Inner Speech in Motor Cortex: Laying the Foundation for Privacy‐Preserving Speech Neuroprostheses

A recent study published in Cell by Erin M. Kunz et al., represents a milestone achievement in brain–computer interface (BCI) research 1. For the first time, decoding via microelectrode arrays in the motor cortex has demonstrated that the ventral precentral gyrus (area i6v) can simultaneously represent signals of inner speech, attempted speech, perceived speech, and silent reading 1. This functional convergence within brain regions associated with articulation may reflect the brain's lateralized architecture, whose core principle is to optimize neural processing efficiency through hemispheric specialization and the reduction of interhemispheric redundancy 2. The study revealed that these language modalities share highly similar geometric structures in their neural representations. They can be accurately distinguished via a content-independent “motor-intent dimension.” The study identified a content-independent “motor-intent dimension” that distinguishes different speech modalities. The team then used this to develop a BCI that suppresses misdecoding of inner speech through “imagery-silenced training”, combined with a keyword lock system, achieving 98.75% accuracy in thought protection. This work reveals the diverse language functions within the motor cortex and offers an efficient communication solution for patients with locked-in syndrome. However, applying inner speech to BCIs has faced both theoretical and technical challenges. Millions of people worldwide are affected by speech disorders, whereas current BCI technology relies heavily on attempted speech 3. This approach is not only prone to fatigue but also has inherent limitations in output speed 4. Although previous studies have shown that motor cortex activity can still be driven by intention after spinal cord injury 5, and the supramarginal gyrus (SMG) has been found to contain shared representations of inner and actual speech, fundamental questions remained unanswered. These included the specific encoding mechanisms of inner speech in the motor cortex, its neural relationship with attempted speech 6, and its feasibility for building efficient and reliable communication systems 7. It is against this backdrop that Kunz et al.'s analysis of the neural geometry of inner speech in the motor cortex marks a significant conceptual and technical advance 1. By parsing neural signals from the motor cortex, the study revealed a core neural representational architecture shared across speech modalities (Figure 1). The researchers found that cortical areas responsible for articulation (e.g., i6v) also encode inner speech, auditory perception, and silent reading, with these representations sharing highly similar geometric structures 8. Inner speech can be viewed as a “scaled-down” version of attempted speech, exhibiting identical neural activity patterns but with weaker modulation strength (Figure 2). This explains why inner speech does not trigger motor output, providing direct evidence for the “activation threshold” hypothesis 9. More profoundly, even in a completely anarthric participant (T17), the motor cortex could still distinguish speech states based on “intent” rather than actual movement 10. This strongly suggests that language-related activity in the motor cortex is fundamentally linked to motor intention itself rather than being dependent on the final execution of movement. Schematic summarizing the experimental approach and key results related to decoding various speech modalities from motor cortex. The instructed-delay paradigm used to collect neural data during different speech behaviors (e.g., attempted speech and inner speech). Results of the decoding analysis (10-fold cross-validation) showing that specific arrays in ventral precentral gyrus (i6v) significantly decoded all tested behaviors in 3 out of 4 participants. Key conclusions: neural representations across behaviors are shared but scaled in intensity (supporting the “activation threshold hypothesis”), and a speech BCI trained on attempted speech can also decode inner speech. Schematic of orthogonal content and intent coding in motor cortex. This conceptual illustration depicts the “content-sharing, intent-separation” principle, where neural representations of three words are shown as colored points. For each word, attempted speech (top cluster) and inner speech (bottom cluster) occupy similar positions in the two horizontal content dimensions but are separated along the vertical intent dimension, indicating that speech content and intention are encoded in orthogonal neural subspaces. Leveraging these mechanisms, further investigations into serial recall and counting mediated by inner speech demonstrated that the motor cortex also participates in naturally occurring cognitive processes 11. Without any instruction, participants' brain activity showed decodable signals corresponding to incrementing number sequences during silent counting (Figure 3). Moreover, structured linguistic tasks (e.g., recalling song lyrics) elicited significantly stronger neural responses than nonverbal tasks 12. These findings indicate that inner speech, as a cognitive tool, produces stable and capturable signals. Based on the identified independent “motor-intent dimension,” the team proposed “imagery-silenced training” to