Brain-computer interfaces (BCIs) create a direct communication pathway between the human brain and external devices. By decoding neural activity into signals that can control computers, robots, or other systems, BCIs allow “thoughts” to be translated into actions or communication in real time. BCIs have enormous potential for assisting people with paralysis or neurodegenerative diseases in communicating and interacting with the world, and they also open novel possibilities for consumer technology in gaming, virtual reality, and hands-free control of devices. While some of the most dramatic BCI demonstrations have involved invasive implants placed inside the brain, non-invasive BCIs are far more practical for widespread use due to their safety and ease of deployment. This article focuses on three leading non-invasive BCI technologies – electroencephalography (EEG), magnetoencephalography (MEG), and functional near-infrared spectroscopy (fNIRS) – and how they enable real-time thought decoding. We explore their principles of operation, capabilities and limitations, and the state-of-the-art applications in both research and consumer neurotechnology. A particular emphasis is placed on how these BCIs integrate with artificial intelligence (AI) algorithms and robotics, enabling advanced systems like brain-controlled prosthetics and AI-driven communication devices. The goal is to provide a comprehensive, up-to-date overview of how EEG, MEG, and fNIRS BCIs are decoding human thoughts non-invasively and turning them into meaningful outputs, from controlling robot arms to typing out sentences by brain activity alone.
Modern advances in machine learning have significantly improved the performance of BCIs, allowing more complex mental states to be recognized quickly and accurately. In parallel, hardware improvements – from novel sensor materials in EEG caps to the miniaturization of MEG and fNIRS devices – are enhancing the comfort and practicality of non-invasive BCIs. As a result, we are seeing BCIs transition from purely laboratory research setups toward real-world use. Tech companies and startups are introducing wearable EEG headsets for consumers, and researchers are demonstrating “mind-reading” prototypes, such as AI models that convert brain signals to text. Brain-controlled robots and prosthetics, once the domain of speculative fiction, are now active areas of development in neuroengineering and rehabilitation. At the same time, these developments bring challenges regarding accuracy, training, and ethics – for instance, ensuring such systems only decode intended commands and respect user privacy.
In the sections that follow, we begin by surveying the key features of EEG, MEG, and fNIRS as non-invasive neural interfaces, comparing their signal characteristics, resolution, and practicality. We then delve into how each modality is used to decode thoughts or mental commands in real time, highlighting representative studies and breakthroughs. Next, we discuss the integration of AI techniques in processing and interpreting BCI data, which has been crucial for recent advances. We also examine the marriage of BCI technology with robotics – enabling brain-controlled wheelchairs, robot arms, drones, and other devices – as well as other research and clinical applications such as neurorehabilitation and assistive communication. Finally, we consider consumer-oriented neurotechnology products and the future outlook, addressing the remaining hurdles and promising directions for making non-invasive BCIs more powerful, accessible, and seamlessly integrated into daily life.
Non-Invasive BCI Modalities: EEG, MEG, and fNIRS
Non-invasive BCIs rely on recording brain activity from outside the skull, avoiding any surgical intervention. Three of the most widely used techniques are EEG, MEG, and fNIRS. Each of these modalities captures a different facet of brain activity – electrical signals, magnetic fields, or blood-flow changes – and comes with its own strengths and limitations for real-time decoding. A comparative overview is given in Table 1 below.
Table 1. Comparison of key non-invasive BCI modalities.
| Modality | Signal Type | Spatial Resolution | Temporal Resolution | Portability | Notable Applications |
|---|---|---|---|---|---|
| Electroencephalography (EEG) | Scalp electrical potentials | Low (cm scale) | High (milliseconds) | High (wearable caps available) | Widely used in research, consumer headsets, neurofeedback, gaming, rehabilitation |
| Magnetoencephalography (MEG) | Magnetic fields from neural currents | High at cortex (mm scale) | High (milliseconds) | Low (requires large sensors & shielding) | Research on high-fidelity decoding (e.g. language and vision), some medical diagnostic use (epilepsy); potential for future wearable MEG systems |
| Functional NIRS (fNIRS) | Near-infrared light absorption (hemodynamic response) | Moderate (millimeters, superficial cortex) | Low (seconds delay) | Moderate (portable optical headsets) | Communication for locked-in patients, workload monitoring, hybrid BCIs, portable brain monitoring in the field |
Each modality records brain activity through completely different physiological processes. EEG measures the electrical voltage fluctuations from neurons firing in the brain, detected by electrodes placed on the scalp. MEG measures the tiny magnetic fields produced by those same neural electrical currents, using extremely sensitive magnetometer sensors positioned around the head. fNIRS infers brain activity indirectly by shining near-infrared light into the head and measuring how much light is absorbed, which changes with the amount of oxygenated blood in the cortex – an indicator of neural activation (through neurovascular coupling). Because of these differences, EEG, MEG, and fNIRS vary in the kind of information they provide and how suitable they are for BCI tasks that require speed, precision, or convenience.
From a performance standpoint, EEG and MEG are similar in offering millisecond-level timing resolution, meaning they can track rapid changes in brain activity in real time as a person thinks, perceives, or moves. MEG generally has better spatial resolution than EEG – on the order of millimeters for sources in the cortex, versus centimeters for EEG which blurs together signals from larger regions. This is because the skull and tissues distort electrical signals more strongly than magnetic fields. MEG can thus pinpoint active brain areas more accurately than EEG, but it also comes with a far steeper cost and complexity: traditional MEG machines weigh several hundred kilograms and cost millions of dollars, requiring shielded rooms and cryogenically cooled sensors. EEG equipment, by contrast, is relatively inexpensive and portable, consisting of an electrode cap and amplifier (even low-cost wireless EEG headsets exist for home use). fNIRS has spatial resolution in between – it can localize activity to within a few millimeters on the cortical surface, but cannot see deep into the brain. Its temporal resolution is much lower: changes in blood flow unfold 1–5 seconds after neurons fire. This intrinsic delay limits fNIRS in fast-paced BCI control, but it can still be effective for tasks that don’t require split-second decisions.
In terms of practicality, EEG and fNIRS are the most mobile. EEG caps can be setup in different environments (some modern EEG systems even use dry electrodes requiring minimal prep), and fNIRS devices can be wearable headbands or helmets containing light sources and detectors. Both can be used with infants, children, and adults safely. MEG has traditionally been a lab-bound modality due to bulky hardware, though as we will discuss, new developments in optically-pumped magnetometer (OPM) sensors are moving MEG toward wearable systems in the future. Another practical consideration is susceptibility to motion artifacts: fNIRS is relatively robust against head movement, whereas EEG and especially MEG require the subject to minimize movement (MEG is very sensitive to any disturbance in the magnetic field, and EEG signals can be contaminated by muscle activity or eye blinks). Only EEG and fNIRS are completely safe for persons with metallic or electronic implants; MEG (and MRI) aren’t suitable for those individuals due to magnetic interference.
Each modality has found a niche in BCI research and applications. EEG is by far the most widely used for BCIs, owing to its combination of real-time responsiveness, relatively low cost, and ease of use. MEG, while less common, has been employed in research settings to push the envelope of decoding performance on complex tasks like continuous speech or imagined sentences. fNIRS has emerged as a promising option particularly for situations where electrical methods fail or where portability is needed outside the lab – for instance, enabling communication with completely locked-in patients who may not produce reliable EEG signals. Sometimes these modalities are even used together as hybrid BCIs, since EEG and fNIRS complement each other (EEG provides speed, fNIRS can add information about brain region activation). In the following sections, we examine EEG, MEG, and fNIRS in depth, looking at how each is used to decode a person’s thoughts or intentions in real time and highlighting examples of their integration with AI and robotics.
Electroencephalography (EEG) in BCIs
Electroencephalography (EEG) is the cornerstone of non-invasive BCI technology. EEG measures the electrical activity of the brain via electrodes placed on the scalp, typically detecting voltage fluctuations from synchronized neural firing (especially from cortical neurons near the surface). EEG was first used to record human brain waves nearly a century ago, and by the 1970s researchers had begun exploring EEG-based communication channels – effectively the first brain-computer interfaces. Today, EEG remains the most prevalent modality for building BCIs, whether in clinical research or consumer devices, due to its real-time responsiveness and relative simplicity. Modern EEG caps can have anywhere from a single electrode up to 256 electrodes; more electrodes generally provide more spatial detail and signal quality, but even a few channels are enough for certain BCI uses.
Real-Time Thought Decoding with EEG: The fast temporal response of EEG (millisecond precision) makes it well-suited for real-time decoding of certain patterns of brain activity. However, EEG does not literally read arbitrary “thoughts” – what it can pick up are patterns correlated with specific mental states or intentions. BCI systems typically define a set of mental commands or cues that a user can produce on demand, which have distinguishable EEG signatures. For example, one classic paradigm is motor imagery: users imagine moving their left hand versus right hand (without any actual movement), producing characteristic changes in EEG rhythms over the motor cortex (mu and beta rhythm desynchronization). By training a classifier to recognize these patterns, a BCI can distinguish “thinking about left” vs “thinking about right” and use that to drive a cursor or robotic arm in two directions. This kind of EEG-based motor imagery BCI has enabled paralyzed users to control wheelchairs or virtual keyboards with some success. Another common approach uses event-related potentials (ERPs) – involuntary EEG responses to stimuli – to select items. The P300 speller, for instance, flashes letters in a grid; when the target letter flashes, the user’s brain emits a P300 wave that the system detects, thereby identifying that letter. Using such EEG spellers, people with severe paralysis have been able to communicate slow but effective yes/no answers or spell out messages one character at a time.
Despite limited resolution, EEG has enabled meaningful communication in real time for those who otherwise could not interact. For example, in the Cybathlon BCI Race (an international competition for BCI technology), teams demonstrated that tetraplegic users could control an onscreen avatar through an obstacle course using EEG-based BCIs, by selecting discrete commands with their thoughts. Over multiple sessions, users and the systems improved together via a mutual learning process, achieving more reliable control. While information transfer rates in such EEG BCIs are modest (on the order of a few bits per second), they underscore EEG’s viability for real-time intent detection when optimized with training and feedback. Research is also pushing EEG decoding to more ambitious goals like decoding imagined speech or internal dialogue. Converting EEG signals to text is extraordinarily challenging due to EEG’s low spatial resolution and noise, but recent studies have begun to show small steps forward. For instance, some experimental systems have reported classifying a limited vocabulary of words or phrases that a user is silently speaking to themselves (i.e. “imagined speech”), using deep learning models trained on EEG data. A 2022 review noted that while EEG-based speech decoding is still far from conversational fluency, improved machine learning and larger datasets are gradually raising accuracy. Nevertheless, fully natural thought-to-text via EEG remains an unsolved problem, underlining the inherent limits of scalp electrodes – which pick up only a blurry composite of millions of neurons at once.
