Can Brain Scans Read Minds? Neural Decoding and Its Limits
Author: Ukrainian Psychological Hub · Published: October 12, 2026 · Editorial Policy
Can a brain scan reveal what someone is thinking? Scientists can sometimes infer a viewed image, a category of words, the general meaning of a story, or an intended utterance from measured brain activity. These achievements are real and increasingly sophisticated. Yet every result depends on the recording technology, the task, the training data, the computational model, and the accuracy standard. The phrase 'mind reading' hides those conditions and often makes a specific laboratory prediction sound like unrestricted access to private experience.
The essential distinction is between detecting a statistical relationship and recovering a person's complete mental life. Modern neural decoding estimates particular features of cognition under defined conditions; it does not provide a universal transcript of thought. The evidence comes from studies of noninvasive functional MRI, vision reconstruction, invasive speech interfaces, and experimental decoding of inner speech. The clearest benchmark is the original semantic reconstruction study by Tang and colleagues, together with subsequent work on internal speech and neuroprostheses.
The short answer: what can a brain scan actually reveal?
Brain activity can carry information about what a person is viewing, hearing, preparing to say, or doing in a controlled experiment. A decoder trained to recognize a limited set of choices may identify which of several pictures was shown. A different decoder may generate a text description broadly consistent with the meaning of a story. In selected patients with implanted electrodes, specialized systems can translate attempted speech into words or synthesized voice. These are distinct tasks with different evidence and practical requirements.
A conventional structural MRI image depicts anatomy, not a person's current sentence or belief. Functional MRI (fMRI) measures blood-oxygen-level-dependent (BOLD) changes associated with neural activity. EEG records differences in electrical potential at the scalp; MEG records magnetic fields generated by neural currents. Surgically implanted arrays can record neural activity closer to its source. A model must be calibrated to map a chosen signal onto a chosen target. See our guide to MRI, fMRI, PET, and EEG for the underlying measurement distinctions.
Established evidence supports classification and reconstruction of selected task information. Preliminary research explores more flexible visual and linguistic decoding, better transfer between participants, and faster communication. The disputed leap is the claim that one device could silently and accurately identify any unexpressed belief, memory, intention, or secret in any individual. Published studies do not establish that capability.
What is neural decoding?
Neural decoding is the use of measured nervous-system activity to estimate a variable of interest, such as a stimulus category, a motor command, a word sequence, a visual feature, or a behavioral state. The decoder is a statistical or computational model. It learns associations from examples in which researchers know both the neural measurement and the relevant stimulus, task, action, or report. Its output is a prediction constrained by what it learned and by what the recording can resolve.
Decoding is often paired with neural encoding. In an encoding model, researchers predict a neural response from known stimulus features. In a decoding model, they estimate stimulus or task features from a neural response. These directions serve related scientific goals but have different errors. A successful prediction also does not imply that the model has discovered the unique biological mechanism generating the mental process. The broader discipline and its inferential logic are introduced in cognitive neuroscience.
Some decoders perform classification: they choose among a predefined set of classes. Others perform regression: they predict a continuous value such as a movement velocity or acoustic feature. Generative systems produce a new sentence, sound, or image conditioned on neural data, frequently aided by a language model or diffusion model. Generative output can be fluent and compelling even when several details were supplied mainly by the model's prior expectations rather than by the recorded brain signal.
How does a neural decoder learn to make a prediction?
A typical experiment starts with a designed task. Participants may watch images, listen to stories, imagine speaking specific words, or attempt a movement. The experimenter records brain signals and gathers labels or ground truth, such as the displayed image, the exact story, the intended target phrase, or the instructed movement. A training procedure estimates a mathematical mapping between signal patterns and target variables. Researchers then assess the model on data withheld from its training set.
The details determine what success means. With two possible pictures, 80% accuracy describes a very different challenge from recognizing one image among thousands. A model that identifies an object category may still be unable to reproduce the original image. A text system that captures general meaning may invent specific names, wording, sequence, or causal relationships. Performance must therefore be reported alongside the number of alternatives, signal type, training duration, subject count, test conditions, and scoring criteria.
High-dimensional models can learn unwanted shortcuts. They may exploit repeated stimuli, temporal autocorrelation, scanner or electrode artifacts, or similarities between the training and test materials. Strict separation between training and testing, held-out sessions, genuinely novel stimuli, and preferably independently collected datasets are crucial. Generalization across different people matters especially because brain anatomy, functional organization, noise, task strategies, and electrode positions vary.