suppress the decoding of inner speech combined with a “keyword lock” system for precise control. This makes it possible to construct a technological pathway that simultaneously ensures “mental privacy” and “communication freedom 4.” Schematic overview of decoding uninstructed inner speech during cognitive tasks and strategies to prevent unintended decoding. (A) Design of motor tasks (e.g., three-element arrows) to elicit inner speech. Successful decoding during the delay period indicated its use for verbal working memory. Evidence that naturally occurring inner speech can be decoded as shown by a significant increase in decoded numbers during counting (p < 1e-8) and more decoded words during verbal versus nonverbal thinking. (B) Two proposed methods to prevent accidental decoding of private thoughts: “Imagery-Silenced Training” for attempted speech BCIs, and a “Keyword Lock System” for inner speech BCIs. In summary, this study represents a key advance in motor cortex speech decoding by establishing the ventral precentral gyrus (i6v) as an integrative hub for multiple speech modalities. Particularly important is the discovery of the “motor-intent dimension,” which provides a novel neural mechanism for distinguishing communicative intent without decoding thought content, thereby directly addressing a central ethical concern in BCI applications. The development of a real-time inner speech decoding system supporting a 125,000-word vocabulary represents a substantive step toward clinical translation 13. Furthermore, the dual-track privacy strategy integrating “imagery-silenced training” and a “keyword lock system” demonstrates a forward-looking design philosophy that embeds ethical considerations into the technical architecture, setting an important benchmark for the future development of neural prostheses 14. Nevertheless, the study also raises a fundamental question: how can shared neural representations and behavior-specific intent signals coexist? Specifically, although the motor cortex represents different speech behaviors with highly similar content encoding, it simultaneously activates a non-overlapping motor-intent signal to distinguish behavioral states. This parallel encoding mechanism—“shared content, separated intent”—poses a new problem for existing theory: how do these two aspects coordinate within the same neural substrate? The “activation threshold hypothesis” proposed in the study, although explaining differences in signal strength, fails to clarify the generation mechanism of such orthogonal encoding 9. Building on this, we hypothesize that a more intricate biological basis may underlie these phenomena, emerging from the synergistic actions of the motor cortex across different spatial and temporal scales. The separation of intention separation within the same anatomical structure likely results from the functional specialization and dynamic integration of the cortical laminar architecture. This specialization enables the formation of distinct functional within neuronal activity manifolds at the population level 15. On a microscale, the differential processing biases between superficial and deep cortical layers, along with dense interlaminar connections, facilitate the mixing of these two types of information within a shared anatomical substrate 16, 17.Their distinct dynamic properties and influence on downstream circuits may, in turn, foster the emergence of nearly orthogonal “content” and “intent” manifolds within the high-dimensional neural activity space. This functional differentiation and plasticity are likely supported by both the macroscopic structural properties of the brain and its genetic background. Research has shown that experience-driven neural adaptations are not only reflected in the reorganization of functional connectivity but also often correspond to stable morphological changes, such as cortical thickness, which provide a structural foundation for the emergence of specific functional manifolds 18. Simultaneously, genetic factors, mediated by polygenic risk scores, influence susceptibility to psychiatric disorders and likely play a role in predisposing individual differences in cortical circuit plasticity 19. Additionally, the inherent functional lateralization of the brain may further constrain or direct the spatial expression and efficiency of these adaptive processes 2.Consequently, the orthogonal encoding capacity observed in area i6v is likely the result of a multiscale process: genetic and developmental programs establish initial boundaries for structural plasticity and functional lateralization, whereas experience-driven adaptations at the microstructural level and network reorganization at the macroscopic scale collectively support the efficient separation of functional manifolds at the meso-scale 20, 21. At the same time, several practical challenges remain for clinical translation. First, the small participant sample size and specific paralysis etiologies limit the generalizability of the findings 10. More critically, the cross-sectional design of this study does not address a key question for clinical translation: how stable are the higher-level neural representations, such as the “motor-intention dimension” decoded from the motor cortex, over the long term? For a practical brain–machine interface (BCI), the core decoder must be resistant to the neural activity “drift” caused