AI in Brain Computer Interfaces
Oceana Li '27
From imagination to reality, a robotic arm under the full control of brain power allowed a man who had lived with paralysis for several years to regain movement. With a lifetime of seven months, this brain-computer interface (BCI), developed by researchers at the University of California (UC), San Francisco, surpassed its predecessor’s limited life span of two days. Through artificial intelligence (AI), neural activity can be captured and translated into real-life actions in substantially shorter times and at higher accuracy. Recent innovations and applications such as UC San Francisco’s robotic arm underscore the increasing integration of AI into BCIs, improving human-to-computer connections and ultimately restoring motor, speech, and other cognitive and physical abilities for individuals with neurological disabilities.
A brain-computer interface (BCI) is a system that establishes a connection between a patient’s neural activity to an external device or machine. The BCI translates neural data that machines may interpret and execute as physical actions, allowing the user to manipulate an assistive device through pure brain power and imagination. A typical BCI architecture contains three main operations: signal acquisition, processing, and translation to application. Of the three, signal acquisition involves the detection of neural activity through electrodes that sense voltage from neuron firing or action potentials. When the signal is detected, a computer translates it into a physical action. AI plays a pivotal role in each stage of this signal processing chain, leading to the advancement and development of BCI applications that assist users with motor control, speech, and other physical functions.
There are three main types of BCIs: non-invasive, semi-invasive, and invasive. Each utilizes different techniques to acquire varying types of neural data, thus presenting their unique advantages and disadvantages. A non-invasive BCI, as evident by its name, is the simplest and safest type of BCI. Its minimally invasive approach does not require any surgical implantation, as it can be applied externally, reducing the risk of infection and providing an affordable, accessible option for patients. However, this approach is limited in capturing data relative to partially invasive and invasive BCIs. Through electroencephalography (EEG), a test that measures the electrical activity in the brain by recording brainwaves, a headset or small electrodes are placed around the scalp to detect tiny voltage changes. Unfortunately, because the skull acts as a barrier between the passage of neural signals to the electrode, the quality of the data diminishes. Despite being disadvantageous in capturing high resolution data, EEG is still the preferred method for its safety, reliability, and accessibility when assessing neural activity.
In contrast, the semi-invasive approach uses electrocorticography (ECOG) to measure intracranial electrical activity, acquiring neural signals from deeper within the brain and therefore requiring surgical operation. Most commonly used in experimentation and clinical trials, this approach serves as a compromise that allows for the acquisition of high resolution data without fully penetrating brain tissue, which is the case for invasive BCIs.
On the further end of the spectrum, invasive BCIs are placed deep into the cortex of the brain. While this BCI is implanted in an area of the brain that is prime for acquiring high resolution neural data, this compromises the user’s safety due to the increased risk of infection and brain tissue damage, and the extensive surgical operation is costly. These risks have made non-invasive and semi-invasive BCIs more preferable in research environments and medical applications.
After the BCI acquires signals and neural data, the data is processed through the signal processing subsystem, which consists of three units: preprocessing, feature extraction, and classification. AI plays a significant role in enhancing each of these units. In preprocessing, the goal is to filter noise from neural data, often contaminated by environmental distractions and other small muscle movements such as eye blinking. Traditional BCIs use the Independent Component Analysis (ICA), a computational method that identifies separate commands or signals from acquired neural data. This denoising method has performed well in traditional models, but is time-consuming and inefficient as it requires individual judgement for each component. However, researchers are utilizing recent advancements in Generative Adversarial Networks (GAN) to complete supervised and unsupervised tasks, fronting a new approach in artificial denoising for EEG data. The neural network model consists of a generator, which uses generative AI, a form of AI that creates new content or information based on existing training data, to create synthetic training data to compete with a discriminator, which determines whether noise is filtered out (Yang et al., 2022). Other novel advancements include Starlab’s Artifact Removal Autoencoder for EEG recordings (LSTEEG), an autoencoder that uses LSTM layers, neural networks that specialize in retaining information over longer processes, to detect and correct artifacts or unwanted noise in EEG data (Aquilué-Llorens et al., 2025). ICA, GAN, and LSTEEG reflect the progression from traditional techniques to leveraging deep learning and AI for signal preprocessing in BCIs, streamlining the removal of noise from neural data.
Leveraging neural networks and AI have also been advantageous for feature extraction in BCIs. Before the introduction of neural network technology, traditional BCIs struggled to interpret the complex characteristics of EEG data, leading to poor performance in signal processing. Fortunately, automatic feature extraction methods overcome the limitations of manual feature identification. Researchers utilize Convolutional Neural Networks (CNNs), a deep learning technique that processes data through several layers, to extract discrete features and spatio-temporal information for later classification of commands (Ding et al., 2025). In addition to CNNs, transformer-based models such as EEG-PatchFormer use a self-attention mechanism to decode different signals from varying regions of the brain while accounting for other logistics, such as temporal information. Once the neural data has undergone preprocessing and feature extraction, AI can improve the accuracy of classifying these signals, determining which command a user wants to execute. Researchers often consolidate CNNs, LSTM, and transformer models into a single model that is capable of highly accurate classification and rapid adaptation to user patterns.
