Telepathy-Based Computing: The Race to Own the Next AI Interface

When an AI agent books a trip, settles a payment, or completes a workflow on a user’s behalf, the industry refers to it as intent-based computing. It means the user sets the goal; the AI determines what needs to be done, breaks it down into tasks, and orchestrates a network of autonomous agents to execute it. The race for this opportunity is already on. Technology companies are building personal assistants that act on behalf of users, and service providers are building agentic platforms that turn stated goals into autonomous actions.

Yet the term hides a contradiction. The user still has to express the intent: type a prompt, speak a command, and describe the outcome. AI has removed the need to specify every step, but not the need to translate the user’s intention into language a machine can parse. That may not be true intent-based computing. What if AI could infer intent from the neural activity associated with that intention? This is where neuroscience enters the AI story, and it points toward something that could eventually be called telepathy-based computing.

Interpreting Intent Before It Becomes a Prompt

To understand why this matters, consider what the brain actually does. The brain is, in many ways, an information processing system. It is radically different from a conventional computer, but it receives information, processes it, constructs representations of the world, makes decisions, and sends signals back to the body.

And despite its extraordinary complexity, its connection to the outside world is mediated through a relatively limited set of biological communication pathways. The cranial and spinal nerves effectively form an interface between the brain and the rest of the body. Sensory information enters through these pathways. Motor commands leave through them. In computing terms, one can think of these pathways as something resembling an API for the brain.

The brain does not receive a raw, objective feed of reality. It receives signals and constructs an internal world model from them. The room around you, the objects you see, the sounds you hear, and even aspects of what you perceive as your own body are not direct copies of the external world. They are interpretations constructed by the brain from incoming signals.

This distinction is important for AI. Because if the brain can transform electrical and chemical activity into perception, intention, language, and action, then perhaps machines do not always need to wait for the final output of that process. They could learn to interpret the signals along the way and increasingly process them closer to the source of human intent.

The Race to Decode Human Intent

The technology is still in its early stages, and it would be premature to suggest that machines can simply read arbitrary thoughts. Human thoughts are far more complex than a stream of easily decodable words, but the direction is real. Previously, near-conversational decoding required brain surgery. But over the past 18 months, the decoding of intended language and movement has transitioned from a lab curiosity to open-source code and consumer products.

Three key patterns are emerging:

1. 1. Data scale is now a main driver of progress: Neural decoding is following the same approach as large language models: pretrain on large brain datasets, then fine-tune for each user. Neuralink’s latest gains came from pretraining on recordings it already had, not from new hardware.
   2. Major technology platforms are building neural input into their products: Whoever owns the input layer owns the intent stream. Touch made the smartphone operating system the gatekeeper of mobile computing. Apple’s BCI input protocol and Meta’s Neural Band are early bids to own the equivalent layer for neural input. For platforms, this is a strategic position, not a unit economics calculation.
   3. Commercial products are emerging first at the body’s periphery: Products on the market today read muscle signals at the wrist and jaw. Brain decoding remains in clinical and lab settings, where safeguards against unintended decoding are being developed alongside it.

These patterns show where vendors are placing their bets. For enterprises, the question is whether neural interaction delivers enough value to justify its higher cost.

The Cost Premium and the Business Case

The economics are unlikely to favor telepathy-based computing on a cost-per-interaction basis alone. Neural interfaces introduce additional costs across hardware, signal processing, decoder training, integration, and privacy and governance. Improvements in model efficiency and lower hardware costs at scale may ease some of this burden, but they are unlikely to eliminate the cost premium over conventional prompting.

The business case, therefore, rests on a different equation. Telepathy-based computing must create value that justifies the additional cost, whether by enabling AI interaction when conventional inputs are impractical, reducing errors, or giving agents signals that improve their decisions.

Business Case 1: Keeping AI Accessible When Hands and Voice Are Busy

The first opportunity is access, not speed. Caltech researchers estimate human behavioral output, through typing or speech, at roughly 10 bits per second. In October 2026, Neuralink reported a cursor-control record of 11.32 bits per second. These measurements don’t directly support the claim that neural interfaces will automatically multiply human communication speed. The stronger opportunity is to interact with AI when typing or speaking is impractical.

For enterprises, the value depends on the cost of an interrupted workflow, a delayed decision, or a missed warning. A field engineer working on equipment may need instructions without putting down a tool. A control room operator may need to approve or stop an agent without having to navigate menus. In these settings, even a more expensive interaction can pay if it avoids a costly delay or error.

The proof point already exists in a few use cases today. For instance, mining companies, including Rio Tinto, Anglo American, and Newcrest, use EEG-based SmartCap technology to warn machinery operators of approaching microsleep.

