronic pain, and an EEG-based system enabling real-time robotic hand control at individual finger level in Nature. Myndlift passed one million neurofeedback sessions; UConn launched a neuromodulation center of excellence for veterans.
July–September. July emphasized neural decoding generalizability—inter-individual and inter-site neural code conversion without shared stimuli, contrastive learning for cross-subject EEG emotion recognition, and hybrid deep learning for EEG-BCI classification—and breadth in neuromodulation: holographic transcranial ultrasound, phase-synchronous tACS for memory, HD-tDCS for insomnia, anodal tDCS for rationality, and minimally invasive extra-epidural electrodes for pain. Onward submitted ARC-EX to the FDA for home use and for a CE mark; Paradromics added an investigator as its BCI trial neared launch; and a Frontiers paper plus the Barcelona AIMD workshop addressed ethics and regulatory oversight for implantable BCI and active implantable devices. August was dominated by inner-speech BCI and mental privacy: Stanfordand BrainGate-led real-time inner-speech decoding from intracranial signals was reported as a communication tool for people with speech impairment and was covered by The New York Times and Scientific American, triggering debate on mental privacy and consent. Stentrode-based control of consumer devices, Paradromics’ first-in-human milestone and trial prep, Neuralink’s first participant narrative, and Corticale’s ultra-dense SiNAPS design underscored multiple invasive paths. ONWARD received FDA IDE for the ARC-IM Empower BP pivotal study; a scoping review on ethical gaps in closed-loop neurotechnology and an ICRC piece on BCIs and international humanitarian law sharpened the policy frame. September put invasive BCI scale and speech decoding at the center: Neuralink reported 12 human implants and a planned trial for speech impairments; Precision shared first-in-human results for its minimally invasive cortical interface; Science Corp opened a patient registry for its sight-restoring BCI. Research advanced linguistic neural decoding with deep learning, transfer learning from distributed brain recordings for reliable speech decoding, and mimed speech as an intermediary between overt and imagined speech in ECoG. ONWARD received CE Mark for ARC-EX in Europe and reported ARC-IM data showing improved blood pressure control after spinal cord injury. Congress introduced a neural data bill; UNESCO and the World Economic Forum addressed neurotechnology ethics and governance; and a Nature paper outlined clinician perspectives on explainability in AI-driven closed-loop neurotechnology.
October–December. October was defined by invasive BCI stepping into the evidence and spotlight phase, regulatory and policy focus on neural data and neurosecurity, and a strong run of methods and interfaces. Neuralink submitted its first peer-reviewed human clinical data and reported a 10,000-person waitlist; Synchron’s Stentrode and Ceribell’s point-of-care EEG were named to Time’s Best Inventions of 2025. U.S. senators proposed the MIND Act to study national standards for protecting consumer neural data; the World Economic Forum advanced technology-neutral neuro-privacy and a “neurosecurity stack”; and commentators pointed to the lack of international model law for neurotechnology. Research included iEEG speech timing, an EEG imagined-speech database for paradigm comparison, silk-based conformal intraventricular interfaces, invasive mapping for personalized OCD neuromodulation targets, and a hair-thin electrode design that roughly tripled recording duration. MindMaze’s $1.1B merger with Relief and NeuroX and ONWARD’s revenue milestone and ~€51M raise underscored consolidation and execution. November put speech and communication BCIs at the centre: Science reported real-time decoding of full-spectrum Mandarin from neural signals; Nature showed progress toward decoding individual words from non-invasive brain recordings; Paradromics received FDA approval for the Connect-One trial of its Connexus speech BCI. UNESCO adopted the first global standard on neurotechnology ethics; the MIND Act framed the U.S. debate on neural data; and Nature asked whether devices that infer preconscious states warrant new safeguards. Synchron raised $200M Series D; analyst reports had invasive BCI funding passing $1B. ONWARD received FDA clearance expanding ARC-EX for home use; Ceribell won FDA clearance for its neonatal seizure-detection algorithm. December put hardware scaling and the clinic in the spotlight: a wireless subdural BCI with 65,536 electrodes and 1,024 channels in Nature set a new ECoG-style density benchmark; an ultra-thin implant (BISC) with tens of thousands of electrodes and AI decoding of movement, perception, and intent was reported with early relevance for epilepsy, paralysis, and blindness. Nature framed the strategic picture with “Brain–computer interfaces race to the clinic.” The first U.S. approval of a tDCS-based depression treatment, the first dedicated BCI clinic at University of Michigan Health, a Neuralink recipient with ALS gaining communication or control capability, Ceribell’s FDA 510(k) for EEG-based delirium monitoring, and an AdvaMed IDE workshop marked regulatory and care-delivery milestones. Neuralink hired the former FDA division lead who had overseen the company; Science Corp added a chief strategy officer; Neurable closed a $35M Series A; Merge Labs (Sam Altman–linked) spun out of a nonprofit; and Mindmaze made a volatile Swiss exchange debut. Methods included long-term unsupervised recalibration for intracortical cursor BCIs, neural decoding of mental imagery with a general semantic model, and open EEG and ECoG datasets for speech, tactile perception, and semantic alignment.