What brain measurements are used for decoding?
Functional MRI: spatially extensive but physiologically indirect
Functional MRI can measure BOLD responses across much of the brain while a person remains in a scanner. These signals arise through changes in blood flow, blood volume, and oxygenation linked to neural activity. They are not direct readings of thoughts or individual neuronal spikes. Their hemodynamic time course is slow relative to the events in a sentence or a rapid sequence of visual experiences. Interpretation also depends on neurovascular physiology and motion control. The review of BOLD mechanisms by Hillman explains why a BOLD pattern must be treated as a complex indirect signal. A dedicated methods article, Functional MRI: How fMRI Measures Brain Activity and Its Limits, explains how the vascular signal is obtained and what it can support.
An fMRI voxel averages information over many cells and vascular structures. Although maps can be spatially detailed, a single voxel does not have a single psychological meaning. Semantic decoding systems sometimes combine thousands of measurements and computational priors to estimate broader patterns. Spatial coverage permits distributed-pattern analysis, but a reconstructed story is an inference from data, not a recording of a sentence stored verbatim at a brain location.
EEG and MEG: fast signals with localization challenges
Electroencephalography records fluctuating electrical potentials at the scalp, arising mainly from coordinated postsynaptic activity of neural populations. Magnetoencephalography measures associated magnetic fields. Their millisecond-scale sensitivity can support temporal analyses and some classifications of perception, attention, speech-related activity, or imagined movement. However, signals from multiple brain sources mix at the sensors, and eye movements, muscles, electrical noise, and source-localization assumptions limit interpretability.
A headset advertised as an EEG 'mind reader' may detect a small set of well-trained actions or mental tasks in a particular setup. That result does not demonstrate sentence-by-sentence access to spontaneously occurring thoughts. Reports should distinguish controlled trial accuracy from dependable performance during ordinary movement, distraction, fatigue, different devices, and use outside the lab.
Implanted electrodes: richer local signals and different risks
Electrocorticography (ECoG) records activity from electrodes placed on the cortical surface, whereas penetrating intracortical arrays measure activity with finer local specificity, including spikes or related population signals. These approaches can support fast, high-information decoding of intended movement and speech. Their clinical possibilities are substantial, but they involve surgery, maintenance, safety considerations, constrained electrode placement, patient-specific adaptation, and small research cohorts.
Evidence obtained with implanted arrays cannot be assumed to apply to a routine hospital MRI scan, a smartphone, or a consumer EEG headband. The information content, access to neural sources, clinical context, and achievable performance are fundamentally different.
What have researchers actually decoded?
Recognizing what someone sees
Early landmark work showed that visual-cortex fMRI could help identify which natural image a participant had viewed. In Kay and colleagues' 2008 study, models learned how patterns of visually driven activity related to image features. The problem was one of identification among candidate images under experimental conditions, not arbitrary retrieval of a hidden private visual experience.
Researchers later produced approximate reconstructions of natural movie content from fMRI signals. The 2011 study by Nishimoto and colleagues used an encoding model and a prior over candidate visual material to infer movie-like structure. The resulting sequences can reflect aspects of motion and scene organization, but they should be evaluated as model-generated approximations. A reconstruction is not a camera recording from inside the head.
Generating images with modern AI models
Image generation systems combine fMRI-derived representations with pretrained generative models. Some impressive reconstructions preserve a recognizable object class or broad composition. Yet visual realism is a poor substitute for faithful recovery of the exact stimulus. The image may contain plausible textures, faces, colors, objects, or details that the measured brain data did not determine. Image-to-image similarity metrics, category matching, and human judgments reveal different aspects of performance.
A strong caution comes from Shirakawa and colleagues' 2025 analysis. It found that some text-guided methods produced misleadingly plausible reconstructions and performed poorly under more demanding tests of generalization. The authors highlighted restricted training data, limited diversity, and the influence of generative priors. This methodological challenge applies especially when media headlines display a striking best example without showing failures, mismatches, or baseline comparisons.
A newer development is Brain-IT, presented for ICLR 2026. Its authors reported improved image reconstruction by integrating brain-voxel clusters with localized visual features, with comparable performance to some previous approaches using far less new-participant calibration data. Those results are relevant to transfer learning but remain bounded by the specific datasets, visual tasks, metrics, and model architecture. The claim is improved reconstruction under an experimental evaluation, not general-purpose decoding of dreams or memories.
