Teaching Machines to Read the Brain: Guan Cuntai on the Future of Brain-Computer Interfaces

Professor Guan is one of the leading international researchers in non-invasive brain-computer interfaces (BCI) and neural AI. Trained in computer science and electrical engineering, his research has helped push BCI beyond the lab toward rehabilitation, assistive technology and a deeper understanding of how the brain recovers after injury. 

In his opening keynote, Professor Guan focused on how AI can make BCIs clinically meaningful — not just technically impressive. He outlined the conditions under which decoding brain activity becomes truly useful in practice: it must be accurate, robust across sessions, fast enough for real-time feedback, easy to use and generalizable across patients. He also showed how neural AI can support gait rehabilitation by identifying brain signals linked to walking, distinguishing healthy neural patterns from altered ones in patients with spinal cord injury, and using those signals to guide personalized therapy, robotic assistance and, potentially, the design of neurostimulation-based interventions. 

- Your work sits at the intersection of artificial intelligence, neural signal processing and healthcare. What first drew you to brain-computer interfaces — and what has kept the field exciting for so long?

- I was trained in computer science and electrical engineering, so I was always interested in data processing and pattern recognition. Early in my career, after my PhD, I worked on speech recognition. Brain-computer interfaces drew me in because they offered a way to apply those skills to something with immediate human value: helping people with neurological and mental disorders.

What has kept the field exciting is that BCI evolves together with AI. Methods that succeed in speech or image analysis often open up new possibilities for brain signals as well. But I also learned quite early that signal processing alone is not enough. If you want to do meaningful work in this area, especially in medicine, you also need to understand physiology and neurophysiology. The brain has structure, connectivity, spatial organization and functional specialization, and all of that shapes the signals we record. That is why I see my work today as a combination of AI expertise and a deep engagement with brain science.

- How has the role of AI in BCI research changed over the past decade? What has been the most important shift in recent years?

- BCI has always tried to adopt the best AI and machine-learning methods available, although the field tends to follow other areas like computer vision or natural language processing. Still, the major shifts in AI have had a strong impact on BCI as well.

One important turning point was the rise of convolutional neural networks. They transformed image recognition, and we found that they were also very powerful for EEG and BCI data. My team was among the early groups to apply them in this area. They improved performance, of course, but they also gave us new ways of thinking about neurophysiology. For us, deep learning was not simply a black box; it also became a tool for insight.

More recently, transformers have become increasingly relevant because brain data are inherently sequential. In BCI, temporal structure matters: what happened before often helps explain what comes next. So if I had to summarize the last decade, I would say the key shift has been toward models that capture neural features more effectively — first spatial, then temporal.

- Your presentation made clear that this research has a very direct practical goal: rehabilitation. How far are we from seeing current clinical trials translate into widely available therapies for patients with spinal cord injuries? And does the future go beyond exoskeletons?

- It depends on the condition and on the function you want to restore. For upper-limb rehabilitation — the shoulder, elbow and other large joints — a clinical trial with several dozen patients may take two to three years to complete. Each patient goes through weeks of intensive training, often several sessions a week.

Once you move to finer motor function, especially the hand, the challenge becomes greater and the timeline longer. The hand is much more complex, so rehabilitation usually requires more time and more precise intervention.

- We have seen experiments with remarkably advanced mechanical hands, capable of very precise movements.

- Yes, but the crucial principle is active training. Patients should not simply have their hands moved for them by a robotic system. They need to participate actively, because that is what drives recovery. One effective approach is to detect motor intention from brain signals and use that to control a robotic glove or another assistive device.

The same logic applies even more strongly to lower-limb rehabilitation, which is often more difficult and more time-consuming. And rehabilitation is only one part of the picture. BCI can also be used in communication and provide interventions for ADHD, mild cognitive impairment, anxiety, and other conditions.

- As AI becomes more capable of decoding neural activity, what new scientific opportunities does that create — not only for therapy, but also for understanding the brain itself?

- That is a very important question. Decoding brain activities for intervention is only one side of the story. The other is understanding the mechanism of recovery.

In our group, as in many others, we use functional MRI and other imaging tools to study brain structure and connectivity. The central idea is neuroplasticity: if rehabilitation is truly effective, the brain should show measurable change — in functional connectivity or even structure. If we can observe that, we know the intervention is doing more than producing a temporary effect.

So BCI is not just a therapeutic tool. It is also a scientific instrument. It helps us understand how recovery happens and why some interventions lead to lasting change while others do not.

- Brain data are among the most sensitive forms of personal data. What ethical and privacy challenges does that create for AI-enabled BCI systems?

- They are significant, and they have to be taken seriously. In medical settings, however, the framework is already well established: ethics approval, informed consent, strict data management, and secure storage. In our work, data are typically kept on secure hospital servers and handled according to medical protocols.