Strengths and Limitations: EEG’s major strengths are its speed, affordability, and portability. Changes in brain electrical activity are registered immediately by EEG, which is critical for applications like closed-loop neurofeedback or fast BCI control (e.g. reacting to a sudden event). EEG setups can be as simple as a headset with dry electrodes or as elaborate as a high-density cap with gel-based electrodes, but even the more complex systems are vastly cheaper and easier to deploy than imaging technologies like MEG or fMRI. This has made EEG the primary choice for home BCIs and commercial brain-sensing gadgets. On the downside, EEG signals are noisy and low-amplitude, typically just a few microvolts by the time they reach the scalp. They are easily contaminated by artifacts – muscle movements (facial expressions, jaw clenches), eye blinks, and electrical interference can swamp the true brain signals. A great deal of signal processing is devoted to filtering and artifact removal to improve the signal-to-noise ratio. Another limitation is low spatial resolution: an EEG electrode picks up overlapping activity from a large swath of brain tissue, and it’s hard to know exactly where a given signal originates. This makes decoding complex or fine-grained information (like distinct words, or multi-finger movements) very challenging with EEG alone. The result is that EEG BCIs often constrain the user to a small set of mental commands or rely on stimuli that evoke known responses, rather than freely decoding any arbitrary thought.
Interestingly, users can learn to generate more discernible EEG patterns with practice, and systems can adapt to users over time – a process known as co-adaptive or mutual learning. With feedback and training, a person might increase the consistency of their imagined movement patterns, and simultaneously the AI model can update to the user’s brain signal idiosyncrasies. Such adaptive calibration helps mitigate EEG’s variability and is a focus area in BCI research (for example, using transfer learning to apply one user’s model to another, or an expert’s model to a novice, jump-starting the learning process).
EEG in Consumer Neurotechnology: Because EEG can be implemented with compact, wireless hardware, it has exploded into the consumer tech space in recent years. There are numerous EEG-based devices marketed for wellness, entertainment, or education. These include simple headbands for meditation and sleep tracking (e.g. the Muse headband with 4 channels, or NeuroSky’s single-sensor devices) and more advanced multi-channel headsets for gaming or research (such as Emotiv’s 14-channel Epoch headset, or OpenBCI’s 8-16 channel open-source systems). The typical consumer EEG device has fewer electrodes than a lab system – often 4 to 16 – and uses dry contacts for convenience (at the cost of slightly noisier signals). With built-in Bluetooth and processing apps, they can stream brainwave data to a phone or computer. Common applications include neurofeedback (training users to reach a relaxed state by providing feedback on their EEG rhythms), attention monitoring, and rudimentary “mind control” interfaces for games or smart home devices. For example, a toy called the Mindflex (using NeuroSky EEG) famously let users move a ball through an obstacle course by concentrating or relaxing, which modulated their EEG and in turn controlled a fan blowing the ball. While simplistic, it demonstrated the principle to the public that brain signals could directly control physical objects.
Modern consumer BCIs are a bit more sophisticated. Some VR and AR companies are integrating EEG to create hands-free control in virtual environments. A notable example was NextMind (recently acquired by Snap), which offered a dev-kit EEG that allowed users to select menu items in an AR/VR display just by focusing attention on them – essentially using visually-evoked EEG responses as a controller. In gaming, a company Neurable demonstrated a prototype where a player in VR could perform actions via EEG-detected intent (like telekinetically picking up objects in a game using brain signals alone). Big tech firms have also shown interest: Meta (Facebook) acquired the startup CTRL-Labs, which works on a related neural interface (EMG from wrist muscles, not EEG, but under the same broader aim of interface innovation), and Valve has explored EEG integration for adapting gameplay in VR headsets. Emotiv, one of the earliest BCI companies, now markets its EEG headsets for uses ranging from improving workplace focus to controlling drones. Emotiv’s software, for instance, includes a “Mental Commands” feature where the user can train the system on distinct EEG patterns to associate with commands like push, pull, or rotate an object; these commands can then be used to interact with games or even real devices via IoT connections. The accuracy is limited and requires training each command per user, but it offers a glimpse of what practical brain control of gadgets might look like.
Crucially, consumer EEG devices still operate within the constraints of EEG physics: they can reliably detect gross states (relaxed vs attentive), rhythmic oscillations (like alpha waves indicating drowsiness), or responses to external cues. They cannot read complex thoughts or sentences from your mind – any marketing claims aside. For example, a user cannot put on a $300 headset and have their internal monologue transcribed to text. But through clever interface design, even a low-bandwidth EEG signal can be useful. One emerging area is using EEG to sense cognitive workload or emotion, for applications like adaptive tutoring systems or market research. Early studies suggest EEG features can reflect when a person is mentally overloaded, which could trigger an AI assistant to adjust the difficulty of a task in real time. In vehicles, EEG-based drowsiness detectors have been prototyped to alert drivers if their brain activity shows signs of microsleep. These are examples of passive BCIs (monitoring the user’s brain state for context) rather than active command-driven BCIs.
Overall, EEG’s ubiquity in BCIs comes from the fact that it is currently the only non-invasive modality that has the trifecta of being affordable, real-time, and easy to deploy. The trade-off is the low fidelity of information. But ongoing improvements in both hardware and AI algorithms are helping to pull more signal from the noise. High-density EEG nets combined with advanced source localization algorithms can approximate which brain region signals are coming from. Deep learning models, trained on large EEG datasets, are showing better accuracy in classifying complex states than the old statistical methods. These trends, along with hybrid EEG-fNIRS systems and integration with other sensors (like eye trackers), promise to keep EEG at the forefront of practical BCIs.
Magnetoencephalography (MEG) in BCIs
Magnetoencephalography (MEG) is a non-invasive technique that measures the magnetic fields produced by neural electrical activity. Whenever neurons fire, they generate not only electrical voltage changes (which EEG picks up) but also electromagnetic fields. MEG uses arrays of highly sensitive magnetometers to detect the faint magnetic signals emanating from the brain. Historically, MEG systems have employed Superconducting Quantum Interference Devices (SQUIDs) that operate at extremely low temperatures, necessitating liquid helium cooling and magnetically shielded rooms to block environmental noise. This made MEG an expensive, stationary technology found only in specialized labs or hospitals. However, MEG provides excellent data quality: it shares the millisecond temporal resolution of EEG while avoiding some of EEG’s blurring, because magnetic fields are less distorted by the skull. As a result, MEG can localize sources of brain activity with a precision of several millimeters on the cortex, giving it a spatial resolving power closer to fMRI but with real-time speed.
MEG for Thought Decoding: The high fidelity of MEG signals means that, in principle, MEG-based BCIs could decode more complex or subtle brain states than EEG-based BCIs. In practice, the use of MEG for BCI research has been somewhat limited by logistical issues, but there are notable demonstrations of its potential. One of the most headline-grabbing recent achievements was by researchers at Meta (Facebook) Reality Labs and collaborators, who created a system called Brain2Text (or “Brain2Qwerty”) that translates brain activity into text using MEG and AI. In their 2025 study, volunteers lay in a MEG scanner and typed sentences, while a deep neural network was trained on their MEG data to predict the text. The system relied on a hybrid architecture: convolutional neural networks (CNNs) to extract spatiotemporal features from raw MEG signals, followed by transformer models and a language model to interpret those features in context. Remarkably, this non-invasive BCI could decode the sentences the participants were typing with an average character error rate of about 32% using MEG signals alone (meaning roughly two-thirds of characters were inferred correctly). The best subjects achieved error rates as low as 19%, which begins to approach the accuracy of slow human typists. By contrast, when the team tested a similar approach with EEG, the error rates were much higher (around 67% CER). This highlighted MEG’s advantage in recording richer neural information – likely due to its ability to pick up more localized and cleaner signals from language-related brain areas, which EEG smeared or missed.
One insight from the Brain2Qwerty MEG study was evidence of the brain’s hierarchical processing of language: the model could see distinct phases in the MEG signals corresponding to the context of a sentence, the semantic meaning of upcoming words, and then the assembly of specific syllables and letters. Essentially, before a word was typed, the brain first represented the general idea/context, then the word, then broke it down into letters for motor execution. This aligns with theories of how the brain formulates speech, but MEG provided a time-resolved window into that process in action. Decoding such a sequence with EEG is far more difficult because the signals overlap in time and space. In this case, MEG’s clarity allowed AI models to start disentangling these overlapping representations.
While Brain2Qwerty was an offline decoding (it decoded sentences after they were typed, not yet live letter-by-letter), it demonstrates that non-invasive devices can achieve levels of decoding complexity previously thought to require implanted electrodes. The limitation, of course, is the hardware – a 500-kg MEG machine costing around $2 million was used. This is hardly something a patient or consumer can use daily. The Meta researchers acknowledged this and pointed to two paths forward: first, hardware miniaturization and new MEG technology to make it portable; second, combining MEG with EEG or other cheaper sensors to balance performance and cost. Notably, they achieved those error rates with MEG, but with EEG the results were poor – suggesting that for now, certain high-bandwidth “mind-reading” goals might only be met by MEG or more invasive options. It’s a reminder that while EEG will remain primary for most BCIs, MEG can be a powerful research tool to push capabilities.
Beyond text decoding, MEG has been used in research to decode other mental states: for example, researchers have used MEG to recognize imagined visual stimuli (reconstructing basic images a person is seeing or imagining, via decoding the visual cortex), to classify different cognitive tasks, and to decode aspects of speech perception or music imagery. Many studies leverage MEG to understand the brain rather than control a device – for instance, using MEG BCI paradigms to map motor cortex activity for neurorehabilitation research, or to examine how brain networks synchronize when a person is multitasking. As an example of MEG’s sensitivity: a 2021 report used MEG to decode inner speech (silently spoken phrases) and achieved significantly above-chance recognition of various phrases by analyzing the magnetic signals from auditory and motor areas involved in internal speech. Such studies hint that MEG could enable BCIs for communication that go beyond yes/no answers, potentially allowing direct neural speech prosthetics in the future, all without implants.
Emergence of Wearable MEG: A major development influencing MEG’s future in BCIs is the advent of OPMs – optically pumped magnetometers. These are small, high-sensitivity magnetic sensors that operate at room temperature, using laser-pumped atomic vapor to detect magnetic fields, and they don’t require the heavy shielding or cooling of SQUIDs. In recent years, prototypes of OPM-based MEG systems have been built that are lightweight and even wearable as a helmet. For instance, researchers at the University of Nottingham and University of Minnesota have demonstrated that a person can move their head (or even walk around) wearing a helmet with OPM sensors, and still record MEG signals, by using real-time magnetic field compensation. This mobility and adaptability (the helmet can fit children or adults, whereas traditional MEG has a fixed dewar size) open the possibility of “MEG outside the lab.” Companies such as Cerca Magnetics in the UK are now offering commercial OPM-MEG systems, aiming to make high-end brain imaging more accessible. If these wearable MEG devices become widely available and cost falls, MEG could enter the realm of practical BCIs. Imagine a portable MEG that you can simply wear like a bike helmet – it could provide high-resolution neural data on the go, which combined with AI decoders might enable capabilities far beyond what EEG headsets allow today.
