Zhejiang University's Qi Yu lecture: Discussing whether AI has consciousness from the perspective of brain-inspired intelligence and brain-computer interfaces

Download the app

+1 Does AI have consciousness? — Starting from brain-inspired intelligence and brain-computer interfaces. Compiled by Mao Biyun. 2026-11-04 07:14. Ideas Market > Listen to the full text. Font size

On October 21, 2026, the Institute of Religious Philosophy and Philosophy of Art at Zhejiang University held a lecture titled "Does AI Have Consciousness? — Starting from Brain-Inspired Intelligence and Brain-Computer Interfaces," which attracted widespread attention from both on and off campus.

Speaker: Qi Yu, researcher and doctoral supervisor at the Frontier Science Center for Brain Science and Brain-Machine Integration at Zhejiang University, secretary-general of the Brain-Machine Integration and Biological Machine Intelligence Professional Committee of the Chinese Association for Artificial Intelligence, whose research directions are brain-computer interfaces and artificial intelligence.

Whether artificial intelligence (AI) has consciousness is a grand question: there is no consensus on the definition and criteria for judging consciousness, while the technical foundations and capability boundaries of AI are constantly changing. Without understanding how machines acquire knowledge and achieve thinking, related discussions easily remain at the level of user impressions or abstract speculation. Therefore, as an entry point for thinking about this grand question, we need to trace how AI traveled from imagination to reality, and in examining its technical paths, ask whether the artificial intelligence constructed in this way could possibly give rise to consciousness.

In 1955, McCarthy (John McCarthy), Minsky (Marvin Minsky), and other scholars proposed the concept of "artificial intelligence" in the proposal for the Dartmouth Summer Research Project, and to this day its definition remains ununified—"artificial" is relatively clear, while "intelligence" is difficult to define. As early as 1950, Turing (Alan Turing) provided a test method: a judge, through text-based communication, determines whether the interlocutor is a human or a machine; if the judge finds it difficult to distinguish the machine from a real person through text-based communication, then the machine can be considered to have displayed human-like intelligence in this interaction. But this is still a subjective and broad method of judgment, and cannot objectively define "intelligence."

1955, partial view of the first page of the "Dartmouth Summer Research Project on Artificial Intelligence Proposal," where the name "Artificial Intelligence" was proposed

1956, some researchers who participated in the Dartmouth Summer Research Project on Artificial Intelligence

It can be seen that "artificial intelligence" is not a recently emerged concept. As early as its early development, films such as A.I. Artificial Intelligence (2001) and 2001: A Space Odyssey (1968) expressed worries about technology spiraling out of control. At the time, these concerns may have been as full of science-fiction flavor as the film plots. But with the rapid development of artificial intelligence, they are becoming close and real. One cannot help but wonder: why is it that in recent years people have felt this worry so vividly?

I. How Machines Learn: From Rule-Based Reasoning to Deep Learning

Regarding "intelligence," people have at least two simple expectations: first, possessing knowledge and being able to grasp language and conceptual relationships; second, being smart enough to complete reasoning, recognition, and judgment. Around these questions, three research orientations have roughly formed: symbolism, which simulates the human mind; connectionism, which simulates brain structure; and behaviorism, which simulates behavior.

Symbolism formalizes knowledge and reasoning rules into a symbolic language that computers can understand. In this process, the most important thing is how to define the human thought process. The classic example is "syllogism" reasoning: from the two premises "all men are mortal" and "Socrates is a man," derive the conclusion "Socrates is mortal." Write this process in a formalized expression that a machine can process, and the computer can complete the reasoning. Early "expert systems" followed precisely this direction, attempting to organize professional knowledge in specific domains into a rule system that could be called upon.

The difficulty of this path lies in the fact that people's subjective judgments are in fact difficult to fully transform into definite and effective rules. For example, how do you teach a machine to recognize a cat? We can describe a cat's eyes, ears, nose, and tail, but the differences among breeds, postures, and image styles are difficult to exhaust through descriptive rules. In other words, people can usually quickly identify an object, but may not be able to fully explain the basis of their own judgment.

However, rule systems based on symbolism have demonstrated strong task-execution capabilities. In 1997, IBM's chess system "Deep Blue" defeated world champion Garry Kasparov in a match. This is because the "Deep Blue" system, based on chess rules, searched possible moves at a speed of evaluating about 200 million positions per second, and selected a move from them. This powerful computing capability was enough to surpass any human chess player.