If BCI moves further into non-clinical settings, those questions will become even more important. But the principles remain the same: secure infrastructure, encryption, anonymization, controlled access, and clear governance. Brain data is highly sensitive, but they are not beyond protection.

- Let me raise the familiar science-fiction concern: is there any real basis for the fear that brain-computer interfaces could one day read a person’s private thoughts?

- Based on current neuroscience, no. There is no evidence that today’s BCI systems — invasive or non-invasive — can decode private thoughts in the way science fiction imagines.

What they detect are relatively high-level functional signals: motor activity, speech-related activity, attention, arousal. These are broad categories, not hidden inner content. So with current technology, the idea of mind-reading is not supported by the evidence.

That said, ethical vigilance remains essential. Even if the science-fiction scenario is not real today, these technologies should still be developed within clear ethical boundaries and for human benefit.

- Your work brings together engineering, computer science, neuroscience and medicine. What have you learned about building truly interdisciplinary teams?

- That it is difficult — and indispensable. We recently completed a five-year programme involving five universities, two hospitals and several research labs, and experiences like that make the requirements very clear.

First, you need strong people: collaborators who are professional, reliable and capable of delivering. Leadership matters a great deal at the team-building stage, because trust matters. Ideally, you know how people work and whether they collaborate well.

Second, you need the right culture. An interdisciplinary team only works if there is openness, mutual respect and no internal competition. That includes intellectual contribution. In this kind of work, no one succeeds alone. Without clinicians, for example, the technical work has no clinical value. So respect cannot be symbolic — it has to shape the whole collaboration.

- But for such collaborations to work, doesn’t someone have to speak both languages — to understand the technical and the clinical sides well enough to connect them?

- Yes, and that bridging role matters a great deal. I have seen many collaborations fail because technical researchers approach clinicians by simply asking: “Tell me what matters. Tell me what to build.” That usually does not work, because the two worlds are too different.

In practice, the technical side often has to make the first move. We do not need to become doctors, but we do need to understand the clinical context — the physiology, the constraints, the real pain points. Only then can we come back with concrete proposals and ask a better question: would this kind of solution actually be useful?

That creates a much more productive conversation. The process is not: tell me what to do. It is: understand the domain, identify the need, then propose solutions that the domain experts can properly evaluate.

- That sounds very much like the “AI plus X” model — except that AI has to actively reach out toward the “X”, whatever that domain may be.

- Exactly. AI researchers have to move towards the application domain, whether that is medicine, materials science or anything else. If you only ask for data and a target outcome, your work remains superficial. To contribute meaningfully, you need to understand the logic of the field, engage with its literature and learn enough to work intelligently with the people who are true domain experts.

- Looking ahead, which advances will matter most for the next generation of BCI: better algorithms, better sensors, richer multimodal data, or closer integration with robotics and clinical practice?

- All of them. The field will not progress through a single breakthrough. Sensors are still not good enough, and when sensors are limited, the data are limited as well. Without rich, high-quality data, even the best algorithms can only go so far.

So the process is iterative: better sensors enable better data, better data enable better models, and better models make clinical integration more realistic. Design matters too, because these are human-centred technologies. Comfort, usability and real-world practicality are all critical. In the end, none of this matters unless the outcomes are clinically meaningful and measurable.

- If you look ten years ahead, what would success look like for AI and brain-computer interfaces in healthcare?

- I am quite confident that within five to ten years, some forms of non-invasive BCI will become mainstream clinical tools. Many systems are still in the lab or in clinical trials, but the necessary pieces are coming together: sensors, algorithms, software and clinical evidence.

I expect non-invasive BCI to be used much more widely in rehabilitation, cognitive improvement and mental health. Some related technologies, such as robotics, are already in use, but they still face challenges in demonstrating strong enough efficacy. BCI has already shown promising early results, and I believe that will lead to broader clinical adoption.

-  One last question: can you imagine a future for BCI beyond medicine — for example, helping healthy people control machines or interact with computers using only their brains?

- In the foreseeable future, I think BCI will remain primarily a medical technology. For controlling devices, the human body already has an extraordinary interface: the muscles. Muscular control is fast, flexible and highly multidimensional. (Non-invasive) brain control is generally slower and less precise.

The brain is very good at initiating intention — deciding to act, choosing a direction — but for fine control, muscles are still superior. Even invasive systems, which offer clearer signals, come with major challenges such as cost, risk and long-term stability.

So for the near future, the clearest value of invasive BCI lies where natural motor control or speech has been lost. That is where it can make the greatest difference. For non-invasive BCI, the immediate value will be shown in rehabilitation, cognitive enhancement, and mental health.