Trends
Invasive BCI: multiple paths to the clinic
The year consolidated several commercial cortical and endovascular BCI paths. Precision Neuroscience’s FDA clearance (April) established a second cleared commercial cortical interface alongside Neuralink; Paradromics’ first-in-human Connexus procedure (June) and later FDA approval for the Connect-One speech trial (November) put a high-channel-count cortical array on a clinical speech path. Neuralink scaled implants (12 by September), raised $650M, added the Miami Project as a trial site, submitted first peer-reviewed human data, and reported a 10,000-person waitlist. Synchron announced native BCI integration with Apple devices (May), raised $200M Series D (November), and had its Stentrode named to Time’s Best Inventions of 2025. Blackrock Neurotech (Tether-backed) was positioned as a more advanced alternative to Neuralink in public statements and expanded non-invasive research access with Cognixion via Axon-R. Corticale’s ultra-dense SiNAPS interface (August) and the December Nature wireless subdural BCI with 65,536 electrodes and the BISC ultra-thin implant illustrated that hardware scaling—density, wireless operation, and minimally invasive form factors—was advancing in lockstep with trials and approvals.
Speech and language decoding: from lab to first-in-human trials
Speech and communication BCIs moved from foundational science to first-in-human and regulated trials. Nature and related work established streaming brain-to-voice decoding, motor cortex encoding of handwriting as sequences of stable neural states, and ECoG encoding of syntactic roles and structure-dependent timing during sentence production. WVU’s first-in-U.S. BCI test for speech and language decoding (April), Stanford/BrainGate real-time inner-speech decoding (August), and Neuralink’s planned speech-impairment trial (September) were followed by real-time decoding of full-spectrum Mandarin in Science (November), Paradromics’ FDA-approved Connect-One speech BCI trial, and progress toward decoding individual words from non-invasive recordings. Open datasets—EEG and ECoG for speech, articulation and coarticulation, natural language comprehension (“Podcast” ECoG), and imagined speech—supported reproducible benchmarks. Mimed speech as an intermediary between overt and imagined speech in ECoG, transfer learning from distributed brain recordings for reliable speech decoding, and an EEG imagined-speech database for paradigm comparison extended the pipeline from invasive to non-invasive and from English to tonal languages.
Closing the loop: sensation, decoding, and adaptive stimulation
Sensory restoration and closed-loop neuromodulation were consistent themes. January’s closed-loop sensory neuroprosthesis for stumble recovery, bionic hands with tactile feedback, and the LUKE Arm trial framed sensation as critical for real-world gait and grasp. Movement-responsive DBS with a remotely optimized neural decoder (June), the USC/UCLA wireless AI-driven chronic pain implant, and intracranial stimulation for psychiatric circuit mapping (June) showed decoding driving stimulation parameters. Brain decoders controlling spinal cord stimulation, artificial tactile feedback for BCI users exploring digital objects, and high-gamma ECoG representing perceived vibration intensity in somatosensory cortex (July) supported closed-loop sensory and motor interfaces. Adaptive DBS with improved artifact removal (December), invasive mapping for personalized OCD neuromodulation targets (October), and thalamic and focused-ultrasound neuromodulation (November) extended the trend toward closed-loop, decoding-informed stimulation across invasive and non-invasive modalities.
Decoder generalizability and calibration-light pipelines
Methods aimed at cross-subject and long-term stability advanced. Alignment of latent dynamics to stabilize BCIs over time (May), inter-individual and inter-site neural code conversion without shared stimuli (July), contrastive learning for cross-subject EEG emotion recognition (July), and hybrid deep learning for EEG-BCI classification supported more scalable, calibration-light or plug-and-play–style pipelines. Circuit-level modeling of crosstalk contamination in brain recordings (May), the Neuroelectrophysiology Analysis Ontology (NEAO) for data sharing (May), and temporal coding for more stable cortical visual representations than firing rate (August) addressed validity and reproducibility. Long-term unsupervised recalibration for intracortical cursor BCIs using a hidden Markov model (December) and adaptive EEG preprocessing to mitigate electrode-shift variability for motor imagery (November) targeted real-world usability and robustness.