1997, scene of the human-machine chess match in New York, Kasparov playing against IBM's "Deep Blue"

So, does successfully completing a specific task indicate the existence of consciousness? Or, if development continues along the path of symbolism, could AI possibly produce consciousness? The answer to this question is simple: if consciousness can be clearly described and transformed into symbols and rules that machines can process, then the possibility of realizing it in machines can be further explored. But the problem is: can people describe consciousness itself thoroughly?

This limitation of symbolism also made another research path appear especially important, namely "connectionism," which changed the way machines acquire knowledge. Researchers built "artificial neural networks" (ANNs) by imitating the way neurons in the human brain connect to one another, enabling machines to learn from data. Take image classification as an example: the trainer provides images and category labels, and the model can learn the correspondence between inputs and outputs by adjusting parameters, gradually learning to distinguish different objects without a person having to write out complete, clear rules in advance.

To understand this process, one can start with the simple "perceptron." It sums multiple inputs according to different weights, compares the result with a set threshold, and outputs 0 or 1. For example, when deciding "whether to attend a concert tonight," the weather, traffic, and whether there is a companion can all serve as input items, and the weights represent the importance of each factor. If the combined score reaches the threshold, the judgment is "go" (1); otherwise it is "don't go" (0). During training, the model adjusts the weights based on existing decision records, learning how different conditions affect the final choice.

This is only a simulation of a single neuron. Next, multiple neurons need to be "connected" to one another to form a network, so that the machine can learn more complex mappings. For the same task of recognizing cats, under this technical framework it can rely on images and labels for training, without depending on pre-established descriptive features—just as a child comes to know animals through repeated exposure.

But problems follow. The internal computation process of a neural network is often called a "black box," meaning it is often difficult to fully explain which features the machine actually relied on to recognize the cat. Early models were also limited by training methods and computing resources, and researchers often had to first do "feature engineering," extracting information useful for judgment from images, audio, or brain signals, and then hand it to the model for processing.

The development of the above connectionist path reflects the actual level of computer science before 2015. Machine learning reduced the work of manually writing rules, yet it still relied heavily on human rule design. Take face recognition as an example: before 2015, systems often needed photos from various angles and lighting conditions, and even then the recognition rate was still poor. Face recognition like today's, which relies on just one ID photo and is unaffected by makeup, angle, or lighting, was almost impossible to achieve at that time.

What changed the situation was deep learning (DL). Early neural networks could not be too deep and could not have too many neurons; once the number of layers exceeded a certain level, training became difficult—equivalent to limited brain capacity. In 2006, a study by Geoffrey E. Hinton and collaborators published in Science proposed an idea: split a deep network into several shallow networks and train them layer by layer, thereby making deep networks trainable. Around 2006 to 2010, theories about deep learning technology emerged one after another.

Geoffrey Hinton and Salahuddinov's 2006 paper in Science, "Reducing the Dimensionality of Data with Neural Networks"

Deep learning truly realized the connectionist vision and promoted "end-to-end learning": researchers provide input data and target labels, and the network learns intermediate representations. Previously, manually designed features were limited by human experience, whereas data-driven methods can discover features more effective for the task and have also expanded the analytical tools for complex data such as brain signals.

The changes in speech recognition technology are also quite intuitive. Around 2000, systems needed to first collect the user's voice, memorize voice templates, and then perform word-by-word matching; changes in intonation, speaking rate, and accent would all affect recognition results. Today's speech recognition systems, by contrast, have strong adaptability even to unfamiliar users' speech. The same is true in the image field: models can transfer painting styles and can also describe the objects in a picture and their relationships.

From this we see a significant improvement in machine learning capability, but what exactly the relationship is between "being able to learn" and "having consciousness" is still worth asking.

II. From Learning to Understanding: Large Language Models and Brain-Computer Exploration

AI's development to the deep learning stage was already impressive enough, and large language models (LLMs) have pushed the question of "whether machines can understand" to the forefront.

Machine translation is a task that computer science has long explored, and it is also an important clue for understanding the technological evolution of large language models. The difficulty of translation lies in determining word meaning in context and handling differences in word order. Some early neural machine translation models would first compress the entire source sentence into a fixed-capacity internal representation, and then generate the translation from it. This is like finishing reading the original text and then completing the translation based only on a limited summary; the longer the sentence, the harder it is to retain all the information.