That said, currently even OPM-based MEG systems are in prototype phase and not something you’d buy at an electronics store. They still require careful calibration, and sensitive control of the magnetic environment (perhaps active shielding). In the context of BCIs, one strategy being explored is hybrid MEG-EEG systems, where a few OPM sensors might target key brain regions (to enhance signal quality for those areas) while EEG electrodes cover the rest of the head – leveraging the information fusion. Another approach is using MEG in lab settings to train decoding models that later could be adapted to EEG for daily use (transfer learning). For example, one could record a rich MEG dataset of a user performing various mental commands, train a robust AI decoder, and then see if a simplified EEG setup can approximate those commands by feeding the model or using it for initialization. This is speculative but highlights how MEG can complement EEG rather than outright replace it.
MEG and AI/Robotics: Most brain-controlled robot demonstrations have used EEG, but one could use MEG in a similar fashion if the setup allowed. For example, controlling a robotic arm by motor imagery should be even more reliable with MEG due to cleaner detection of motor cortex signals. The main hurdle is that a person typically must sit still with their head inside a MEG helmet, so it’s not very practical to move around and control a robot simultaneously. However, in a purely lab setting, there have been proofs of concept – such as studies where MEG was used to control computer cursors or simple devices, basically to benchmark control accuracy under ideal signal conditions. MEG’s role is more pronounced in the decoding of high-dimensional information (like continuous hand trajectories, speech, etc.), which could feed into robotics. For instance, decoding a user’s intended 3D hand movement path from their brain could drive a robot arm to trace that path. MEG has shown better performance than EEG for such continuous movement decoding because it captures subtle neural oscillation patterns related to movement kinematics with more fidelity.
One particularly exciting application is using MEG to understand brain network dynamics during human-robot interaction – for example, measuring how a person’s brain responds in real time as they use a BCI to control a robot, potentially allowing feedback to improve the interaction. This veers into neurofeedback territory: a closed-loop where the brain, the AI decoder, and the robotic device are all influencing each other. MEG’s richer data could help researchers fine-tune such loops.
In summary, MEG is like the high-performance sports car of non-invasive BCIs: extremely powerful under the hood, capable of things EEG can only dream of, but currently expensive and impractical for everyday use. With ongoing technical advances (like OPMs) aiming to make MEG more wearable and affordable, we might see MEG playing a larger role in BCI research this decade. For now, its main contributions are in pushing the boundaries – showing what’s possible in terms of decoding accuracy and complexity – and informing the development of algorithms and hybrid systems that might trickle down to more accessible modalities. The integration of AI is essential for MEG as well, since the data richness goes hand in hand with complexity; only advanced machine learning models can fully exploit the fine-grained temporal-spatial patterns MEG captures. As one scientist noted about the Meta MEG studies, “They are definitely doing as much as we can do with current technology in terms of what they can pull out of these signals”. In the coming years, we anticipate MEG contributing to breakthroughs especially in high-bandwidth BCI communication and in understanding the brain’s coding schemes for thought, which can inspire new non-invasive BCI strategies across the board.
Functional Near-Infrared Spectroscopy (fNIRS) in BCIs
Functional Near-Infrared Spectroscopy (fNIRS) is a non-invasive optical technique for monitoring brain activity by tracking blood flow changes. fNIRS systems emit near-infrared light (typically 700–900 nm wavelength) into the scalp and skull and use detectors to measure the light that bounces back. Because oxygenated and deoxygenated hemoglobin in blood absorb light differently, fNIRS can infer the concentration of oxygenated blood in the cortex underneath each sensor. When a specific brain region becomes active, neurons consume oxygen and the local blood vessels respond by increasing flow to that area – a phenomenon known as the hemodynamic response. fNIRS captures this response, providing a signal somewhat analogous to functional MRI (which also relies on blood oxygenation levels). While fNIRS cannot see as deep or with as high spatial detail as fMRI, it has key advantages: it is portable, quiet, and tolerant of movement, making it possible to use in natural environments and with a variety of subjects.
In a BCI context, fNIRS signals are slower than EEG/MEG – often peaking ~5 seconds after the onset of neural activity. This naturally limits the speed of an fNIRS BCI, since you might have to wait a few seconds for a clear “yes” vs “no” response, for example. However, fNIRS BCIs have been very valuable for certain user groups and applications. One major area is assistive communication for completely locked-in patients. EEG-based BCIs sometimes fail with completely locked-in syndrome (CLIS) patients, perhaps because of abnormal or very weak electrical signals, whereas fNIRS has shown success in enabling basic communication in such cases. For instance, in 2017 a team led by Niels Birbaumer reported a groundbreaking result: four late-stage ALS patients in CLIS (unable to move any muscle, even their eyes) learned to use an fNIRS BCI to answer yes/no questions reliably by modulating blood flow in the frontal cortex. They would think “yes” or “no” and the system measured subtle changes in oxygenation that corresponded to their responses, achieving over 70% accuracy in communication. This was the first time fully locked-in individuals could communicate at all, albeit slowly (each response took tens of seconds). The system used a wearable fNIRS cap developed at SUNY Downstate and was eventually envisioned for home use, since it was non-invasive and relatively easy to don. The success is modest in terms of bit-rate, but enormous in terms of impact for patients who otherwise have zero channels of interaction.
fNIRS BCI Paradigms: Most fNIRS BCIs are based on sustained mental tasks that produce a hemodynamic change. A common approach is to have the user perform a mental task for a few seconds to indicate a selection. For example, to answer “yes,” the user might do a task like mental arithmetic or motor imagery which is known to increase blood flow in certain cortical areas; to answer “no,” they relax and do nothing. The fNIRS system then classifies the resulting blood oxygen level pattern as either task or rest. These tasks can be personalized to the user’s abilities—some might find it easier to sing a song in their head or visualize a scene as the active task. It’s slow (each “bit” of communication may require several seconds of task plus a rest period), but can be very robust. Studies have shown that fNIRS can achieve around 70–80% accuracy for binary choices in many patients, which, through protocols like spelling (binary tree spelling systems), can be used to spell out words one bit at a time.
Another paradigm is using cyclic changes: for instance, driving a BCI-controlled wheelchair by alternately thinking and resting in a certain pattern. One could imagine an fNIRS BCI where the user triggers “go” by performing a task (causing a detectable rise in blood flow) and triggers “stop” by resting. Indeed, researchers have built simple brain-controlled vehicle demos using fNIRS in controlled conditions. The low information rate means that fNIRS is often combined with intelligent automation – the BCI might just give high-level commands and the wheelchair’s onboard AI handles the low-level navigation.
Interactive Applications and Hybrid Setups: Despite its slow pace, fNIRS has been tested in interactive scenarios including gaming. A recent study (Ghalavand et al. 2025) benchmarked deep learning models on fNIRS signals from people playing a virtual tennis game. They classified periods of active gameplay vs rest with very high accuracy (95–97%) using methods like CNNs and LSTMs on the fNIRS time-series. This suggests that, for detecting relatively distinct mental states (concentrating on a game vs idle), fNIRS combined with modern machine learning can be extremely reliable. The result is promising for using fNIRS in real-world settings like training simulations or VR experiences, where the system can adapt based on whether the user is actively engaged or needs a break. Deep learning was particularly effective at extracting features from the fNIRS data automatically, underscoring the benefit of AI in fNIRS BCI analysis.
Hybrid EEG-fNIRS BCIs are an active area of research because the modalities naturally complement each other. EEG can provide quick event detection and fNIRS can provide additional confirmation and context. For example, a hybrid system might use EEG to get an initial quick guess of a user’s intent and fNIRS to verify it over a longer window, which can reduce false positives. One study on semantic decoding (distinguishing whether a person is imagining an animal vs a tool) explicitly combined EEG and fNIRS, noting that fNIRS could help overcome EEG’s low spatial resolution by indicating which cortical regions were involved. The two together improved overall accuracy of deciphering the imagined category. Another advantage of fNIRS is that it’s immune to electrical noise and avoids certain EEG artifact issues. This makes it attractive for use alongside EEG in environments where EEG quality degrades (for instance, in motion or outdoors). In a hybrid setup on a mobile user, EEG might drop out due to movement, but fNIRS – being more tolerant to movement and unaffected by EM interference – could still function, providing a fallback signal.
Consumer and Portable fNIRS Systems: Traditionally, fNIRS devices were research instruments in labs, but there is a trend toward portable and even consumer-grade fNIRS. For example, a company called Kernel has developed a headset called Kernel Flow, which is a high-density time-domain fNIRS device. It looks like a bicycle helmet and can record from many channels with good signal quality, intended for both research and eventually wellness applications (such as tracking cognitive activity in daily life). Although expensive, such devices hint at a future where one could wear an unobtrusive fNIRS cap and have continuous monitoring of certain brain states. Some startup projects have even explored fNIRS-based neurofeedback for stress management, since it can monitor the prefrontal cortex activation that correlates with stress or concentration. Compared to EEG, fNIRS has the benefit that you don’t need to worry about electrical contact or hair preparation – functional NIRS sensors can even work through hair to some extent, and newer ones use high-bandwidth light sources for better penetration.
One creative use of fNIRS is in augmented reality (AR) and human-computer interaction. Researchers have developed AR applications where fNIRS is used to assess the user’s mental workload. For instance, an AR headset might include embedded fNIRS optodes on the forehead measuring the prefrontal cortex. If it detects a pattern indicating high workload (maybe the user is overwhelmed by the AR task), the system could alter the interface or provide assistance. These kinds of context-aware interfaces leverage the unique ability of fNIRS to measure brain states continuously and relatively comfortably. A 2020 proof-of-concept by Luu et al. showed an AR fNIRS BCI for a simple control task: users in AR could perform a hands-free selection of virtual objects by modulating their fNIRS signals via mental tasks. Though rudimentary, it points toward combining brain sensing with AR/VR tech for more immersive control schemes.
Robotics and fNIRS: Using fNIRS to directly control robots is less immediate than EEG, but it has been explored in high-level control scenarios. The Brain-Robot project at Maastricht University, for example, developed a mobile, non-invasive fNIRS brain-robot interface to let a user’s intentions control an autonomous robot. The concept was that the user would have high-level goals (like “pick up the object” or “move to location”) that they communicate via fNIRS by thinking in certain ways, and the robot’s autonomous system (AutInEx: automated intention execution) would then carry out the complex sequence of actions to fulfill that goal. The fNIRS BCI acted as a trigger or selector of pre-programmed actions, rather than manually guiding the robot continuously. This approach of shared control is well-suited to fNIRS because of its low bitrate – the person doesn’t control every joint of the robot in real time (which fNIRS could not update fast enough), but rather issues a command when needed which the robot confirms and executes. Initial tests of the system were done offline (replaying recorded fNIRS data) and then moved to real-time, showing that users could indeed select tasks for the robot via changes in their fNIRS signals. Although slower than EEG, the fNIRS BCI had the benefit of being robust and requiring little training, and the user did not need to wear an uncomfortable cap (fNIRS optodes can be embedded in a normal cap or headband).