Non-invasive BCI and neurofeedback at scale
Non-invasive BCI and neurofeedback reached finer control and larger deployment. An EEG-based system for real-time robotic hand control at individual finger level (Nature, June), transformer-based and temporal-fusion-attention motor imagery decoding (June, September), and low-latency handwriting recognition from EEG on an edge device (November) showed non-invasive motor interfaces closing in on dexterous control. Simultaneous EEG+fNIRS for semantic and motor-imagery decoding, wearable and dense fNIRS for real-time and precision neuroimaging, and 40-class SSVEP spellers with beta-range stimulation (November) expanded the non-invasive toolkit. Myndlift passed one million neurofeedback sessions (June); VA-backed individualized neurofeedback for concussion (April), cognition-guided EEG neurofeedback for response inhibition (June), and neurofeedback for executive function in ADHD and for treatment-resistant PTSD (July, September) illustrated clinical and institutional adoption. Muse (Interaxon) expanded its AI sleep platform (October); Neurable raised $35M and launched smart headphones with brain health tracking (November, December).
Neuromodulation: approvals, home use, and evidence
Spinal cord and transcranial neuromodulation had a strong regulatory and clinical year. Onward submitted ARC-EX to the FDA for home use and for a CE mark (July), received CE Mark for ARC-EX in Europe (September), and received FDA clearance expanding ARC-EX for home use (November); the company also received FDA IDE for the ARC-IM Empower BP pivotal study (August) and reported ARC-IM data showing improved blood pressure control after spinal cord injury (September). The first U.S. approval of a tDCS-based depression treatment (December) was a landmark for transcranial stimulation. Evidence accumulated for tDCS (dysphagia, depression, cognition, fibromyalgia, post-stroke motor recovery), tACS (memory, insomnia, Alzheimer’s), theta-tACS for sleep in insomnia, HD-tDCS for insomnia and conflict processing, and personalized and precision TMS; the RATES Delphi consensus on reporting standards for transcranial electrical stimulation (September) supported reproducibility and trial design. Holographic transcranial ultrasound (July), focused ultrasound for deep brain circuits (August), and low-intensity focused ultrasound for thalamic conscious perception and nucleus accumbens reward sensitivity (November) extended non-invasive neuromodulation to deep targets.
Ethics, policy, and neural data protection
As commercial BCIs and neurotech devices advanced, ethics and policy gained structure. New America’s analysis of neurotech growth and brain-data privacy and security risks (March), a Frontiers paper on bridging ethics and regulatory oversight in implantable BCI human subjects research (July), the Barcelona AIMD workshop on active implantable devices (July), and Unicef’s neurotechnology and children’s rights framing (July) were followed by the ICRC’s question of BCIs and international humanitarian law (August), a scoping review on ethical gaps in closed-loop neurotechnology (August), and Congress’s neural data bill (September). October and November brought the proposed MIND Act to study national standards for protecting consumer neural data, the World Economic Forum’s technology-neutral neuro-privacy and “neurosecurity stack,” UNESCO’s first global standard on the ethics of neurotechnology, and Nature’s question of whether devices that infer preconscious states warrant new safeguards. Secure pairing for implantable devices using vibration and motor-behavior entropy (August), clinician perspectives on explainability in AI-driven closed-loop neurotechnology (September), and the AdvaMed IDE workshop (December) connected policy to device design and regulatory practice.
Hardware and interfaces: durability, density, and minimally invasive design
Implant durability, density, and less invasive design progressed across the year. Silicon IC degradation in the body and PDMS elastomer barriers (January), mechanically adaptive and resveratrol-eluting probes (July), and material and design strategies for chronically implantable neural probes (August) addressed long-term stability. Transparent PEDOT:PSS for electrophysiology and two-photon imaging (April), clinical translation of ultrasoft Fleuron probes (April), silk-based conformal intraventricular interfaces (October), and a hair-thin electrode extending recording duration roughly threefold (October) emphasized minimally invasive and chronic recording. The December wireless subdural BCI with 65,536 electrodes and 1,024 channels and the BISC ultra-thin implant with tens of thousands of electrodes set new density and scalability benchmarks; scalable high-density cortical microelectrode arrays, spike compression for high-density implants, and monolithic multimodal probes for sustained recording and stimulation (September) pointed to smaller, denser, and more sustainable implants for the clinic.
Suggested Titles
- Invasive BCIs race to the clinic: Precision, Paradromics, Neuralink, and the year of first-in-human milestones
- From inner speech to Mandarin: how speech and communication BCIs defined 2025
- 65,000 electrodes, UNESCO ethics, and the MIND Act: hardware, policy, and neural data in 2025
- Closing the loop—sensation, movement-responsive DBS, and the convergence of decoding and stimulation
- Second cortical path, first BCI clinic, first tDCS for depression: regulatory and care milestones in 2025