Artificial intelligence development and subjective well-being of older adults: Evidence from the China longitudinal aging social survey

Abstract

Objective

With the intensification of population aging, the subjective well-being of older adults has become a topic of social concern. Meanwhile, artificial intelligence (AI) is penetrating various fields of society and transforming the living environment of older adults. However, research on the association between AI and the subjective well-being of older adults remains relatively scarce, and this study aims to explore the relationship between AI development and their subjective well-being.

Methods

Based on the data from the China Longitudinal Aging Social Survey (CLASS), this study uses a two-way fixed effects model to examine the association between AI development and the subjective well-being of older adults. For robustness checks, it employs methods including double machine learning (DML), instrumental variable approach, variable replacement, addition of control variables, and adjustment of clustering levels.

Results

(1) AI development is positively correlated with the subjective well-being of older adults. (2) Older adults’ use of elderly care services and their social participation are the channels through which AI development is associated with improved subjective well-being among older adults. (3) The digital literacy of older adults plays a positive moderating role. (4) The positive association between AI development and subjective well-being is significantly larger in magnitude for the old-old group than for the young-old group.

Conclusion

This study provides robust empirical evidence for a positive association between AI development and the subjective well-being of older adults. Based on this, efforts should be made to promote the in-depth integration of AI with elderly care services, expand AI-enabled channels for social participation, launch initiatives to improve older adults’ digital literacy, and refine policy safeguards for AI-assisted elderly support.

Keywords

artificial intelligence subjective well-being older adults elderly care services social participation digital literacy

1. Introduction

Population aging is a major social transformation challenge faced by many countries worldwide, and this is especially true in China. Affected by the family planning policy implemented since 1971, the degree of population aging in China has continued to deepen in recent years, making it the country with the largest elderly population globally. Data from the National Health Commission shows that by the end of 2024, the population aged 60 and above in China reached 310.31 million, accounting for 22.0% of the total population. Population aging has become a long-term and hard-to-reverse social issue in China. Against this backdrop, the subjective well-being of older adults has gradually become a core concern of society. Subjective well-being, as a subjective evaluation of life quality based on one’s own feelings, expectations, and values,¹ is directly related to older adults’ sense of happiness in their later years. Enhancing the subjective well-being of older adults is crucial for effectively addressing aging challenges and promoting the process of active aging.

Another major social trend parallel to population aging is digitalization, especially the breakthrough progress of a new generation of artificial intelligence (AI) and its deep integration with the economy and society. The Chinese government attaches great importance to the innovative development and promotion of applications in the AI field. In 2017, the State Council issued the New Generation Artificial Intelligence Development Plan, elevating AI to the national strategic level for systematic layout. In 2025, the State Council issued the Opinions on Further Implementing the “AI+” Action, proposing to promote the extensive and in-depth integration of AI with various industries and fields of the economy and society. With national strategic leadership and policy support, China’s AI field has developed vigorously. According to the report of the World Intellectual Property Organization, by the end of June 2025, China had become the country with the most AI patents globally, accounting for 60%. AI is constantly permeating all areas of social life,² changing the living environment of older adults and is bound to impact their subjective well-being. Then, is this association positive or negative? Clarifying the relationship between AI development and older adults’ subjective well-being is of great research value and practical significance for China, which is in the stage of rapid population aging and rapid social digitalization.

Subjective well-being, as an important indicator representing an individual’s overall cognitive and emotional judgment of their life state,³ has always been the focus of academic attention. Scholars usually use indicators such as life satisfaction and happiness to measure it.^4,5 Compared with traditional economic welfare indicators like income and consumption, subjective well-being not only covers objective material living conditions but also reflects older adults’ subjective experiences, thus having a more comprehensive and profound connotation. Existing studies have extensively explored the influencing factors of older adults’ subjective well-being, covering multiple dimensions such as older adults’ individual characteristics, family support, and social support.^6–8 In the context of the digital economy era, people’s daily life patterns have undergone subversive changes, and the impact of digitalization on older adults’ subjective well-being has received extensive academic attention. According to research perspectives, existing literature can be divided into two categories: the individual behavior perspective and the living environment perspective. Studies from the individual behavior perspective are conducted based on older adults’ use of a certain digital technology. Some scholars have investigated the impact of older adults’ use of mobile Internet, mobile payment, APPs, etc. On their subjective well-being.^9–11 Studies from the living environment perspective examine the impact of the digital environment where older adults live on their subjective well-being, such as digital rural construction, rural digital development, and digital inclusive finance development.^12–14 Overall, different digital technologies have significantly different impacts on older adults’ subjective well-being, including positive impact, negative impact, and no significant impact. This complexity means that targeted empirical investigations on the well-being effects of AI development are necessary, rather than simply and roughly classifying it into the category of digital technologies.

