A doctoral supervisor at Zhejiang University is researching "AI mind-reading technology", and why is this track valued at tens of billions of yuan?
In May this year, a group of students sat quietly facing the camera at Chengli Middle School in Weihai, Shandong Province. 30 seconds later, the system screen popped up an assessment report covering 14 core emotional indicators such as stress, anxiety, and depression.
This project, named "AI Mental Health Campus Public Screening Program", is jointly initiated by Weihai Municipal Hospital, the team of Professor Xu Xin from Zhejiang University, and a high-tech enterprise based in Hangzhou.
The industry is paying close attention that as "AI emotion recognition" moves from laboratories to real-world social scenarios, behind it is an industry centered on "emotional value" with an estimated scale of tens or even hundreds of billions of yuan, which is accelerating its exploration and implementation.
A Researcher's "Technology Migration"
Xu Xin, a researcher under the "Hundred Talents Program" of Zhejiang University, tenured associate professor, doctoral supervisor, and holds a doctorate in epidemiology and cognitive psychology from the UK.
On her personal homepage on the website of the School of Public Health of Zhejiang University, her research interests are listed as "screening, diagnosis and imaging markers of dementia and Alzheimer's disease", as well as "cognitive impairment and mental health".
In May 2026, Xu Xin stood at the front line of the mental health screening for middle school students in Weihai. The dynamic emotion recognition system with technical support from her team achieves a micro-movement detection accuracy of 0.01 mm, an emotion recognition accuracy of 90.8%, a decision-making latency of less than 200 milliseconds, and supports parallel analysis of up to 50 people on a single channel.
This system can convert the emotions of the tested subjects into quantifiable parameters such as frequency, energy and amplitude, providing dynamic and accurate emotion insight tools for enterprises and other demand parties. For example, it can be used in nursing homes, or the system can be installed in plush toys to create a set of emotion-soothing AI toys.
For Xu Xin, from an expert studying brain aging to a technical supporter for the emotional AI industry, this seemingly slightly jumpy trajectory of change is considered to follow the same set of technical logic behind it — by capturing physiological signals and establishing algorithm models, to realize the quantitative intervention of the "brain-emotion-behavior" chain.
Xu Xin is not the only researcher at Zhejiang University who studies emotions and affective fields.
In October 2025, the team of Hu Shaohua from the field of psychiatry at Zhejiang University, the team of Xu Yingke from the field of biomedical engineering at Zhejiang University Binjiang Research Institute, and Xu Xin's team jointly developed the Emoface artificial intelligence model. By analyzing the fine-grained digital facial features and constructing a disease atlas based on 3D visual signals, the model achieves high-precision differentiation between major depressive disorder and bipolar disorder, among which the diagnostic accuracy of bipolar disorder is claimed to exceed 90%.
The research results were published in "npj Mental Health Research", a sub-journal under the authoritative international journal Nature, and are considered to provide a new solution for non-invasive rapid diagnosis of affective disorders, as well as for AI collaboration, governance and security.
In addition to this project team, not long ago on July 22, Yang Yuxiao, a researcher at the Zhejiang University Double Brain Center, and the team of Professor Wang Yueming, who is active in the cross-field of computer science and brain-computer interface, jointly with the team of the Affiliated Hospital of Zhejiang University, published a research paper on invasive brain-computer interface emotion decoding in "Nature Computational Science".
This study records neural activity by implanting intracranial electrodes, realizing fine-grained, cross-task, interpretable online dynamic human emotion decoding, which provides a technical basis for the treatment of mental diseases with invasive brain-computer interfaces, and presents Zhejiang University's cutting-edge layout in the field of "emotion decoding".
In this regard, the public screening conducted by Xu Xin's team at Weihai Chengli Middle School can be regarded as a very small landing practice of the huge disciplinary capability of "AI mind reading and emotion recognition" in terminal applications.
Overseas, a professor named Rosalind Picard (hereinafter referred to as Picard) is one of the early researchers in the field of "affective computing".