Commercial applications will emerge in stages, starting with detecting fatigue or lapses in attention, moving to simple commands, and eventually decoding more complex thoughts. Useful applications can develop well before machines can translate free-form thoughts into text.

The trade-off is that prompt-based computing wins on cost per interaction, while telepathy-based computing aims to win on value per interaction. At a desk, with a keyboard readily available, conventional prompts remain cheaper and effective. Telepathy-based computing has little advantage unless it delivers a meaningful improvement in access, safety, or interaction quality. Its initial market is, therefore, not every knowledge worker, but the situations in which conventional input becomes a constraint.

Business Case 2: Improving AI Decisions With Signals Users Never Articulate

The four quadrants of knowledge point to a second opportunity: neural input could give AI agents feedback that users may never put into words.

Prompts work well for what people know and what they know they need to learn. But people also make tacit judgments without consciously explaining them. They notice when something looks wrong or recognize a relevant result. Some neural responses associated with these judgments can be measured without an explicit response:

1. 1. Detecting errors: In an MIT CSAIL study, researchers used EEG signals to detect when people noticed a robot making a mistake during an object-sorting task. In a later version of the system, combining brain and muscle signals helped improve the robot’s accuracy from 70% to more than 97%. The potential implication for AI systems is that a user’s reaction could provide a feedback signal without a written or spoken command.
   2. Recognizing relevance: In a Finnish research study, EEG signals recorded while participants read Wikipedia articles helped researchers estimate which words were relevant to their interests and retrieve related documents, without requiring a separate search query. A similar approach could eventually help AI systems identify what a user finds relevant without waiting for them to formulate a prompt.

This suggests that brain signals could give AI another source of feedback, beyond what users consciously choose to express. An agent might detect that a result has caught the user’s attention or that a decision appears wrong, then ask for confirmation or investigate further.

The evidence remains preliminary as these were controlled experiments with specific tasks, not demonstrations of general-purpose application. Neural signals can indicate a response without reliably explaining its meaning. And the fourth quadrant remains a limit: neither prompts nor neural input can directly reveal what a person does not know. Agents must still explore, test assumptions, and surface risks the user has not recognized.

What Must Be True for This to Work

Telepathy-based computing will not succeed simply because neural signals can be decoded. It must distinguish deliberate intent from passing thoughts, prove that it is reading signals rather than generating plausible guesses, and protect information people never intended to share. Three challenges will determine whether it can become a trusted interface for AI agents.

1. 1. Distinguishing intention from thought: Not every neural pattern is a command. A system that acts on passing thoughts would be unpredictable and unsafe. The solution is to require deliberate activation and confirmation before an agent acts. Stanford researchers have explored a mental-password mechanism to gate neural decoding, while research from the University of Texas showed that its non-invasive fMRI decoder worked only with trained, cooperative participants. When participants deliberately thought about something else, the output became unusable. These findings suggest that user control is not just a safety feature; it is fundamental to making neural input work.
   2. Protecting mental privacy: Neural data can reveal more than the command a user intends to send. Depending on the signals collected and the models applied, it may also support inferences about attention, emotional responses, or health. Consent to collect neural data is not necessarily consent to infer mental states, retain those inferences, or use them for another purpose. Regulation is beginning to address these risks. US states, including Colorado and California, have introduced protections for neural data, while UNESCO adopted its first global Recommendation on the Ethics of Neurotechnology in November 2025. The framework calls for safeguards around privacy, human rights, and informed consent. For telepathy-based computing, privacy must be built into the architecture through data minimization, on-device processing where feasible, clear consent for specific uses, and strict limits on secondary use. Users should also be able to pause decoding, revoke access, and delete their data.
   3. Proving the system is decoding, not guessing: Language models can produce convincing text even when the underlying neural signal is weak. A fluent response, therefore, does not prove that a system has accurately decoded a user’s intent. Credible evaluations must separate information recovered from neural activity from information supplied by the model’s context and predictions. This means testing against shuffled or unrelated signals, evaluating performance on users excluded from training, and measuring how often the system misinterprets intent. Until vendors demonstrate this independently, claims of thought decoding should be treated cautiously.

Conclusion

The first gains may come from narrow applications where conventional input falls short, from hands-free interaction to feedback that never makes it into a prompt. But as decoding advances, the interface itself could become a strategic layer in the AI stack. The transition will be gradual. Wearables that interpret muscle signals are already bringing neural-adjacent input into everyday computing, while implanted interfaces are restoring communication for people with paralysis. Conduit’s Naomi Bashkansky expects AI models to consume neural representations directly by 2030 and two-way read-and-write interfaces by 2035. These timelines are a bet, not a forecast, but they show where the capital is pointed. For technology leaders and investors, the question is no longer whether neural signals become an AI input modality. It is who owns the data, the decoders, and the trust layer when they do.

By Chandrika Dutt, Research Director, Avasant