Finally, the filtered, classified data is translated into commands for real-life applications. An example of this is Neuralink’s BCI, an AI-based invasive BCI that excels in translating user thoughts into tangible actions. Elon Musk founded Neuralink in 2016 with a goal of creating a device that enables the remote control of external devices through a user’s thoughts. Eight years later, Neuralink’s BCI was successfully implanted into the brain of their first patient. Using Neuralink’s device, a paralyzed user regained autonomy in controlling a computer cursor through their thoughts, allowing them to access the internet and even play video games. Looking ahead, Neuralink and other neurotechnology companies anticipate applications in vision restoration, an achievement that would be life-changing for many individuals.
Other medical applications have also revolutionized prosthetic technology and rehabilitation after paralysis, which occurs as a result of a disconnection between the brain’s motor cortex and the central nervous system. In this regard, BCIs introduce the possibility of regaining mobility through assistive devices and rehabilitation. Researchers at UC San Francisco created a robotic arm that could be controlled via the user’s imagination. After training to imagine and visualize simple movements, the participant’s neural activity trained an AI model to decode his commands into movement. With each practice session, the accuracy of the AI model’s translations improved, and the achieved motor functions became increasingly complex. In stroke rehabilitation, too, AI-powered BCIs have significant potential to restore motor functions and communication. By predicting a user’s intentions, the BCI can stimulate damaged brain areas and successfully replicate and restore lost neural pathways.
As the rise of AI transforms the face of medicine, medical professionals and patients alike anticipate vast improvements and capabilities in technologies like BCIs. These devices not only serve as bridges for neurological disconnections, but can also enhance cognitive abilities such as improving working memory. While AI-powered BCIs are still in their experimental stages, hopes to bring them into real-world settings have promising impacts on the lives of individuals living with neurological impairments. Before these devices can be released for public medical use, researchers must address several ethical considerations. Many are concerned about the privacy of users’ thoughts as improvements in precise data collection and translation increase access to user’s emotions and thoughts, potentially revealing confidential or sensitive information. This intrusion of a patient’s privacy violates the fundamental guidelines of scientific research such as patient autonomy and informed consent. Additionally, authoritarian regimes may utilize the power to monitor thought for malicious purposes such as surveillance, posing a severe threat to human rights. Patient safety also remains a concern, as using invasive BCIs like Neuralink risks brain damage and infection. These threats to patient safety and privacy call for the necessary regulation of BCI development and usage. Finally, developers must consider the affordability of these devices and BCIs, as machine learning infrastructure and training is expensive. Despite these challenges, advancements in BCIs through increased computational
power and the integration of AI are revolutionizing human-computer interaction, enabling BCIs to have profound impacts on medicine, cognitive enhancement, and accessibility.
References
An, Y., Lam, H. K., & Ling, S. H. (2022). Auto-Denoising for EEG Signals Using Generative Adversarial Network. Sensors, 22(5), 1750. https://doi.org/10.3390/s22051750 Aquilué-Llorens, D., & Soria-Frisch, A. (2025). EEG Artifact Detection and Correction with Deep Autoencoders. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2502.08686
Brain-Computer Interfaces and AI: Technology Overview and Current Capabilities. (2024, December 30). BAYLOR AI. https://baylor.ai/?p=5303
Ding, Y., Lee, J. H., Zhang, S., Luo, T., & Guan, C. (2025). Decoding Human Attentive States from Spatial-temporal EEG Patches Using Transformers. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2502.03736
Haseltine, W. A. (2024, August). The Need for Ethical Regulation of Brain-Machine Interface Technologies. Inside Precision Medicine. https://www.insideprecisionmedicine.com/topics/translational-research/the-need-for-ethic al-regulation-of-brain-machine-interface-technologies/
Hossain, K. M., Islam, Md. A., Hossain, S., Nijholt, A., & Ahad, M. A. R. (2023). Status of deep learning for EEG-based brain–computer interface applications. Frontiers in Computational Neuroscience, 16. https://doi.org/10.3389/fncom.2022.1006763
Javaid, M. A. (2013). Brain-Computer Interface. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.2386900
Lewington, L. (2025, March 23). The man with a mind-reading chip in his brain, thanks to Elon Musk. BBC. https://www.bbc.com/news/articles/cewk49j7j1po
Marks, R. (2025, March 6). How a Paralyzed Man Moved a Robotic Arm with His Thoughts. How a Paralyzed Man Moved a Robotic Arm with His Thoughts | UC San Francisco. https://www.ucsf.edu/news/2025/03/429561/how-paralyzed-man-moved-robotic-arm-his thoughts
Shih, J. J., Krusienski, D. J., & Wolpaw, J. R. (2012). Brain-Computer Interfaces in Medicine. Mayo Clinic Proceedings, 87(3), 268–279. https://doi.org/10.1016/j.mayocp.2011.12.008
Zhang, X., Ma, Z., Zheng, H., Li, T., Chen, K., Wang, X., Liu, C., Xu, L., Wu, X., Lin, D., & Lin, H. (2020). The combination of brain-computer interfaces and artificial intelligence: applications and challenges. Annals of Translational Medicine, 8(11).