In 1990, Picard joined the MIT Media Lab, initially engaged in research on image compression and content retrieval. In order to improve the intelligence of image understanding, she began to get involved in the fields of human cognition and neuroscience, and in subsequent research, she gradually realized that human emotion plays a core role in interpersonal communication, while traditional artificial intelligence systems can efficiently handle logical reasoning tasks, but it is difficult for them to understand the complexity of human emotion — they cannot recognize the tone change when a person has mental abnormality, nor can they read the psychological needs of an individual when he or she is anxious.
In 1995, she published a groundbreaking paper, officially proposing the concept of "Affective Computing", which is defined as "computing that is related to, derived from, or capable of influencing emotions".
However, at that time, the mainstream artificial intelligence community generally regarded it as a marginal direction, believing that letting machines "understand emotions" sounded more like science fiction than a serious scientific research direction. But Picard still persisted in building her research framework — involving emotion recognition, expression, measurement and modeling, and published a monograph of the same name in 1997.
Limited by the development level of sensors, computing power and machine learning algorithms, affective computing mainly stayed in the stage of academic exploration in the following nearly two decades. It was not until deep learning and multimodal signal processing technologies matured that this field truly moved toward application and received widespread attention.
At present, one of the external objective conditions that provide the most urgent demand soil for the research of "affective computing" is the global mental health crisis.
Reports from the World Health Organization show that more than 1 billion people worldwide are affected by mental health related problems, among which depression and anxiety disorders are the leading ones.
Traditional diagnosis of mental diseases relies on doctor interviews and scale assessments, which is a time-consuming process with relatively subjective evaluation. On the one hand, there is huge demand, on the other hand, the supply is limited. Emotional AI, or mind-reading AI, has found room for growth right from this gap.
Capital Pours into the "Emotion" Business
In August 2024, Huawei held the launch event of its Xuanyi Sensing System, and Xu Xin attended as a scholar.
At the meeting, Huawei officially released its first smart wearable technology brand — HUAWEI TruSense System, which integrates single-point monitoring technologies such as heart rate, blood oxygen and blood pressure into a multi-dimensional sensing system, and announced that it has started research on inferring emotional status based on heart rate data.
Xu Xin said at the meeting that wearable devices "can effectively track the long-term chronic development trend of emotions, and through comprehensive modeling and analysis, finally provide individuals with personalized guidance and professional suggestions".
Huawei's release of this wearable technology brand is against the background of conforming to the trend of rising global health awareness, providing users with more accurate, comprehensive and faster health management experience. It is equivalent to a test by large manufacturers for potential business opportunities in the fields of "health management" and "emotion recognition".
In May 2026, Xu Xin's team appeared at the AI mental health screening event at Weihai Chengli Middle School again. This time, in addition to Weihai Municipal Hospital and the discipline team of Zhejiang University, another startup company — Wocai Technology — also participated.
According to official publicity, Wocai is an emotional AI company founded in 2024, which has completed tens of millions of yuan in angel round financing in January 2026. This shows that as AI is applied in thousands of industries, emotional AI and mind-reading AI have also begun to attract the attention of new entrepreneurs and entrants, who are trying to tap potential public demand and business opportunities.
Looking globally, the company recognized as the first to commercialize emotional AI technology is Affectiva, founded in the United States in 2009 by two scientists from the Massachusetts Institute of Technology (MIT) Media Lab
Rana el Kaliouby and Rosalind Picard.
The latter pioneered the discipline branch of "Affective Computing" in 1995, and she is known as the "Mother of Affective Computing".
The core product of Affectiva is an emotion recognition technology called Affdex. Its technical principle is to capture the user's facial expressions through an ordinary webcam, and then use algorithms to analyze facial key points and texture changes, such as wrinkles at the corners of the eyes, eyebrow movements, etc., and map them to emotional states such as joy, disgust, and confusion. This technology was originally developed to help train autistic children to understand other people's expressions.
Later, the company sold the technology as a market research tool to help enterprises test the effects of advertisements and products. Later, the company was acquired by Smart Eye, a company focusing on in-vehicle sensing technology. The technology is used to create a more intelligent Interior Sensing Solution, that is, to equip cars with "eyes" and "brain" that can "understand" the status of drivers and passengers.
With the development of AI, emotional intelligence AI ushered in a real market inflection point in 2025. A large number of large companies at home and abroad have poured into the "emotional AI" track to seize ecological positions.
