Inside the human mind: Can AI read our thoughts?

​AI can already write, code, see and speak. The next frontier may be considerably more intimate: understanding what we are thinking before we say it. Researchers are getting better at turning brain activity into words, while startups are betting that "telepathy" could become usable within years. How close are we really to machines that can read our minds?

From OpenAI to "telepathy"

In July, Naomi Bashkansky left OpenAI to join Conduit, a startup developing thought-to-text models capable of translating brain activity into language without requiring surgery.

"We're building telepathy: thought-to-text models, trained on non-invasive neural data," Bashkansky wrote when announcing the move. She envisions early thought-to-text tools emerging around 2027, with more capable brain-to-AI interfaces developing over the following decade.

View post on X Loading...

The tweet was deleted by the author.

But we saved everything 🙂.

The word "telepathy" sounds like science fiction. But the underlying idea is becoming less far-fetched as researchers learn to translate brain activity into increasingly meaningful information without surgery.

AI is already turning brain activity into words

In 2023, researchers at the University of Texas at Austin used functional magnetic resonance imaging (fMRI) and an AI decoder to reconstruct continuous language from people's brain activity. The system could recover the meaning of speech people heard and, more strikingly, language they imagined.

The results were far from perfect transcripts. The decoder reconstructed semantic meaning rather than extracting every word passing through someone's mind. That distinction matters. Decoding imagined or intended language is not the same as accessing arbitrary thoughts, memories or emotions.

The technology has since moved further. Meta's Brain2Qwerty v2, unveiled in June, converts MEG recordings collected while participants type into sentences. In tests involving nine participants, the system achieved 61% average word accuracy, while its best participant reached 78%.

Brain implants are further ahead. In 2025, researchers used 256 implanted microelectrodes to turn the attempted speech of a man with amyotrophic lateral sclerosis (ALS) into audible speech almost instantaneously.

Why AI changed the problem

Scientists have been measuring brain activity for decades. What is changing is their ability to make sense of it.

Brain signals are extraordinarily noisy, and thoughts do not produce neat sequences corresponding to individual words. Modern deep-learning models can identify patterns in neural data that traditional decoding methods struggled to capture.

Language models add another advantage. They can use context to determine which words and sentences are plausible even when the neural signal is incomplete. A future system may therefore not need to capture every detail of a neural signal perfectly if AI becomes sufficiently good at reconstructing its likely meaning.

This is not mind reading. At least not yetToday's technology still comes with enormous limitations. Meta's system requires an MEG scanner, hardly something consumers can wear while walking around. It also required approximately 10 hours of individual brain recordings from each participant for training.

The University of Texas experiment provides another reality check. Researchers found that decoding required the participant's cooperation. People could deliberately disrupt the system simply by thinking about something else, while a decoder trained on one person performed poorly when applied to another.

Lead researcher Jerry Tang has also cautioned about the potential for misuse. "We take very seriously the concerns that it could be used for bad purposes and have worked to avoid that," he said.

From keyboards to brain-to-AI

The more realistic promise is not unrestricted mind reading. It is removing some of the steps between human intention and a computer.

Today, someone thinks of an instruction, turns the thought into words and types or speaks those words to an AI system. A sufficiently capable neural interface could eventually translate an intended message directly from brain activity.

The immediate applications are medical. People who have lost the ability to speak or move could regain a more natural way to interact with the world. Neuralink has already demonstrated implanted brain-computer interfaces that allow people with paralysis to control computers through neural signals.

Conduit's much harder bet is that useful brain-to-computer communication could eventually work without surgery. If non-invasive systems become smaller, more accurate and require less individual training, the technology could move beyond medicine and become another way of interacting with computers.

What happens to mental privacy?

Brain-to-AI communication also creates a privacy problem. Browsing histories, messages and locations reveal enormous amounts about people, but brain activity sits one step closer to the source.

Neural systems could eventually infer intentions or words that a user never intended to store permanently. Who owns that information? Can companies use it to train AI models? Under what circumstances could someone else access it?

Those questions are already reaching lawmakers. In 2024, Colorado expanded its privacy law to explicitly cover neural data as part of protected biological data, defining it as information generated by measuring activity in the central or peripheral nervous system. Researchers are thinking about the problem too. The University of Texas team wrote that brain-computer interfaces "should respect mental privacy."

Will AI actually read our minds?

The first widely used brain-to-AI systems are unlikely to resemble science-fiction telepathy. They are more likely to translate deliberate attempts to communicate rather than expose everything someone happens to think.

But keyboards, touchscreens and voice assistants all solve the same problem: turning human intention into something a computer can understand. If AI eventually learns to decode that intention directly from the brain, the next revolution in computing may not be about what machines can do. It may be about how little we need to do to tell them.