In rehabilitation robotics, fNIRS is sometimes used as a monitoring tool rather than the control signal. For example, in a study on a BCI-controlled soft robotic glove for stroke patients, EEG was used to control the glove (the patient imagines moving their hand, EEG triggers the glove to move), while fNIRS was used to measure cortical activation changes before and after training. The fNIRS data showed how the BCI-driven therapy increased activity and connectivity in motor-related brain areas over time, correlating with improved motor function. This is a great example of how fNIRS can add value in BCI systems by providing neuroimaging evidence of recovery and engaging the brain’s plasticity. For actual control, EEG was still used due to its immediacy, but fNIRS validated that the intervention was causing beneficial brain changes (increased oxygenation in sensorimotor cortex when using the BCI glove, versus no change in a control group using the glove without BCI).
Advantages and Drawbacks: fNIRS’s advantages include portability, safety, and relative insensitivity to motion artifacts and electrical noise. It’s one of the few modalities you can use in scenarios like standing/walking, which makes it valuable for studying brain activity in natural actions (hence “real-world BCIs” interest). It’s also completely silent and poses no magnetic/electrical hazards, so can be used on people with implants, small children, etc., where other methods are risky. From a user comfort perspective, fNIRS sensors are generally comfortable to wear (just small discs or fibers on the scalp), though high-density setups can feel warm or tight. On the other hand, fNIRS has notable limitations for BCIs: the delay in response means command throughput is low; also, physiological factors like systemic blood flow changes can affect signals (e.g., if your heart rate or blood pressure shifts, it might reflect in the fNIRS data). Researchers work around this by clever signal processing and choosing tasks that give strong localized responses. Additionally, fNIRS only measures the outer cortex up to maybe 1.5–2 cm deep, so it cannot directly capture activity from deeper brain regions (important signals from, say, the thalamus or deeper cortical layers are invisible to fNIRS).
Yet, fNIRS continues to rise in popularity for BCIs because it offers something EEG and MEG do not: it’s a hemodynamic measure, which is slower but often clearer for certain cognitive states. When one is engaged in a mental task, the slow ramp-up of blood flow can serve as a distinct signature that complements the fast oscillatory changes in EEG. For tasks that last a few seconds (mental arithmetic, meditation, etc.), fNIRS can robustly detect “task on vs off” states. As deep learning methods improve, decoding more nuanced patterns in fNIRS signals (like distinguishing different types of mental tasks or emotional states) is becoming feasible.
In summary, fNIRS-based BCIs fill an important niche: they enable brain-computer interfacing for populations and contexts where EEG/MEG are less effective or impractical, and they pair well with AI and automation to overcome their slower response. While you wouldn’t use fNIRS alone to play a fast-paced video game, you might use it to answer a question when you cannot move or speak, or to provide context to an AI about how stressed or engaged you are. With companies and labs working on miniaturizing fNIRS (even towards chip-scale NIRS sensors) and increasing channel counts, we can expect its sensitivity to improve. There is even speculation about combining fNIRS with transcranial brain stimulation in future headsets – one device to read and also modulate brain activity. For now, the proven value of fNIRS in BCIs lies in communication for the severely disabled, monitoring during neurorehab, and acting as a synergistic partner to EEG in hybrid systems for more robust brain-controlled interfaces.
AI and Machine Learning for Real-Time Thought Decoding
Advances in artificial intelligence have been a driving force behind the recent breakthroughs in brain-computer interfaces. In the past, decoding EEG or other signals often relied on hand-crafted features and relatively simple classifiers (like linear discriminant analysis or support vector machines). These methods work for basic BCI tasks, but they struggle as we ask BCIs to interpret more complex or subtle mental phenomena. Modern machine learning, especially deep learning, has significantly improved the accuracy and scope of thought decoding from non-invasive signals. AI algorithms excel at recognizing patterns in high-dimensional, noisy data – exactly what brain signals are. Here we highlight how AI is integrated at various stages of BCI: from preprocessing and artifact removal to classification and even data augmentation, as well as how it enables sophisticated systems by combining neural decoding with context and prediction (like language models).
Signal Processing and Feature Learning: One of the first challenges in BCI is filtering out noise and artifacts. AI-based approaches, such as independent component analysis and deep autoencoders, can automatically separate neural signals from noise better than manual filtering. For example, convolutional neural networks can be trained to take raw EEG time-series as input and implicitly learn to ignore eye-blink artifacts while amplifying relevant brain-wave patterns. In the Meta Brain2Qwerty architecture, the initial stage was a CNN that learned spatial and temporal filters on the multi-channel EEG/MEG data, effectively doing an AI-driven feature extraction instead of relying on pre-defined features. This allowed the network to isolate patterns linked to each keystroke in the typing task. The use of deep CNNs for EEG has become popular following models like EEGNet, which can achieve strong performance on tasks like motor imagery classification by learning frequency-specific filters similar to neuroscientist-defined bandpower features, but in a data-driven way.
Sequence Models: Many thought decoding problems involve time sequences – e.g., a series of EEG samples corresponding to a sequence of imagined words. Recurrent neural networks (RNNs) and transformers shine here. They can maintain context over time and model how brain signals evolve. For imagined speech decoding, for instance, researchers have used long short-term memory (LSTM) networks to capture the temporal dependencies as a person “speaks” a sentence in their mind. Transformers, which are state-of-the-art in natural language processing, have also entered BCI: in Brain2Qwerty, a transformer module was used to interpret the sequence of character probabilities from the CNN, allowing the system to predict full words by considering context. This drastically improved accuracy, since the model could, say, infer that after “hell” the next letter is likely “o” to form “hello,” even if the raw signal for “o” was a bit uncertain. Effectively, AI has enabled BCIs to do autocompletion and error correction, similar to a smartphone keyboard, but based on brain signals.
Pretrained Models and Transfer Learning: An exciting trend is using pretrained deep learning models (from outside the BCI domain) and transferring their knowledge. One example is the use of language models in decoding text from brain activity. By aligning neural signals with embeddings from a pretrained language model, decoders can leverage linguistic knowledge to ensure the outputs make sense. Meta’s brain-typing system included a large language model component to refine the output into valid sentences, essentially acting as an autocorrect that knows about grammar and common phrases. This is crucial when decoding thoughts like speech or semantics, because it injects world knowledge that pure signal decoding lacks. Another area is using computer vision models to decode visual imagery from EEG/MEG: researchers have begun using generative adversarial networks (GANs) or stable diffusion models conditioned on brain signals to literally generate images of what a subject is seeing, based on fMRI or even EEG features. While still in early stages for EEG, these approaches indicate how powerful AI priors can fill gaps in noisy neural data.
In terms of transfer learning across users, AI helps mitigate the notorious problem of BCI calibration for each new user. Deep models can be pre-trained on data from many individuals (learning generalizable representations of EEG signals) and then fine-tuned with a small amount of new user data. This greatly cuts down the tedious calibration many BCI systems historically required. For example, a deep network trained on dozens of people’s motor imagery EEG can give a new user usable control almost immediately, adapting on the fly to that user’s brain patterns with just a few minutes of data (sometimes called “few-shot learning” for BCIs). Similarly, transfer learning across sessions helps address day-to-day variability in signals.
Enhanced Classification and Regression: For traditional BCI classification tasks (like which of N commands the user intends), deep networks often outperform classical methods once enough data is available. They can combine spectral, spatial, and temporal features optimally. In one 2021 study, a deep CNN achieved higher accuracy in classifying 4 different mental tasks from fNIRS data than any classical algorithm, by converting time-series into image-like representations and applying convolution (a technique using Gramian Angular Fields). Another example is in motor decoding: deep regression models can map EEG signals to continuous 2D coordinates to move a cursor, performing smoother control than was previously possible with linear decoders. Even in the Cybathlon BCI race, teams that incorporated deep learning (with careful regularization and training on EEG) showed improved robustness, which contributed to their pilots successfully completing tasks under competition conditions. One case reported is an LSTM-based classifier that helped a quadriplegic user continuously control a virtual race car with EEG, by maintaining context of recent signals to avoid erratic predictions.
AI for Error Detection and User Feedback: Another integration of AI is detecting when the BCI might be wrong or when the user is not actually issuing a command. This is critical for reliability. Machine learning models can monitor signals for signs of attention or intention. For instance, if the system expects a certain brain response for a command but doesn’t see it strongly, it can withhold action instead of outputting a random guess. Some BCI spellers now incorporate “error-related potentials” – a distinctive EEG pattern that occurs when the user’s brain subconsciously recognizes the system made an error. A classifier can detect this ERP and trigger the system to auto-correct or undo the last output. These meta-detectors are essentially AI oversight, adding a layer of safety to thought decoding so that unintentional thoughts are less likely to trigger actions.
Data Augmentation and Synthesis: Training data for BCI deep learning can be limited (brain data is time-consuming to collect), so researchers use AI to generate more. Techniques like for EEG augmentation include adding noise, morphing signals in time or frequency, or using generative models to simulate plausible EEG segments. This helps make models more robust. There’s also interest in “zero-shot” decoding, where a model generalizes to decode a type of task it hasn’t explicitly been trained on, by leveraging common representations. One futuristic scenario: an AI model might understand the general neural signature of “yes” vs “no” from prior data and be able to apply that to a new user who imagines nodding vs shaking their head, even if that specific prompt wasn’t in the training data. Progress toward such flexible decoders is being made with representation learning approaches in which the model learns fundamental building blocks of brain signals (perhaps an analogy to phonemes in speech, or components of thoughts) that can be recombined.
AI and Multimodal Integration: AI also facilitates combining multiple input streams – EEG + fNIRS + perhaps other sensors like eye-tracking or EMG – to yield a more accurate overall BCI system. Deep networks can take multimodal inputs and learn the optimal fusion, weighting the more reliable modality at each moment. For example, during movement, EEG may degrade due to muscle artifacts but fNIRS remains clean; the AI can dynamically rely more on fNIRS then, and switch back when the user is still. This kind of intelligent fusion was implemented in some prototypes (e.g., a continuous robotic arm control with hybrid EEG and computer vision, where a neural network decided when to use EEG vs when to trust the vision system’s autonomous grasping).
Brain-State Aware AI: Besides decoding user commands, AI can use brain data to adjust how it interacts with the user. If a deep learning model detects the user is confused (from EEG frontal theta rhythm increase, say), an AI assistant might alter its explanation or simplify a task. This is a more implicit use of BCI, turning brain signals into a user experience modifier. It’s done via machine learning models trained to map EEG/fNIRS features to cognitive states like “high workload” or “low engagement”. Companies in neuroergonomics use such AI-powered BCI systems for measuring operator fatigue in real time.
Generative AI: The Future of Decoding? One can imagine eventually having generative AI that directly interfaces with brain signals: for example, an AI that listens to one’s EEG and generates text not by classification, but by literally using a large language model conditioned on neural input. Some research teams have started exploring connecting brain recordings to GPT-like models, where the model fills in text that the brain data suggests. Similarly, for vision, using something like DALL-E or Stable Diffusion guided by brain activity from the visual cortex to depict what the user is imagining. These approaches, while experimental, show how the line between brain data and high-level AI understanding is blurring.