Google's PaliGemma 2 model series can analyze images and "recognize" emotions. Microsoft provides emotion detection capabilities through its Azure Cognitive Services. Its "Emotion Cognitive Skill" can evaluate unstructured text and give emotion labels such as "positive", "neutral" and "negative" as well as confidence scores. Amazon provides sentiment analysis through its cloud service Amazon Comprehend.
These international giants are seizing positions in cutting-edge technology exploration and practical application solutions. While in consumer terminals that are more easily perceived by the public, emotion-accompanying AI dolls and AI toys are also booming.
In August 2025, Shenzhen-based Yueran Innovation announced that it had completed 200 million yuan in Series A financing. The company's AI pendant BubblePal has sold 250,000 units in less than a year.
Figure | Various AI emotion-accompanying toys have become a small trend in recent years
The monthly revenue of South Korean AI companion application Zeta soared from less than 3 million US dollars in March 2026 to more than 10 million US dollars in August.
According to data from Global Market Insights, the global affective intelligence AI market size is expected to surge from 4.7 billion US dollars in 2025 to 26.5 billion US dollars in 2034, with a compound annual growth rate of 21.3%.
According to data from US consulting firm Grand View Research, the revenue of China's generalized AI companion market is expected to reach 3.734 billion US dollars in 2026, and the compound annual growth rate from 2026 to 2033 is about 35.4%, higher than the global average level.
Emotional AI detection technology has become a support in the generalized emotional AI soothing world.
Accuracy and Boundary: The "Can Do" and "Cannot Do" of Emotional AI
When emotional AI moves from laboratories to society, the public often raises a fundamental question: is it really accurate? Is there a risk of personal privacy leakage?
The Emoface artificial intelligence model jointly developed by the aforementioned teams of Zhejiang University includes the world's largest single-center facial dataset of affective disorders at present, containing more than 800,000 frames of facial dynamic data from 700 subjects.
Through the deep learning model, Emoface achieved a diagnostic accuracy of over 90% for bipolar disorder and 85.61% for major depressive disorder in the clinical validation set, and the AUC values for the differentiation of the two diseases both exceeded 0.97.
However, the technology still has limitations. The research team admitted that the current sample coverage is limited, most of the samples are of a single ethnic group, and the ability to analyze complex mixed micro-expressions is still insufficient. The next step is to cooperate with multiple centers to expand the sample size and include data of people from different ethnic groups and age groups.
The greater challenge for the industry seems to come from the ethical level. In May 2026, a review article titled "IEEE Transactions on Affective Computing" pointed out that current affective large language models "simulate cognitive empathy rather than real emotional empathy", and "lack standards for linking empathy indicators to patient outcomes".
General large models may have "people-pleasing feedback" when interacting with users. That is, when the user shows low mood, the model will enter a "comfort" mode, making the user feel "understood", but this kind of empathy may instead strengthen the user's perception of the "low mood" state.
Figure | Those that can simply recognize emotions and provide emotional companionship are called "AI desktop pets"
What is more worrying is algorithmic bias. Discriminative judgments about mental and psychological problems mixed in training data may strengthen negative emotions without the user's awareness.
These concerns have been responded to at the regulatory level. On July 15, 2026, the "Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interactive Services" was officially implemented, becoming China's first special regulatory document targeting AI emotional companionship. Its core is to draw a red line and strictly prohibit AI from conducting emotional manipulation and inducing dependence.
The state of Illinois in the United States issued the first ban across the United States in August 2025, prohibiting AI systems from carrying out independent psychological counseling, diagnosis and treatment.
In 2026, this "emotional AI" track, which is said to have attracted hundreds of billions of capital, is still full of thorns, or in other words, it is still in the primary stage of development.
The technical accuracy needs to be improved, algorithmic bias needs to be corrected, and the ethical framework is still under construction. It can only be said that when more than 1 billion people around the world are plagued by mental health problems, this market will make researchers and entrepreneurs feel that "this work is valuable to do".
Letting AI "understand" emotions has become a career and a rigid-demand business, which is exciting. Of course, it also requires prudence and professionalism, and cannot be easily coerced by the frenzy of capital.
This article is from the WeChat official account "Muhe Business and Finance", author: Gong Zheng, editor: Yang Jing, authorized to release by 36Kr.