Estimating the meaning of heard or imagined language
In 2023, Tang and colleagues published a noninvasive semantic decoder that generated intelligible language related to the meaning of stories participants heard. They also investigated imagined storytelling and silent videos. The system combined fMRI activity with a computational model that evaluated plausible language sequences. Its outputs could capture the gist rather than reproduce the original wording. Semantic similarity should not be confused with word-for-word transcription.
The study's practical conditions are central to interpreting it. Training involved extensive participant-specific fMRI data, including many hours of listening to narrated stories. Decoding from models trained on different participants performed much worse. Participants could also reduce successful decoding by performing competing mental tasks. These findings demonstrate a research capability and the importance of cooperation in that experiment; they do not prove an immutable technical barrier that will protect everyone under every possible future system.
The language model contributes substantial prior knowledge about what sentences are likely. An apparently coherent sentence can be an optimal guess among model candidates, even where the neural signal provides insufficient evidence for particular words. Researchers therefore need to evaluate both whether a prediction is semantically associated with the test stimulus and which details are genuinely supported by the brain recording.
Decoding attempted speech for communication
The clearest clinical progress concerns people who cannot speak because of paralysis or neurological injury but can attempt or intend speech. These systems generally decode activity related to the production of speech, rather than the full contents of private thought. In a 2023 Nature study by Willett and colleagues, intracortical recordings supported high-performance attempted-speech decoding in a research participant. Another 2023 study by Metzger and colleagues used cortical surface recordings to enable text, synthesized speech, and avatar control.
More recent systems have advanced the timing and expressivity of output. Littlejohn and colleagues (2025) reported continuously streaming synthesized speech from high-density cortical recordings in a person with severe paralysis. A separate 2025 Nature study demonstrated low-latency brain-to-voice synthesis from implanted microelectrodes in a participant with ALS. Such systems are early clinical research technologies, and individual demonstrations should not be mistaken for broadly established efficacy across patients, diagnoses, and care environments.
What about silent inner speech?
Inner speech is language-like activity experienced without overt speaking. Its neural representations can overlap with attempted or overt speech, although the exact relation depends on task and brain region. Wandelt and colleagues (2024) found that a small predefined vocabulary of imagined words could be decoded above chance from implanted recordings in two participants. Importantly, decoding scores varied substantially by participant, and the test was limited to instructed items.
The 2025 Cell study by Kunz and colleagues provided stronger evidence that elements of inner speech can be decoded from implanted motor-cortex recordings, including sentence-level tasks. It also found that some private, unrequested inner speech might appear during other tasks, raising a concrete design problem for assistive systems. The team investigated safeguards that distinguished intended output from inner speech and permitted controlled unlocking. These results justify privacy-by-design, while their participant numbers, device type, calibration, and evaluation setting still sharply limit generalization.
Why decoding success does not equal mind reading
The decoder only predicts what it was designed and trained to predict
Every decoder has a target. If the target is 'which of four images appeared,' an accurate answer provides evidence about that experiment, not an unrestricted description of what the participant believes. A decoder trained on attempted speech can exploit speech motor representations even if it offers little information about other cognitive activity. A model trained to detect a mental state from a narrow task may not generalize to different causes of the same signal.
Brain activity is distributed and context dependent
Complex perceptions and mental acts emerge through interacting neural populations and systems. One cortical region often participates in multiple processes, and the same process can recruit different distributed configurations depending on task, learning, physiological state, and individual variation. A colorful activation map does not establish that the highlighted region contains a discrete thought. For an explanation of network-level organization, see neural circuits and brain networks.
The classic problem of reverse inference arises when someone sees activation in a brain area and concludes that a specific mental process must be occurring. Poldrack's 2006 analysis shows why the strength of such inferences depends on the selectivity of the association and the alternatives. Decoding models can sometimes produce valid predictions in a narrow context; that does not turn a statistical correlation into a uniquely identified psychological cause.
Generative models fill informational gaps
Modern AI models make decoding outputs appear more precise by supplying fluent language and richly detailed imagery. The generating system combines neural observations with knowledge it acquired from large datasets. When the observation is weak or ambiguous, the prior can dominate. A model's ability to produce a realistic face or grammatical sentence does not demonstrate that the person actually saw that face or silently formed those exact words. Failure analysis and error calibration matter more than a handful of attractive examples.