by learning, fatigue, neural plasticity, or disease progression. Existing research shows that the group neural activity used to decode limb movement intentions changes significantly over time, necessitating frequent recalibration of the decoder. Therefore, to account for the potential drift in “intention” signals, future studies must include longitudinal tracking of the same participants for several months or even years, specifically evaluating the stability of the “content” and “intention” neural manifolds, and developing adaptive decoding algorithms that can accommodate such potential drift. Only after validating longitudinal stability can this intention-decoding-based BCI paradigm move toward clinical practicality. Second, the reliance on invasive surgical implantation of microelectrode arrays itself constitutes a major obstacle to clinical adoption. Furthermore, the study does not specifically report the overall decoding delay of the system, which is a fundamental challenge for achieving a natural interaction experience. Research has shown that the human conversational turn-around gap is around 200 milliseconds 22, while even high-performance BCI systems typically have a character selection delay on the order of seconds 23. Reducing decoding delay by several orders of magnitude to achieve near-real-time interaction is another core performance bottleneck that must be overcome for the technology to become practically viable. Additionally, although the privacy protection mechanism has demonstrated high accuracy, its robustness in real-world scenarios requires further strengthening. Although the 98.75% accuracy reported for the keyword lock system is impressive at the concept validation stage, its practical utility depends on the balance between false positives (incorrect unlocking that exposes private thoughts) and false negatives (failure to unlock when needed, hindering communication). In high-risk scenarios, such as communication involving sensitive personal or medical information, even a low false–positive rate may be ethically unacceptable. On the other hand, a system that is too strict and frequently produces false negatives would hinder normal user interaction and reduce the effectiveness of BCIs as reliable communication tools. This tension highlights the importance of privacy protection in neural technologies. Future iterations may need to incorporate adaptive safety thresholds or context-aware unlocking mechanisms that adjust sensitivity based on the user's environment or intention, thereby embedding ethical flexibility into the technology's design. In conclusion, these findings advance our understanding of the cortical brain. They redefine the human motor cortex not as a passive “motor command center,” traditionally responsible for translating high-level plans into low-level muscle commands, but as an active “intelligent intention interface.” This means that the motor cortex's function is not only to generate movement outputs but also to dynamically represent and integrate behavioral intentions, abstract content, and behavioral states. For neuroprosthetic design, this suggests a shift from focusing narrowly on “decoding motor commands” to interacting with a high-level “intention interface.” Future speech neuroprosthetics with privacy protection should aim to dynamically adapt to the encoding rules of this interface (such as content-intention separation) to enable more natural and robust communication. For cognitive neuroscience, this challenges the traditional view of the motor cortex as a peripheral part of the cognitive network, repositioning it as a central hub that connects abstract cognition and embodied action, providing a novel framework for understanding thought, language, and their neural bases. Overall, this study not only provides a groundbreaking proof-of-concept but also, by clarifying the theoretical tensions and technical challenges, lays a forward-looking and responsible foundation for the next generation of privacy-preserving speech neuroprostheses. Chenzhitong Chen: conceptualization, investigation, validation, visualization, writing – original draft, writing – review and editing, supervision. Weiqi Wang: visualization, writing – original draft, writing – review and editing, resources. Xingyu Mou: investigation, data curation. Tingting Tang: formal analysis, investigation. Xinling Chen: investigation, funding acquisition. Qian Wang: conceptualization, funding acquisition, methodology. Xi Yang: visualization, supervision, formal analysis. Xiaosong Hu: visualization, validation, project administration. Zheng Yang: writing – original draft, supervision, resources, writing – review and editing. The figure was generated using BioRender.com. This work was supported by the Sichuan Province Training Program of Innovation and Entrepreneurship for Undergraduates (Grant Nos. S202313705058, S202313705077, and S202313705054) and the National College Students' innovation and entrepreneurship training program (Grant No. 202213705009). This article does not involve any studies with human participants or animals performed by any of the authors. The authors declare no conflicts of interest. No new data were generated or analyzed in support of this commentary. All data discussed are derived from the cited publications and are available from the corresponding sources referenced within the text.