In summary, AI has become the decoder and interpreter of the brain’s whispers. Without AI, the raw signals from EEG, MEG, or fNIRS are largely indecipherable for complex tasks. With AI, especially deep learning, we can translate those signals into increasingly rich outputs – whether it’s selecting a letter of the alphabet, moving a robotic limb through a smooth trajectory, or reconstructing a sentence a person is thinking of. As computing power and algorithms continue to improve, we expect BCIs to benefit from ongoing AI advancements like more efficient neural networks (important for running decoders in real time on wearable devices), explainable AI (to understand what patterns the model is using and ensure they make neurophysiological sense), and continual learning (AI that adapts over months and years to an individual’s changing brain). The partnership between BCI and AI is a virtuous cycle: better AI yields better BCIs, and the demand of BCI problems spurs new AI innovations tailored to noisy, real-time, human data.
Brain-Controlled Robotics and Neuroprosthetics
One of the most inspiring applications of BCIs is using them to control robots and prosthetic devices – effectively bypassing the body to let the brain act on the world directly. Non-invasive BCIs have enabled various forms of robotic control, from moving cursors on a screen to driving wheelchairs and operating robotic arms. While invasive BCIs (like implanted electrode arrays) currently offer the most precise control for prosthetic limbs, non-invasive methods have achieved remarkable feats without the need for surgery. Here we explore how EEG and other BCIs interface with robotics, the levels of control achievable, and examples in assistive technology and beyond.
Wheelchairs and Mobile Robots: An early target for BCI control was the powered wheelchair, giving paralyzed individuals a way to move independently. EEG-based wheelchair control has been demonstrated using different strategies. One common approach is a thought-based directional command: the user imagines an action corresponding to a direction (e.g., imagine moving left hand to go left, right hand to go right) and a classifier maps EEG patterns to those directions. Additional imagined actions or states (like pull both hands = “stop”) can increase the command set. In practice, continuous control purely by EEG is challenging, so many systems use a semi-autonomous mode. For instance, the user might issue a command “go forward” via BCI, and the wheelchair’s sensors (lidar, cameras) handle obstacle avoidance on the way. This shared control relieves the user from micromanaging every turn, which is difficult at EEG’s bit rates. Even so, in clinical experiments, patients were able to navigate simple courses with BCI-driven wheelchairs, albeit at slow speeds and with considerable concentration. The Cybathlon competition’s BCI race (2016 and 2020 editions) essentially simulated such a scenario in a virtual environment: pilots had to use EEG to steer an avatar through gates representing “turn left,” “turn right,” etc., which is analogous to controlling a powered wheelchair in a track. The success of multiple pilots completing the race showed that with training and robust signal processing, EEG can drive mobility systems. One team reported that their tetraplegic pilot used the BCI at home for months leading up to the event, building proficiency, and they highlighted the importance of mutual learning (the pilot learning to modulate brain signals, and the algorithms tuning to the pilot) for achieving reliable control.
Robotic Arms and Manipulators: Controlling a robotic arm or a prosthetic hand is more complex than wheelchair navigation because it involves multiple degrees of freedom and more continuous motions. Non-invasive BCIs have achieved coarse but useful control in this domain. A notable example is the Brain-Robot interface with AR experiment described earlier. Researchers combined an EEG-based motor imagery decoder with an augmented reality interface and an intelligent robot arm. The user would look at an object (using eye-tracking to select it) and then imagine a hand movement to pick or place, which the EEG decoder interpreted as a chosen action. The robot then executed the action autonomously. This system highlights how BCI can work in tandem with other inputs (eye gaze) and robotic intelligence (autonomous precise movement) to accomplish a task that neither could do alone: the user’s brain provided intent (what object, what to do with it) and the robot handled execution. Performance varied among users, with some achieving nearly 100% task success and others struggling due to misclassifications. The key takeaway is that shared control and multi-modal interfaces greatly enhance BCI-robot interaction. The brain doesn’t have to explicitly control every joint of the robot arm – a high-level command is enough, and the robot’s control system fills in the details.
Another approach to controlling a robot arm is via continuous EEG signal features. Instead of discrete commands, algorithms can map features like sensorimotor rhythm amplitude to continuous velocities. For example, an increase in beta rhythm suppression (indicating stronger motor imagery effort) could make the robot arm move faster. Some groups have managed 2D or 3D control by combining two mental tasks (like imagining left hand vs right hand movement to move in X-axis, and maybe foot imagery for Y-axis, etc.). However, juggling multiple continuous control signals is mentally demanding and often less accurate. A hybrid BCI can help here too: one system combined EEG with eye-tracking where gaze set the spatial target and EEG confirmed the action – making it easier than using EEG for both direction and confirmation.
One celebrated demonstration was a few years ago where a quadriplegic man controlled a robotic exoskeleton suit to walk (mentioned earlier, but that used implants, not non-invasive). For non-invasive, there have been attempts like using EEG to control exoskeletons for gait rehabilitation. Typically, EEG picks up the intent to walk (from motor cortex or frontal signals) and the exoskeleton then moves the legs in a pre-defined walking pattern. This is less fluid than invasive systems but can be enough to initiate steps. In clinical rehab, even partial control or just the sense of agency (knowing one’s own brain triggered the movement) can improve engagement and outcomes.
Drones and Telepresence: Flying a drone via BCI has captured public imagination. Researchers have held “drone races” where pilots wear EEG headsets and attempt to fly small UAVs through checkpoints by thinking about movements. These setups often use P300 or SSVEP paradigms: e.g., flashing arrows on a screen and when the pilot focuses on the desired direction, the corresponding EEG evoked potential triggers the drone to turn. Several teams have successfully raced drones using such EEG interfaces, though with much slower responsiveness than manual sticks. It’s as much a tech showcase as a practical tool, but it spurs improvement in making BCIs more responsive. A more practical telepresence example is controlling a camera or robot in a remote location. A user could wear an EEG and control where a telepresence robot looks or moves simply by gazing at interface options or using motor imagery for directional commands. It has applications for people with mobility impairments to visit remote places virtually by controlling a surrogate robot.
Prosthetic Devices: For amputees or individuals with paralysis, BCIs offer a route to control prosthetic limbs or computer cursors when muscles cannot. EEG-based control of a prosthetic hand has been trialed, though invasive methods (EMG from residual muscles or implanted electrodes) are typically preferred for fine control. One non-invasive approach is using residual neural signals above a spinal injury: some quadriplegic patients can still generate EEG patterns associated with attempted limb movement. BCIs can harness those signals to control a prosthetic or even stimulate the muscles (functional electrical stimulation, FES) to move the paralyzed limb again. There have been cases where a patient with paralysis uses EEG to trigger FES in their own hand, allowing them to grasp an object – a form of neuroprosthetic bypass. In a study with stroke patients using an EEG-controlled hand orthosis, those who received feedback where their imagined movement actually caused the hand to move showed better motor recovery than those who just imagined without feedback. This is likely due to the closed-loop engaging neuroplasticity: the brain’s attempt leads to visible motion, reinforcing the brain-muscle connection over time.
Industrial and Collaborative Robotics: Outside of personal assistive devices, BCIs are being explored to control or collaborate with robots in work settings. For example, an assembly line worker could use a BCI to control a robotic assistant hands-free, perhaps by selecting tasks for the robot to do while the worker’s hands are busy with something else. In multi-robot scenarios, a concept called Brainet or multi-brain BCIs envisions multiple people jointly controlling multiple robots via shared neural control – though this is highly experimental. A 2024 arXiv paper (Ouyang et al.) titled “BRIEDGE: EEG-Adaptive Edge AI for Multi-Brain to Multi-Robot Interaction” outlines a system where two or more users’ EEG signals are combined by an AI to command a team of robots. The idea is to pool decisions, perhaps for improved reliability or to allow cooperative control of complex tasks.
Challenges in BCI Robotics: Non-invasive BCI-controlled robots generally operate slower and with less precision than manual or invasive controls. There is often a trade-off between decoder complexity and usability – a simple two-command BCI is easier to use but limits functionality, whereas a multi-command continuous control BCI offers more freedom at the cost of high mental load and likely more errors. To make BCI robotics practical, a lot of assistance from the robot’s side is used: intelligent control algorithms, obstacle avoidance, target locking, etc. This way the user can issue sparse commands and the robot infers the rest. This shared autonomy is a key paradigm. It’s been very effective, for example, in Brain Painting applications (users select high-level brush strokes via BCI, and software smooths and applies them to create artwork beyond the raw input capability).
Another challenge is user training and fatigue. Operating a BCI for a long period can be mentally exhausting. Users often need breaks after a few tens of minutes, as concentrating on generating clear signals is tiring. For robots that need continuous control, this is an issue. Incorporating passive control modes, where the robot can take over routine actions, helps reduce fatigue.
On the flip side, BCIs can also monitor brain signals during manual robot control to improve safety. For instance, in operating hazardous machinery, an EEG could detect if the human operator is losing focus, and slow the robot or trigger an alert (a different aspect of human-robot interaction enabled by BCIs). Likewise, error potentials in the human brain when observing a robot’s mistake can feed into the robot’s learning (a form of human-in-the-loop training).
Real-World Impact: While still not mainstream, brain-controlled robotics have had some powerful real-world demonstrations. One case was a paraplegic man who, using a BCI, was able to drive a vehicle on a race track in 2016 – he used a cap with EEG and a custom interface to accelerate and steer a NASCAR car (with safety systems) by thought. Another is the ongoing development of mind-controlled prosthetic arms for upper limb amputees using EEG signals from motor imagery combined with residual muscle signals. Although not as fluid as invasive muscle decoders, even semi-autonomous prosthetics that open/close or switch grip modes via EEG command can restore some functionality.
Looking forward, improvements in signal acquisition (maybe more electrodes, better placement, possibly novel sensors on the scalp) and smarter AI will gradually increase the bandwidth of non-invasive control. If MEG-like sensing becomes wearable, that could vastly improve motor decoding for robotics. Also, integrating BCI control with voice assistants or other interfaces can make a system more practical – e.g., use voice when possible and fall back to BCI when voice is not an option (for someone who can still speak some, this hybrid approach could combine the strengths of each).
Ethical and societal aspects: Brain-controlled robots also raise questions: Is the person responsible if the robot causes harm while following brain commands? How do we ensure the robot only does what the user intends (and not an accidental thought)? These require failsafes and transparency. Most systems use confirmatory steps for critical actions (like a double confirmation via BCI or a secondary modality). There’s also the aspect of giving people agency – a user controlling a prosthetic or wheelchair via BCI often reports a great psychological benefit, feeling “re-connected” with the world.
In summary, BCI robotics is bridging mind and machine, allowing actions to be carried out by robotic embodiments of the user’s intentions. Non-invasive BCIs, while currently limited in precision, have demonstrated the ability to empower users in meaningful ways: moving, manipulating objects, interacting in virtual and physical spaces. As both BCI decoding and robotic autonomy improve, their synergy will lead to smoother and more intuitive control. One can envision a future assistive robot that almost feels like an extension of the user – partially controlled by brain signals, and partially by its own AI, working together seamlessly. Every year, competitions like Cybathlon and new research prototypes bring us a step closer to that reality, showing that even through a skull cap, we can drive machines with our thoughts and thereby regain abilities or enhance our interaction with the environment.