Training and generalization remain major barriers
Training on the same person, repeated stimuli, or the same scanning configuration can work well while cross-person and real-world performance remains weak. Structural and functional alignment is an active research area, and models with shared parameters may reduce individual calibration requirements. Yet a strong result in one dataset does not establish transfer across ages, neurotypes, languages, scanners, sleep states, clinical populations, or independent laboratories. The 2025 review of linguistic neural decoding by Wang and colleagues surveys these technical and evaluation challenges.
Thoughts are not all verbal, reportable, or stable
Conscious experience can include language, mental imagery, affective and bodily sensations, memories, intentions, partial associations, and fleeting attention shifts. A reportable sentence is only one possible target. Experiments that reconstruct perceived stories do not thereby measure the totality of subjective experience. Nor can decoding of neural correlates alone settle philosophical questions about consciousness or establish a unique mapping from all mental content to a readable signal.
How should neural decoding accuracy be evaluated?
A reported percentage is meaningful only with a clear denominator and baseline. For classification, researchers should report the size and distribution of the candidate set, chance performance, confusion patterns, performance on unseen examples, and uncertainty intervals. For speech, they may report word or character error rate, communication rate, latency, vocabulary size, and whether language-model corrections are included. For visual reconstruction, quantitative similarity measures should be supplemented by tests of faithfulness and out-of-distribution generalization.
The most revealing question is whether a decoder retains predictive value after reasonable alternative explanations have been ruled out. Could it infer the answer from repeated stimuli, prompts, overt behavior, task timing, auditory cues, gaze, or dataset regularities? Did the model train on related material? Were best-case images chosen for display? Did independent participants or laboratories obtain comparable results? Was an ablation performed to reveal how much output came from the generative prior rather than the brain?
Population-level decoding should not be presented as an individual clinical test without evidence of reliability, calibration, diagnostic accuracy, and benefit in its intended clinical setting. Predicting a laboratory label, identifying a clinical symptom, diagnosing a disorder, and improving health outcomes are separate evidentiary claims. Brain imaging is frequently useful in medical diagnosis for neurological and structural indications, but these established applications do not validate 'thought reading' as an individual mental-health diagnostic service.
Can someone read your thoughts without your consent?
Present-day published noninvasive semantic decoding does not support covert, distance-based, arbitrary thought extraction. An fMRI scanner is large, requires a participant physically inside the device, and in the foundational semantic decoding experiments needed extensive personal training data and cooperation. EEG and MEG also require sensors and appropriate measurement conditions. A phone camera, ordinary security camera, or satellite does not provide the neural recordings used in these studies.
Privacy concerns nevertheless deserve serious treatment. Implanted interfaces and future more transferable decoding models could expose information that users did not intend to communicate. Kunz and colleagues' 2025 research made this issue experimentally concrete by measuring unintended inner speech in certain tasks and testing technical safeguards. The correct inference is to develop consent controls and protections before systems scale, not to claim that everyday devices already reveal everyone's hidden thoughts.
Neural data can also be sensitive independently of successful sentence reconstruction. A dataset may contain health, identity, task-response, or behavioral information; linkage and secondary use can create risk. Data minimization, purpose limitations, meaningful consent, on-device computation where feasible, retention controls, access security, auditability, and user-controlled output are sensible governance requirements. The UNESCO Recommendation on the Ethics of Neurotechnology, adopted in 2025, frames mental privacy and human rights as core concerns.
Could brain scans reveal lies, political beliefs, hidden memories, or criminal intent?
A successful experiment decoding perception or instructed speech does not validate a general-purpose lie detector. Deception is context sensitive; brain responses related to conflict, arousal, memory, and decision-making are shared with many nondeceptive situations. Inferring a unique lie from a multivariate brain pattern would require direct evidence of accuracy in realistic, adversarial circumstances with independent validation and defensible error rates. Neural decoding studies of stories and images do not supply that evidence.
Likewise, the presence of a memory-related or emotional response cannot establish the legal truth of a recalled event, infer intent independently of context, or certify a person's moral belief. People may imagine scenarios they reject, experience involuntary thoughts, or recognize information without endorsing it. Treating an uncertain model output as authoritative would impose substantial risks of false accusation and intrusion.
Can brain scans read dreams?