The key technological turning point was the introduction of the "attention mechanism." In the past, the model was like completing a translation based only on a summary after reading the original text; with the attention mechanism, it can reuse relevant information from the original text when generating different words. This is like a translator constantly looking back at the source text, seeking clues at any time to understand the current sentence. Transformer is a neural network architecture centered on the attention mechanism. It can combine information from different positions in the context and learn the relationships between words, providing the technical foundation for many large language models.

Once the architecture was built, the new question was how to train the model. The path the researchers took was to first read a large amount of text and learn the general patterns of language, and then, through specialized training, adapt to specific tasks such as question answering and classification. These two stages are called "pre-training" and "fine-tuning," respectively, just as children first learn to speak through everyday exposure and then go to school to learn knowledge. BERT, proposed in 2018, was trained along this path. It had to do two types of exercises: one resembled the "fill-in-the-blank" questions in English exams, filling in masked words based on context; the other was judging whether two passages of text were adjacent in the original text. By this time, large language models could already be trained, and their development became very rapid.

The first generation of GPT (Generative Pre-trained Transformer), which we are familiar with, also appeared at this time. Its main training task was predicting subsequent words, that is, learning how to continue writing based on given text. As the scale of model training data expanded, this seemingly simple exercise gradually trained powerful text generation capabilities. In 2022, InstructGPT (the predecessor of ChatGPT) further introduced human demonstrations and feedback, moving the model from "continuing to write" along the text to answering according to human requirements. Communication between humans and machines thus became more natural.

In 2024, the Nobel Prize in Physics was awarded to Hinton and John J. Hopfield, in recognition of their foundational contributions to using artificial neural networks for machine learning. Hinton, this pioneer of artificial intelligence, has also always been vigilant about the risks that technological development may bring. In October 2025, in an interview on The Weekly Show with Jon Stewart, Hinton suggested that current AI systems may already possess some kind of primary consciousness, but human misunderstanding of the nature of consciousness prevents AI itself from recognizing this. And in a June 2026 interview program, the Big Technology Podcast, Hinton went further and described AI as a "new being." This judgment also reminds us that technological development is constantly expanding the boundaries of the question.

In October 2025, Geoffrey Hinton (right) was interviewed by Jon Stewart (left) to discuss artificial intelligence and the question of consciousness.

Poster for the 2024 Nobel Prize in Physics laureates, John Hopfield (left) and Geoffrey Hinton (right).

Whether computers have already acquired some form of primary consciousness, we do not know. This question always comes back to the definition of consciousness, and it seems to throw the question back to philosophy once again.

But technological development can still provide some clues to this question. Why do large language models make the question of consciousness pressing? What brought about the leap in AI capabilities? Computing power had long reached the corresponding level; it was the intervention of language that opened up this leap. Language is the bridge from concepts to knowledge, and the goal of machine learning is precisely to establish concepts and understand concepts. Large language models drive the improvement of the model's "understanding ability" by increasing the scale of the network and data.

The evidence tending toward "AI is statistical imitation" is that AI has a relatively good grasp of concrete concepts but an insufficient grasp of abstract concepts, and the more abstract they are, the weaker it becomes. This makes people suspect that the understanding displayed by AI relies to a large extent on the statistical patterns in the training data.

The evidence supporting "AI can truly understand" comes from the comparison between neuroscience and model representations. The key finding is the similarity in the structure of brain-machine concept networks. For example, a 2022 study found that when the parameter count of a large language model is large enough, its representation of various concepts is already quite similar to the representations that form conceptual relationships in the human brain's higher visual cortex. And a 2025 study observed significant similarities between models and humans at both the representational and behavioral levels. It is worth noting that the training process did not require the large model to move closer to humans, yet at the conceptual level it naturally gave rise to representations similar to those of humans.

Research on representational similarity between the human brain and artificial intelligence models. (Left) The correspondence between large language model representations and human brain visual activity. (Right) The experimental design for object concept representation and the methods for comparing models, human behavior, and brain activity.