Research and Clinical Applications
Non-invasive BCIs are not only about directly controlling gadgets; they also serve as valuable tools in neuroscience research and clinical settings beyond robotics. Here we discuss how EEG, MEG, and fNIRS BCIs are used for studying the brain, diagnosing conditions, and treating patients through neurofeedback and rehabilitation. These applications often prioritize insights and therapeutic outcomes over raw control speed.
Neuroscience Research and Cognitive Monitoring: BCIs provide a unique window into the human mind by enabling real-time readout of certain aspects of cognition. Researchers have employed BCI paradigms to investigate attention, memory, decision-making, and more. For example, “mind wandering” can be detected via EEG in real time – if a subject is supposed to focus on a task but their brain transitions to an idle or daydreaming state, a classifier can catch that and perhaps trigger a prompt to re-focus. This has been used experimentally to study attention in students, with the BCI delivering a gentle nudge when EEG indicates loss of focus. Another area is using BCI to study conscious vs unconscious processing. By requiring a person to communicate only via BCI (for instance, in locked-in patients), researchers learn about which brain signals correlate with consciousness and intention. In one famous 2006 study, a completely locked-in patient was able to answer yes/no questions correctly using an EEG-based BCI, confirming they were conscious and understanding despite clinical assessments to the contrary – a finding that has profound ethical implications for patients in vegetative states.
Communication for the Paralyzed: Arguably the most impactful use of BCIs in a clinical sense has been as assistive communication devices for those who cannot speak or move (ALS, brainstem stroke, etc.). Non-invasive BCIs like the P300 speller have allowed such patients to select letters one-by-one by merely paying attention to flashing choices. While slow (typically 1 character per 5-10 seconds), it has given voice to people in “locked-in” conditions. fNIRS systems, as mentioned, extended communication even to completely locked-in patients, enabling responses after all muscle-based methods (like eye blinks) failed. These systems are being refined for speed and accuracy, and importantly, for ease of use by caregivers at home. A BCI that requires a team of PhDs to calibrate isn’t practical for everyday communication; thus, a lot of research goes into algorithms that auto-calibrate and headsets that can be donned quickly. One promising development is combining BCIs with other minimally active channels – for instance, a system where a patient uses an eye-tracker when they can, and if eye control is lost, seamlessly switches to BCI. This kind of adaptive multimodal interface can provide continuity in communication ability as diseases like ALS progress.
On the horizon, as we saw with the Meta MEG work, is the possibility of BCIs that can output not just letters but whole words or sentences from neural activity. Imagine a patient thinking of the sentence they want to say, and the system outputting it (with some delay). Achieving this non-invasively with high reliability is extremely challenging, but even a rudimentary partial decoding could speed up communication rates (perhaps going from 5 words per minute to 20). That’s still far slower than natural speech (~150 wpm), but could dramatically improve quality of life. Collaborations between neuroscientists and engineers are actively exploring imagined speech decoding with EEG and fNIRS for this reason.
Neurorehabilitation: BCIs are making waves in stroke and spinal cord injury rehabilitation. The principle is simple but powerful: use the patient’s intent to move (detected by BCI) to assist or stimulate actual movement, thereby engaging the brain in relearning motor control. For instance, a stroke patient tries to move their paralyzed arm; the EEG picks up the attempted movement brain patterns and triggers a device like a robotic exoskeleton on the arm or electrical stimulation of arm muscles to execute the movement they attempted. This closing of the loop – brain intent leads to feedback of movement – is thought to drive neuroplastic changes that help the brain rewire and regain function. Numerous studies have shown improved outcomes when BCI feedback is added to rehab compared to physical therapy alone. Patients achieve greater motor improvements and also show brain reorganization (confirmed by fMRI or fNIRS) indicating that areas of the brain took on new roles to compensate for damaged regions. Over weeks of training, BCIs essentially help patients practice moving even when their body can’t, which slowly strengthens the neural pathways needed for voluntary control.
In addition to motor rehab, BCIs have been used in cognitive rehabilitation. Disorders like ADHD, anxiety, or mild TBI can sometimes be helped with neurofeedback – where the patient is shown a real-time visual or auditory display related to their brain activity and learns to modulate it. A common example is ADHD: children with ADHD may have atypical EEG patterns (like an excess of theta waves vs beta). Neurofeedback games can encourage them to shift their brain activity toward a more focused state – e.g., a rocket in a game only flies when their beta (focus-related) waves increase and theta (drowsy-related) decrease. Over many sessions, some studies find this can reduce ADHD symptoms as the child internalizes better control of attention (though results are mixed and still under investigation). Similar neurofeedback protocols exist for anxiety (training to increase alpha relaxation waves), depression, and even as adjunct therapy for epilepsy (trying to reduce aberrant activity that might lead to seizures). All these can be considered BCIs since they involve reading brain signals and feeding back to the user (though not necessarily outputting to an external device beyond a screen). The success of closed-loop BCIs in inducing neuroplasticity is a major research theme. For example, pairing brain activity with stimulation – if a BCI detects a certain pattern and then triggers a peripheral nerve stimulus – can potentially re-wire brain circuits. This is being explored in stroke: detect motor intention with EEG, and time a stimulation of the limb at just the right moment to reinforce that brain-muscle connection.
Mental Health and Biofeedback: A burgeoning area is using BCIs for mental wellness. Devices like Muse (EEG headband) provide meditation feedback by translating real-time brain activity into guiding sounds (e.g., calm waves when the user is in a meditative state, stormy sounds when the mind is wandering, nudging them back) as a form of mindfulness training. Thousands of people use these at home, effectively making EEG neurofeedback mainstream. There are also research projects looking at using fNIRS to monitor stress in real time and guide breathing exercises, or using EEG to detect early signs of a panic attack and trigger relaxation protocols on a paired smartphone app. While these are not “BCI” in the traditional sense of controlling an external device, they are brain-computer interfaces in that they interface the brain with a computer to alter the user’s experience for therapeutic purposes. The non-invasive nature of EEG/fNIRS makes them suitable for such wellness applications with no significant risk.
Diagnostics and Monitoring: BCIs can assist in diagnosis of certain conditions by providing interactive brain testing. For instance, an EEG-based BCI can assess if a patient with disorders of consciousness (coma or vegetative state) has awareness by seeing if they can follow commands in their brain signals (like “imagine playing tennis” vs “imagine resting” – an fMRI paradigm adapted to EEG). A positive result indicates covert awareness (as in some famous cases by Adrian Owen using fMRI). EEG BCIs may be a cheaper bedside alternative for this purpose. Another diagnostic angle is using BCIs to quantify cognitive decline – a sort of “brain vital sign.” Some researchers present stimuli and measure BCI responses (like a P300 to an oddball sound) to gauge brain health, and track changes over time or with treatments.
In sleep research, EEG systems (like personal headbands or more advanced polysomnography BCIs) are used to detect sleep stages. There’s even experimentation with lucid dream induction via EEG: detecting REM sleep onset and then providing an auditory cue to the sleeper to trigger awareness in the dream (which could be considered a kind of brain-computer interaction across the sleep-wake boundary).
Education and Training: BCIs are being used experimentally in education to personalize learning. An EEG headset on students could monitor engagement; an AI tutor adjusts difficulty if it senses boredom or confusion, as inferred from brain metrics. Another concept is using BCIs to train skills – for example, in sports or military training, showing trainees their brain activity live to help achieve optimal focus (the concept of entering “the zone” could be reinforced by feedback). Some pilot studies with archers and golfers gave neurofeedback to help them reach a desired EEG pattern associated with peak performance, resulting in improved outcomes compared to no feedback.
Neuroethics and Privacy: With all these applications, especially ones dealing with patient communication and monitoring, ethics are paramount. Non-invasive BCIs, unlike implants, are generally low risk physically, but the data they collect is sensitive. Real-time thought decoding raises questions: Could someone’s private inner speech be decoded without consent? (Currently, no – the tech isn’t that capable without cooperation and training – but in principle one day maybe). Ensuring BCIs decode only intended communications is critical; Meta’s team, for instance, stressed that their Brain2Qwerty system decodes only attempted typing, not random thoughts. In clinical use, issues of consent for patients who can’t signal normally have to be handled carefully (BCI might be the means to ask consent). There’s also a push for a “neurorights” framework that includes right to mental privacy, which would cover BCI data and its protection.
Validation and Reliability: For clinical adoption, BCIs must be tested rigorously. A device used for communication by a locked-in patient must work consistently and be user-friendly. Some EEG spellers (like the g.tec IntendiX) have received medical device approval in Europe, reflecting progress toward standardized tools. Clinical trials for stroke rehab BCIs have shown positive results, but larger studies are needed to convince health systems to integrate BCIs into standard rehab protocols. As of now, many rehab BCI successes are in controlled research settings. Bridging that to everyday clinical practice (with busy therapists, limited training time, cost constraints) is a challenge being actively worked on.
In conclusion, research and clinical BCIs are expanding what we can measure and heal in the brain. We are using them to ask and answer research questions about brain function, to restore communication and movement to those who lost it, and to potentially modulate brain states for better health. Non-invasive BCIs are at the forefront because they can be deployed widely and ethically with minimal risk. They might not be as dramatic as surgically implanted chips that make headlines, but their steady integration into assistive technology and therapy could impact far more people in the near term. Every incremental improvement in speed or accuracy directly translates to better quality of life for users relying on a BCI to speak or move. And every novel use – be it helping a child with ADHD focus or a stroke patient regain arm use – broadens the societal acceptance and knowledge of this technology, paving the way for brain-computer interfaces to eventually become as commonplace as voice interfaces are today.
Consumer Neurotechnology and Daily-Life BCIs
While research and medical uses of BCIs are crucial, an emerging frontier is the integration of BCIs into everyday consumer tech for generally healthy users. This realm, sometimes called “neurorobotics” or “personal neuroinformatics,” envisions brain sensors as a standard input modality in our devices, alongside touchscreens, cameras, and microphones. We have touched on some consumer applications under EEG, but here we consider the broader picture: how BCIs are entering daily life and what opportunities and challenges they bring.
Current Consumer BCI Products: The consumer BCI market has several players offering wearable EEG devices intended for personal use. To recap a few notable ones:
- Muse: A slim headband with 4 EEG electrodes, marketed for meditation and relaxation. It provides real-time feedback (like sounds of weather) correlated to the user’s brain state, helping them learn to meditate more effectively.
- NeuroSky MindWave/MindLink: Low-cost devices (1 or 2 electrodes) that measure basic brainwave stats (attention/meditation scores) and were used in toys and simple games. One famous toy was a Star Wars-themed device where users “levitated” a small ball by concentrating (i.e., increasing certain EEG activity raised the ball via a fan).