Some sleep and imagery studies have investigated correspondences between neural patterns, stimulus categories, and reports collected after awakening. A model might infer certain broad categories under specially controlled conditions. Yet dream reports are themselves incomplete recollections, and sleep-state measurements differ from waking imagery experiments. The leap from category decoding to a reliable, frame-by-frame dream video remains unsupported by the studies discussed here. Generative reconstructions should not be marketed as direct playback of dreams.
Can neural decoding help people who cannot speak?
Speech neuroprostheses are among the most compelling translational applications. If a person retains cortical processes for planning or attempting speech but cannot produce audible words, a trained neural interface may transform some of that activity into text or voice. The intended purpose, informed participation, personal calibration, and user feedback distinguish these systems from sensational accounts of involuntary mind reading.
Current reports involve small numbers of carefully selected participants, different medical conditions, implanted sensors, specialized clinical teams, and varying language and accuracy measures. Research outcomes cannot be generalized automatically to every person with paralysis, stroke, ALS, aphasia, or another communication disorder. No single published demonstration establishes an approved universal treatment. For readers seeking to understand the underlying techniques, the 2024 review of speech neuroprostheses synthesizes the field's neurobiological and translational foundations.
What is the role of artificial intelligence in mind-reading headlines?
Machine learning enables high-dimensional prediction that simpler models often cannot achieve. Language models can rank candidate sentences and generate plausible word sequences, while diffusion models synthesize images conditioned on neural features. Better machine learning may increase usable information and reduce the amount of calibration data, as suggested by Brain-IT. However, computational performance is jointly determined by the brain signal, model assumptions, training distribution, and evaluation. Increasing model fluency alone cannot recover information that was never reliably measured.
Brain-inspired AI and neuroscience have productive reciprocal relationships: neuroscientists use algorithms to test neural representations, and AI researchers sometimes borrow conceptual architectures from brain science. Equating an artificial model's output with a participant's exact conscious content confuses a statistical reconstruction with the phenomenon it is intended to estimate. The practical question is how much information comes from the recording itself, and how that contribution is validated against genuine alternatives.
Common claims and their scientific status
'An MRI can read anyone's private thoughts': unsupported. Standard MRI measures anatomy; specialized fMRI decoding requires appropriate recordings, task-specific models, and validation. 'Researchers can infer the gist of a story from fMRI': demonstrated experimentally under extensive participant-specific training, with important errors and constraints. 'AI can reproduce exactly what you saw': unsupported as a general claim; generated visual content can be plausible while unfaithful.
'Brain implants can translate intended speech into communication': supported by early clinical studies in selected participants. 'An implant can sometimes detect inner speech': supported in specific, limited research settings. 'Neural decoding is already a reliable psychiatric test': unsupported as a general claim. 'A neural decoder's output proves that a person had a specific belief or intention': unsupported without direct, context-specific validation. The terms mind reading, telepathy, neural decoding, and brain–computer interface should not be treated as interchangeable.
How to read a brain-decoding headline critically
First identify the recording modality: ordinary structural MRI, fMRI, EEG, MEG, or implanted electrodes. Then ask exactly what was decoded: a category, an instructed item, the meaning of a presented story, attempted speech, or spontaneous private experience. Identify whether the system required hours of calibration, whether the participant was cooperating, how many people took part, and whether the study used data from unseen subjects and tasks. Finally, inspect the performance metric, errors, model priors, independent replication, and comparison with behavioral or non-neural information.
Headlines that say 'scientists read thoughts' often compress all these questions into one memorable phrase. A more faithful account names the signal, the target, the limits, and the reason the technique is useful. For example, 'a calibrated fMRI system reconstructed aspects of story meaning in cooperative research participants' communicates the 2023 semantic decoding result without promising universal access to mental contents.
Frequently asked questions
Can an MRI scan show what I am thinking?
A routine structural MRI does not show a transcript or image of your thoughts. Specialized fMRI experiments can sometimes predict task-specific information from blood-oxygenation patterns when paired with a trained decoder. Results depend on the study design and do not represent free access to all mental activity.
Can scientists decode an exact sentence from my brain?
Implanted speech interfaces may produce relatively accurate text from attempted or imagined speech in selected, trained participants, with measurable errors. Noninvasive semantic fMRI decoding can produce an intelligible paraphrase or approximate meaning in an experimental setting, not a guaranteed verbatim sentence.
Does neural decoding work on strangers without training?
The 2023 noninvasive semantic decoder performed far worse when trained using different people's data. New methods attempt cross-person generalization, including work reported in 2026 on visual reconstruction, but that is not equivalent to dependable decoding of arbitrary thoughts from an unknown person.