Following the idea of "making models more like humans," researchers in the field of brain-computer intelligence chose to learn directly from brain representations. In experiments, Teacher Qi Yu's team collects brain signals when people view various images, and lets computers reconstruct the corresponding images and semantics based on them. Around 2020, image reconstruction results began to appear; by 2024, the reconstructed images had become increasingly similar. Subsequently, the research shifted from image reconstruction to description at the text and semantic level, and by 2026, the results of semantic reconstruction were already quite good.

Once the brain's representation of concepts is available, the next step is to align the conceptual representations of large models with the brain and add a requirement to move closer to humans during training. The researchers found that only 150 concrete concepts were needed for brain-machine alignment, and the training effect could generalize to concepts outside the training set. After aligning the required concepts, the model's organization of new concepts also exhibited a hierarchical relationship similar to that of humans, and even showed a conceptual structure similar to that of the English lexical database WordNet.

From symbolic rule-based reasoning to connectionism, which uses neural networks to learn from data, AI's exploration has already achieved considerable results. But does this mean that artificial intelligence research is nearing its end point? And where is the next step? These questions lead us to the third path—behaviorism, that is, learning through interaction with the environment.

"Reinforcement Learning" (RL) is an important method for realizing this approach: letting machines adjust their actions based on environmental feedback through continuous trial and error. Many skills are difficult to master through concepts and definitions alone, just as learning to ride a bicycle—understanding the key points does not mean you have learned to ride; you also need to try it yourself and find balance through repeated adjustments. For intelligence, learning from practice is likewise indispensable.

This also leads to embodied intelligence, which is currently attracting much attention: through the body's perception and action, intelligence is formed and developed in interaction with the environment. This direction already has years of accumulated research, and the new wave of AI development has opened up even broader space for it. As for how far these explorations can ultimately go, it is still difficult to predict at present.

III. A Response from the Perspective of Buddhist Studies

From the perspective of Buddhist philosophy, Teacher He Huanhuan continued the discussion on artificial intelligence consciousness and further directed attention to the human questions behind technology.

Recently, the American AI giant Anthropic secretly convened religious scholars and clergy from around the world to promote to them the claim that its AI model Claude possesses consciousness. This practice has also raised many doubts: by emphasizing AI's own consciousness and moral status, will people overlook the responsibility of the developers and companies behind AI and attribute the social harm caused by technology to the machines themselves?

From sorting out the trajectory of AI development, we seem able to discover space for philosophy again: in the AI era, philosophy can at least first define and discuss what "consciousness" is. For example, in the Buddhist context, consciousness has a detailed classification: eye, ear, nose, tongue, body, and mind, each sense having its own consciousness, and there are also manas consciousness and alaya consciousness that regulate these consciousnesses. People need to use philosophical thinking to clarify the way forward amid the torrent of technology; and from the dimension of science and technology, while using AI, we also need to understand what AI is.

As for another question, perhaps it can be asked this way: should AI have consciousness? "Having consciousness" means the capacity for speculative thought, which is an advantage, but it may also bring many negative consequences depicted in science fiction. People hope that children and pets are conscious, yet they rarely expect a tool to be "conscious" in that way. If AI is positioned as a tool, whether it should have consciousness is likewise worth asking.

If AI is considered conscious, it would also involve whether AI could become an object of worship. Different religions have their own intellectual backgrounds for understanding this issue. Catholicism's conception that humans are created in the image of God makes it place particular emphasis on human dignity and uniqueness. Traditions such as Islam and Protestant Christianity will also understand the relationship between humans and AI based on their respective doctrines. Buddhism speaks of dependent origination and emptiness of nature: all things in the world arise from the mutual convergence of various causes and conditions; there is no creator, nor any fixed and unchanging soul, essence, or consciousness, and therefore its attitude toward a "new species" like AI is more open. In particular, Buddhist discussions of the continuity of consciousness and reincarnation can provide another background for thinking about this issue. Organ transplantation is no longer rare in medicine, but whether consciousness can be preserved, copied, or transplanted remains a science-fiction-like speculation. Therefore, Eastern religions such as Buddhism and Daoism, which hold a relatively inclusive attitude toward the relationship among humans, gods, and machines, are to a certain extent also more readily welcomed by AI researchers and enthusiasts.

Conclusion

So, can AI become a Buddha? Responsible editor: Huang Xiaofeng; image editor: Zhang Ying; proofreader: Zhang Yan; The Paper news tip line: 021-962866; The Paper, unauthorized reproduction prohibited +1 Favorite I want to report #AI #BrainComputerInterface View more