- Emotiv Epoc/Insight: Higher-channel wireless EEG headsets (5, 14 channels) that come with software for both developers and consumers. They can do things like facial expression detection (via EEG electrodes picking up muscle signals) as well as some mental command recognition. Emotiv’s apps include games where you can push or pull virtual objects with your mind and a brain-controlled wheelchair demo for showcase.
- OpenBCI: An open-source project that sells EEG headsets (the Ultracortex) and biosensing boards. While targeted at hobbyists and researchers, some consumer DIYers use these to experiment and build custom neurotech – like brain-controlled art installations or home automation triggers.
- NextMind (now part of Snap): A unique EEG-based headpatch that was placed on the back of the head to pick up visual cortex activity. It allowed a user to control interfaces by focusing on visual elements; for example, in a demo, users could unlock a smart door by staring at a specific icon and “thinking” it open (the device detected the brain’s response to a flashing pattern in the icon the user was focusing on). This approach essentially turned EEG into a eye-gaze selection tool without using the eyes themselves.
- Neurosity Crown: A newer entrant, a crown-like 8-channel EEG device that claims to track focus and cognitive state, connecting to productivity apps to help users time their deep work sessions or take breaks when concentration wanes.
These devices indicate that the consumer appetite for brain tech is real, albeit currently niche. Their primary uses revolve around self-tracking (quantified self), wellness (mindfulness, stress reduction), and novelty/entertainment (mind-controlled games or experiences). The prices range widely – from under $200 for simple headsets to several thousand for research-grade ones – making some accessible to average consumers and others mainly to enthusiasts or professionals. Reviews often note that while fascinating, the tech is early: results can be inconsistent, and many people question “is it actually reading my mind or just picking up noise?” Ensuring a good signal (proper fit, minimal movement, etc.) is a new requirement that consumers have to learn, akin to when heart-rate monitors first came out and people had to figure out how to wear them correctly.
Wearables and Integrations: We are also seeing EEG sensors being integrated into other wearables. For example, some augmented reality glasses prototypes include dry EEG electrodes in the frame that touch the user’s scalp – potentially for context sensing or simple commands. Similarly, earbuds have been tested with EEG electrodes that touch the ear canal (capturing brain signals there). While ear-EEG is lower quality than scalp, it could detect certain states (like when the user is falling asleep listening to audio, the earbuds could pause the music). There’s a conceptual product idea of “neural headphones” that both play audio and read brainwaves, personalizing your music or learning from your cognitive responses.
Smart Home and IoT Control: Envision a scenario where you could turn on your lights or TV just by thinking. Companies have demoed smart home control with BCIs: e.g., focusing on a symbol for a lightbulb on a screen and the light toggles (using a P300 or SSVEP BCI). In research labs, similar setups have allowed paralyzed users to control their home environment – like selecting “TV” vs “Fan” on a BCI-controlled interface and then controlling those devices. For the average consumer, a fully BCI-driven home is overkill, but selective use could be helpful. Perhaps a sleep monitoring BCI automatically turns off your IoT devices when your brain signals show you’ve slept. Or a BCI alarm clock that wakes you at an optimal moment in your sleep cycle.
Gaming and Virtual Reality: Gaming is a space that always pushes interactive tech. While no AAA game uses BCI as an input yet, there are niche games built around EEG (like throw the fireball if you can enter a focused state). Valve Corporation has reportedly researched BCIs to incorporate into VR, not for direct control but to adapt gameplay. They filed patents on using physiological signals (including EEG) to adjust game difficulty or narrative direction based on player engagement. Imagine a horror game that becomes more intense when your brain signals show you’re not scared (to try to scare you more) and eases off when you’re very highly aroused. That could create a more immersive experience responsive to your actual fear rather than just scripted. BCIs could also enable new game mechanics; for instance, a co-op game where players “synchronize” their brainwaves to unlock a power (forcing a kind of team meditation exercise!).
In VR/AR, BCIs can provide hands-free input where controllers are inconvenient. There’s ongoing exploration into “active VR” where you might wear a lightweight EEG inside a VR headset to pick menu items without using your hands – useful if your hands are busy or you want extra degrees of input (like controlling a game character’s powers by thought while hands do movement). The NextMind example essentially did that, and it’s likely that AR glasses of the future could incorporate something similar for quick commands (maybe thinking “click” while looking at a virtual button).
Neuro-Marketing and Research: Some companies use consumer EEG headsets for market research – measuring brain responses to advertisements, products, or movie trailers to gauge emotional engagement beyond what people self-report. This is controversial and an evolving field known as neuromarketing. The quality of insights is debated, but it shows another angle of consumer neurotech: not user-facing novelty, but quietly gathering data on what subconsciously attracts attention. Big advertisers have been interested in this, though rigorous science is needed to interpret EEG in such complex real-world stimuli.
Training and Cognitive Enhancement: We touched on neurofeedback for performance (focus training for athletes, etc.). For everyday consumers, there are now apps with EEG headbands claiming to improve concentration or memory through regular brain training exercises. Some are packaged like games that one controls via attention level – e.g., keep a spaceship flying by staying focused. This verges into territory of cognitive enhancement – using BCI to potentially achieve a mental edge. While the evidence for lasting benefits is mixed, users often do report short-term improvements or at least heightened self-awareness of their mental state.
Concerns and Skepticism: With consumer neurotech, one must be cautious about overhyping. Many devices are marketed in ways that blur science and sci-fi, leading to unrealistic expectations. For instance, calling a headband a “mind reading” device is misleading; it might detect stress, not read specific thoughts. Education is needed for consumers to understand what brain signals are and are not. Privacy is also a concern: Could companies collect EEG data and infer things like mood patterns for targeted advertising? The resolution is low for now, but one can’t ignore the questions.
Additionally, there’s a risk of security: in theory, a malicious actor could try to spoof BCI inputs or a poorly secured BCI toy could be hijacked (imagine a mind-controlled drone being taken over by someone else’s transmitter). These are technical issues that standard cybersecurity can address – encryption, authentication for BCI device linking, etc.
User comfort and fashion are non-trivial too. People might not want to wear electrodes on their forehead in public unless they’re nearly invisible or built into normal-looking apparel. Google Glass faced backlash partly for being socially awkward; a brain-sensing headband could face similar challenges. Thus, companies work on making sensors discreet and ergonomic. Dry electrodes have improved but still sometimes require some pressure to maintain good contact, which can cause discomfort over long hours.
Future Consumer Scenarios: Looking ahead, one can imagine:
- Brain-enabled personal assistants: Your digital assistant (Alexa/Siri) not only listens to your voice but also has access to a feed of your brain state (through your smartwatch or AR glasses). It knows when you’re overwhelmed in a meeting and can reschedule conflicting tasks automatically, or it detects you’re in a creative flow and silences notifications. This requires robust and interpretation of brain signals for complex states – a challenge but not implausible as algorithms and sensors improve.
- Entertainment and media: Movies that adapt to your emotional responses in real time, immersive experiences that literally monitor your brain’s engagement and adjust pace or content accordingly. Some very initial experiments have done “neuroadaptive music” where EEG drives music parameters so that it shifts to calm you if you’re stressed.
- Education and skill acquisition: Headsets that, while you practice something (like piano or language learning), observe your brain patterns and tailor drills optimized for how your brain learns best, perhaps even detecting when a memory has been successfully encoded or if you’re about to make a mistake due to lapse in attention.
- Everyday convenience: Small things like mentally clicking through slides during a presentation when you have a BCI-equipped AR lens, or authenticating to devices via your brain’s unique signatures (brainprint as a biometric, though that is a research area with issues like variability).
The trajectory of consumer BCIs will likely mirror that of other tech: initial novelty and early adopter use, gradual improvements, integration into other multipurpose devices, and eventual normalization if value is proven. For example, early digital cameras were standalone exotic devices; now every phone has one seamlessly integrated because it clearly adds value. If brain sensors find a “killer app” that other sensors can’t provide, they might become a standard component of wearables.
Right now, the killer app is not definitively identified – relaxation training is helpful but niche, games are cool but gimmicky to some, hands-free control is nice but can often be done via voice or eye-tracking. The coming together of AI, AR, and BCI might produce something greater than the sum of parts. For instance, maybe it’s the combination (as mBrainTrain’s example) of eye-tracking + brain signals that yields a magical effortless interface for AR. Big companies like Meta, Apple, and Microsoft have all done low-profile research on brain sensing for their AR/VR divisions, so we might see features quietly integrated soon (even if just as supplementary sensor for attention tracking or health metrics).
In the end, if consumer BCIs do become commonplace, it could transform how we interact with technology – making it more empathic and personalized. Devices could know when we are frustrated, confused, bored, or excited without us having to explicitly say so, and respond accordingly. This has profound implications for human-computer interaction, essentially making it more “human-aware.” Of course, it raises new ethical requirements to ensure that such intimate data is handled with consent and care.
For now, we are in the early adopter and exploratory phase. The neurotechnology market is growing, and non-invasive BCIs are at its forefront since invasive devices are not ethical or feasible for general consumers. As the technology matures, we will likely see a winnowing of ideas to those that truly stick and improve user experience in a noticeable way. Much like how touchscreens existed for decades but only truly took off with the right interface (smartphones), BCIs may need that combination of hardware capability and use-case maturity. The next decade will be telling as neural interfaces quietly slip into more products and we learn from both successes and failures in the consumer realm.
Challenges and Future Outlook
Despite the impressive progress in non-invasive BCIs, significant challenges remain before they become ubiquitous, high-performance tools. Addressing these challenges is the focus of many ongoing research efforts. At the same time, future developments hold promise to greatly expand BCI capabilities. In this final section, we consider the key hurdles – technical, user-centered, and ethical – and discuss emerging innovations and potential directions for the field in the coming years.
Signal Quality and Reliability: The foremost technical challenge for non-invasive BCIs is the low signal-to-noise ratio of brain signals. EEG electrodes pick up microvolt-level signals that are easily masked by noise from muscles (EMG) and external electrical sources. fNIRS signals can be confounded by systemic blood flow changes unrelated to neural activity. MEG, while cleaner, is highly sensitive to magnetic interference and motion. These noise issues cause variability and limit the fidelity of decoding. Progress is being made on multiple fronts: better hardware (like active shielding for MEG, improved optode design for fNIRS, novel electrode materials for stable EEG), better signal processing (sophisticated filtering, ICA, and AI denoising techniques), and artifact mitigation strategies (like combining BCIs with eye trackers to subtract out eye-blink artifacts in EEG, or built-in accelerometers to tag movement artifacts).
The ultimate hardware dream is to approach invasive-like signal quality without being invasive. That might involve new sensor modalities – for instance, some researchers are exploring ultrasonic approaches to sense neuronal activity via skull vibrations, or hybrid EEG/functional ultrasound. Advances in materials might yield dry EEG electrodes that perform as well as wet gel ones (which currently still reign for best quality). Nanomaterials, graphene-based electrodes, or micro-needle (almost-but-not-quite invasive) electrodes that just barely penetrate the skin could dramatically improve EEG contact quality.