Can EEG headphones read my thoughts?
EEG measures scalp electrical signals. Under designed tasks, it may permit prediction of a selected mental or motor state. Everyday, unrestricted sentence-level mind reading through ordinary consumer EEG has not been established by the cited clinical and laboratory studies.
Can a brain implant read thoughts I do not wish to say?
Specific invasive research has found neural information about inner speech, including some activity during tasks where participants were not trying to communicate it. These findings motivate user-intent detectors, privacy locks, and controlled output, but do not establish complete or infallible access to all thoughts.
Can AI reconstruct images from memories or dreams?
AI can generate approximate images conditioned on certain neural measurements, especially in experiments involving viewed pictures. The realism of a reconstruction does not prove faithful recovery of unseen memories or dreams. Greater generalization and independent validation are required.
Can neural decoding diagnose depression, anxiety, or another disorder?
Neural decoding of laboratory stimuli is separate from clinical diagnosis. A clinically useful test requires validated performance for its intended population and decision, appropriate sensitivity and specificity, and evidence that it improves care. Research-level brain associations should not be represented as individual diagnoses.
What is the most important practical takeaway?
Evaluate every claim using the five elements of the experiment: signal, task, training, output, and validation. Real advances in decoding are meaningful for brain science and assistive communication. What they reveal is specific information from measured neural activity under defined conditions.
Conclusion
Brain scans can support genuine neural decoding, and implanted interfaces can restore elements of communication in selected experimental circumstances. The strongest evidence concerns precisely defined stimuli and tasks, participant-specific training, and validated predictions. Modern generative AI can make outputs more fluent and visually persuasive, so interpretation must separate signal-derived information from statistical expectation. Neural decoding is a powerful scientific method whose achievements become most understandable when its measurements and limitations are described with equal precision.
Related Articles
References
Beliy, R., Zalcher, A., Kogman, J., Wasserman, N., & Irani, M. (2026). Brain-IT: Image reconstruction from fMRI via Brain-Interaction Transformer. ICLR 2026, conference paper; arXiv:2510.25976. Primary manuscript.
Hillman, E. M. C. (2014). Coupling mechanism and significance of the BOLD signal: A status report. Annual Review of Neuroscience, 37, 161–181. DOI: 10.1146/annurev-neuro-071013-014111.
Kay, K. N., Naselaris, T., Prenger, R. J., & Gallant, J. L. (2008). Identifying natural images from human brain activity. Nature, 452, 352–355. DOI: 10.1038/nature06713.
Kunz, E. M., et al. (2025). Inner speech in motor cortex and implications for speech neuroprostheses. Cell, 188, 4658–4673.e17. DOI: 10.1016/j.cell.2025.06.015.
Littlejohn, K. T., et al. (2025). A streaming brain-to-voice neuroprosthesis to restore naturalistic communication. Nature Neuroscience, 28, 902–912. DOI: 10.1038/s41593-025-01905-6.
Metzger, S. L., et al. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature, 620, 1037–1046. DOI: 10.1038/s41586-023-06443-4.
Nishimoto, S., et al. (2011). Reconstructing visual experiences from brain activity evoked by natural movies. Current Biology, 21, 1641–1646. DOI: 10.1016/j.cub.2011.08.031.
Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10, 59–63. DOI: 10.1016/j.tics.2005.12.004.
Shirakawa, K., et al. (2025). Spurious reconstruction from brain activity. Neural Networks, 190, 107515. DOI: 10.1016/j.neunet.2025.107515.
Tang, J., LeBel, A., Jain, S., & Huth, A. G. (2023). Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience, 26, 858–866. DOI: 10.1038/s41593-023-01304-9.
UNESCO. (2025). Recommendation on the Ethics of Neurotechnology. Official adopted instrument.
Wandelt, S. K., et al. (2024). Representation of internal speech by single neurons in human supramarginal gyrus. Nature Human Behaviour, 8, 1136–1149. DOI: 10.1038/s41562-024-01867-y.
Wang, Y., et al. (2025). Progress, challenges and future of linguistic neural decoding with deep learning. Communications Biology, 8, article 1350. DOI: 10.1038/s42003-025-08511-z.
Willett, F. R., et al. (2023). A high-performance speech neuroprosthesis. Nature, 620, 1031–1036. DOI: 10.1038/s41586-023-06377-x.