Wearability and Convenience: Non-invasive BCIs must become as easy to put on and wear as a pair of glasses if they are to be widely adopted outside the lab. Today’s systems still often involve cumbersome caps, time-consuming placement, or uncomfortable pressure points. The future likely holds fully wearable BCIs that look like simple headbands, hats, or even earbuds. Flexible electronics and conductive fabrics could allow sensor arrays to be embedded in a normal-looking cap, conforming to the head shape and perhaps self-adjusting for optimal contact. Companies are already working on EEG electrodes that don’t require any gel and can be worn for hours comfortably (some use spongy or bristle designs that bypass hair). The UT review noted that soft, stretchable BCIs are a growing trend, aiming for long-term stability and comfort. When BCIs can be donned in seconds and forgotten about, their use will expand massively.
User Training and Usability: Many BCI systems to date require users to undergo training to learn to control them. This training might take days or weeks of practice, which is a barrier for both clinical and consumer adoption. Efforts to reduce or eliminate user training include adaptive algorithms (that learn from the user instead of vice versa) and intuitive control paradigms. For example, BCIs based on natural responses (like a P300 when you recognize a target) require virtually no training – the user just does the task and the brain naturally generates the signal. Future BCIs may lean more on these implicit signals rather than expecting users to consciously modulate abstract brainwaves. And for cases where training is needed (like motor imagery skill), “assistive training” using virtual feedback, guidance, or even AI coaches could shorten the learning curve. Also, as mentioned, transfer learning can apply an existing model to a new user to get initial decent performance, so the user isn’t starting from scratch.
From a user-interface perspective, making BCIs more usable also means integrating them with other interfaces to reduce cognitive load. The brain should not have to do all the work. We will see BCIs working in concert with voice commands, gestures, eye gaze, etc., picking up slack only where needed. This multimodal approach ensures users aren’t frustrated by trying to do everything with a limited BCI channel.
Speed and Bandwidth: Non-invasive BCIs still lag far behind natural communication and movement speeds. A major future goal is increasing the bandwidth — how many bits of information per minute can be conveyed. This could come from better decoding of continuous signals (e.g., decoding even partial neural correlates of speech could boost word rates by predicting likely phonemes or words), or from combining multiple signals (an EEG+fNIRS hybrid could, for instance, transmit two parallel streams – one fast binary signal and one slower multi-class signal). However, there may be upper limits to non-invasive bandwidth that only invasive methods or leaps in sensor tech can overcome. If those limits are hit, another approach is to let AI fill the gaps and guess what the user likely means with minimal input (similar to predictive text on phones). As language models and context-aware AI get better, perhaps a BCI can get away with only sparse, low-bit inputs and still achieve flexible communication (the AI figures out the rest). This is both exciting and a bit concerning, because heavy reliance on AI to infer intent could lead to mistakes if the AI guesses wrong. So the partnership between user-driven input and AI assistance must be well-balanced and transparent.
Universality and Individual Differences: Brain signals are highly individual – the patterns one person produces for a given thought can differ from another’s. This variability means a one-size-fits-all BCI is tough. Solutions will include more personalized models (fine-tuned on each user), but also making BCIs that self-optimize each time they’re used. Possibly long-term BCIs might continuously learn your neural patterns over months, building a profile as unique as a fingerprint, which can improve both performance and security (because then someone else’s brain signals wouldn’t operate your device). Large datasets like EEG recordings from many people are being compiled in open repositories to help train generalizable decoders and understand population-level variations. The hope is to identify features that are stable across most humans (like certain ERPs or frequency responses) and design BCIs that leverage those robust features, minimizing reliance on idiosyncratic ones.
Ethical, Legal, and Social Issues: As BCIs become more capable, ensuring ethical use is paramount. Privacy of one’s neural data is a key concern – already, companies like NeuroSky and Emotiv collect EEG data, and questions arise about how it’s stored and whether it could be misused. Regulations may be needed to classify certain brain data similar to health data, granting it protections. There’s active discussion of “neurorights” such as the right to cognitive liberty (not to be forced to use a neural interface, and to have autonomy over one’s own brain states) and freedom from algorithmic bias in neurotech (ensuring BCIs work equally well for different groups, and that underlying AI models don’t disadvantage some users). Legal frameworks might also need to clarify liability – if a person’s BCI-controlled robot injures someone, is it device fault, user fault, or nobody’s fault (an accident)? And if neural data is used in court (e.g., a BCI communication from a locked-in patient as testimony), how is its validity established? These are complex issues yet to be fully addressed.
Public Acceptance: The success of BCIs will also hinge on public perception. Invasive BCIs face fear and skepticism (for good reason, given the risks), but non-invasive ones might be more easily accepted if they show clear benefit. Nonetheless, some people may feel uneasy about devices that read brain activity, even if non-invasive. Transparency about what BCIs can and cannot do, and positive narratives (e.g., focus on how they help disabled individuals or improve wellness) can help in gaining acceptance. The more BCIs can be framed as empowering tools rather than mind-reading gimmicks, the better the adoption. The gradual introduction via harmless fun applications (like games and meditation) can acclimate society to them.
Synergies with Other Tech: BCIs will not develop in isolation. They’ll ride the wave of improvements in AI (as discussed), but also in other fields: wireless tech (for untethered BCIs transmitting data to phones or cloud), battery technology (for longer use wearables), and even neuroscience (deeper understanding of how information is encoded in brain signals). There’s also synergy with neural stimulation (e.g., transcranial magnetic or electrical stimulation). We might see closed-loop systems that not only read but also stimulate the brain in response, potentially optimizing cognitive states or treating mental illness on the fly (for example, detecting depressive neural patterns and stimulating to counteract them). Some experimental therapies for depression already use EEG to guide transcranial stimulation timing for better effect – a precursor to such closed loops.
Regulatory Pathways: For medical BCIs, regulators like the FDA will require evidence of safety and efficacy. As more clinical trials show benefits (e.g., in stroke rehab or assistive communication), we’ll likely see a wave of approved BCI devices. Insurers might then consider covering them if cost-effective. Widespread clinical use could then provide more data and drive improvements, in a virtuous cycle. On the consumer side, regulation is looser, but if BCIs start getting integrated in mainstream products, standards for safety (electrical, optical) and interoperability might emerge.
Long-Term Vision: Envisioning 10-20 years ahead, one could imagine that non-invasive BCIs become as common as smartwatches. People might wear subtle neural sensors that continuously monitor brain health (flagging early signs of neurodegenerative disease perhaps), augment communication (like silently sending a message just by thinking “text John: I’ll be late”), or enhance experiences (learning new skills faster by brain-state optimization). The division between “assistive” and “augmentative” BCIs could blur – the technology developed to aid those with disabilities may augment abilities for everyone. For instance, a memory prosthesis concept that started with clinical goals could become a consumer memory assistant that cues your brain when you’re trying to recall a name.
Fully non-invasive “mindwriting” of text at speeds near typing, or controlling devices as naturally as limbs, may or may not be achievable – but the progress shown by MEG decoding studies and advanced AI suggests we shouldn’t rule it out. If not with current tech, maybe with some yet-to-be-invented modality (e.g., some optical or electromagnetic technique that gives much higher resolution through the skull). It’s worth noting that the brain is immensely complex, and decoding entire thoughts might inherently require reading deep and distributed activity that non-invasive means can’t capture. So likely, non-invasive BCIs will focus on capturing key bits that are accessible (cortical surface activity, certain waves) and teaming up with intelligent algorithms or contextual systems to make the most of that limited info.
Conclusion: Non-invasive BCIs have journeyed from rudimentary experiments to sophisticated systems interfacing with AI and robotics. They already improve lives in important ways, and their potential is only beginning to be tapped. Overcoming current limitations will require interdisciplinary collaboration – neuroscientists, engineers, computer scientists, clinicians, ethicists all working together. As BCIs become more robust, user-friendly, and powerful, they could fundamentally change how we communicate and interact with machines. The ideal future is one where BCIs are seamlessly integrated, giving people with disabilities new independence and offering everyone optional new channels of control and expression, all while respecting individual rights.
The road ahead has many hurdles, but the progress of the last decade in EEG, MEG, and fNIRS-based interfaces gives reason for optimism. With careful development and ethical guiding, non-invasive BCIs could truly usher in a new era of human-computer synergy – one in which technology adapts to us as much as we adapt to technology, bridging the biological and digital in harmony.
References
- Romero, Luis. “Meta’s Mind Reader: Brain2Qwerty Translates Thoughts Into Text.” Forbes, 19 Feb. 2025.
- Ware, Skyler. “New AI model converts your thought into full written speech by harnessing your brain’s magnetic signals.” Live Science, 10 Mar. 2025.
- The Evolving Landscape of Non-Invasive EEG Brain-Computer Interfaces. Department of Biomedical Engineering, University of Texas at Austin, 2 Jan. 2025.
- Jovanovic, Jelena. “BCI with AR: A Game-Changer for Hands-Free Control and Automation.” mBrainTrain Blog, 24 Feb. 2025.
- Noronha, A. “Differences between EEG, NIRS, fMRI and MEG.” BrainLatam Blog, 23 Dec. 2019.
- Ghalavand, Mohammad, et al. “Real-World fNIRS-Based Brain-Computer Interfaces: Benchmarking Deep Learning and Classical Models in Interactive Gaming.” arXiv preprint arXiv:2505.10536, 15 May 2025.
- Rybář, Milan, Riccardo Poli, and Ian Daly. “Simultaneous EEG and fNIRS recordings for semantic decoding of imagined animals and tools.” Scientific Data 12, 613 (2025).
- “Locked-In ALS Patients Answer Yes or No Questions with Wearable fNIRS Device.” Neuroscience News (SUNY Downstate Medical Center press release), 13 Mar. 2017.
- Perdikis, Serafeim, et al. “The Cybathlon BCI race: Successful longitudinal mutual learning with two tetraplegic users.” PLOS Biology 16.5 (2018): e2003787.
- Murad, Saydul A., and Nick Rahimi. “Unveiling Thoughts: A Review of Advancements in EEG Brain Signal Decoding into Text.” arXiv preprint arXiv:2405.00726, May 2024.
- Farnsworth, Bryn. “EEG Headset Prices – An Overview of 15+ EEG Devices.” iMotions Blog, updated 2025.
- Consumer brain–computer interfaces. Wikipedia, Wikimedia Foundation, last edited 24 Jul. 2023.
- Emotiv. “Brain Controlled Technology using EMOTIV’s Algorithms.” Emotiv Blog, 2021.
- Holmes, Niall, et al. “A novel, robust, and portable platform for magnetoencephalography using optically-pumped magnetometers.” Imaging Neuroscience 1.1 (2024).
- Ji, Xiang, et al. “Effects and neural mechanisms of a brain–computer interface-controlled soft robotic glove on upper limb function in patients with subacute stroke: a randomized controlled fNIRS study.” Journal of NeuroEngineering and Rehabilitation 22.1 (2025): 171.
Get the URCA Newsletter
Subscribe to receive updates, stories, and insights from the Universal Robot Consortium Advocates — news on ethical robotics, AI, and technology in action.
Leave a Reply