Compared with traditional technological progress, AI has stronger universality and permeability in various industrial fields.¹⁵ Currently, the development of AI has entered a new stage of deep integration with the economy and society, and its penetration and reshaping of various social fields have attracted scholars’ attention. However, current empirical research on the relationship between AI development and older adults’ subjective well-being remains relatively scarce. Existing literature mainly focuses on the practical effects of AI in the field of intelligent elderly care products. By using intelligent elderly care products, older adults can monitor their physical conditions and lifestyles in real-time, enhance their sense of security and control over their health conditions, and alleviate their inner anxiety.^16,17 The social interaction and emotional companionship functions based on natural language processing technology enable older adults to obtain emotional comfort and establish positive concepts, thereby alleviating their feelings of loneliness.^18,19 However, most existing studies focus on a single intelligent product. Although they have confirmed the positive effects of AI in local scenarios, it is difficult to systematically reveal the comprehensive impact of the overall development of AI technology on older adults’ subjective well-being. In addition, existing studies are mainly carried out from the individual behavior perspective, investigating the impact of older individuals’ active use of AI products on their subjective well-being. In fact, as a general-purpose technology, AI has penetrated into many aspects of older adults’ lives. Even if older adults do not actively use AI products, they will inevitably be passively affected by the changes in the living environment brought about by the integration of AI. Ecological systems theory points out that the living environment has an important impact on people’s physical and mental conditions, and the state presented by an individual is the result of the interaction between the individual and the environment.²⁰ However, there are few studies conducted from the perspective of older adults’ living environment at present. Then, what kind of impact will this external AI development have on older adults’ subjective well-being? This remains to be further empirical analysis.

This paper conducts an empirical analysis based on the micro-data from the Chinese Longitudinal Aging Social Survey (CLASS) and city-level artificial intelligence patent data, aiming to examine the association between macro-level AI development and the subjective well-being of micro-level older individuals. Compared with existing literature, this paper makes three main contributions. First, it expands the research perspective. Previous studies mostly focused on the individual behavior perspective, concentrating on older adults’ use of a single intelligent product. While this can reflect the practical role of AI in certain specific application scenarios, it is difficult to capture the comprehensive association of overall AI development with older adults’ subjective well-being. This paper starts from the perspective of older adults’ living environment, regards AI as a collective concept, and examines the association between external macro-level AI development and the subjective well-being of older individuals, thus making up for the deficiency in relevant research perspectives. Second, starting from the perspective of older adults’ living environment, this study deeply analyzes the association of AI development with older adults’ subjective well-being based on ecological systems theory. It further examines the potential mediating roles of elderly care service utilization and social participation, as well as the moderating role of older adults’ digital literacy, which enriches the theoretical interpretation of how AI may relate to older adults’ subjective well-being. Third, it provides more rigorous and reliable empirical evidence for understanding the relationship between AI and older adults’ subjective well-being. In previous empirical studies, older adults’ use of AI products and their subjective well-being may be interdependent, which could introduce severe endogeneity concerns and affect the reliability of empirical results. This study uses the macro-level AI development level as an explanatory variable for regression analysis, which is less interfered by the endogeneity problem and can obtain more convincing research conclusions.

2. Theoretical analysis and research hypotheses

2.1. The direct effect of AI development on the subjective well-being of older adults

According to ecological systems theory, the living environment has a significant impact on people’s physical and mental conditions, and the state an individual presents is the result of interaction with the environment.²⁰ Compared with other age groups, older adults are more likely to face risks such as mobility impairments and cognitive decline. These characteristics make the elderly group more sensitive to environmental factors, and the connection between their physical and mental conditions and the living environment is closer.^21,22 As a major general-purpose technological innovation, AI penetrates into all areas of social life, which will undoubtedly change the living environment of older adults. By reviewing relevant literature, we can see that the application of AI in many fields brings many conveniences and improvement factors to the lives of older adults. For example, in the transportation field, AI-based transportation systems significantly reduce travel risks such as accidents, falls, and injuries for older adults during travel,²³ and improve the inclusiveness and safety of their travel.²⁴ In the medical and health field, hospitals can use intelligent medical systems for remote diagnosis, treatment, and physical examination to provide effective medical services for older adults.²⁵ In the environmental protection field, intelligent environmental monitoring systems can manage the environment efficiently, helping to create a green and healthy natural environment.²⁶ Therefore, with the development of AI, a series of intelligent application scenarios jointly form a more supportive living environment for older adults. Older adults living in such an environment can feel the benefits brought by AI development, and their subjective well-being is improved.

In addition, it is worth noting that some literature points out that social digitalization may pose inconveniences to digitally vulnerable groups represented by older adults, thereby exerting a negative impact on their well-being.²⁷ However, for China, where the digital transformation is still in its infancy, the marginal benefits brought by the application of digital technology are higher than the marginal costs, and the overall impact on residents’ well-being is promotional.²⁸ Furthermore, in terms of technical characteristics, AI technologies such as natural language processing, speech recognition, and image recognition can effectively lower the digital threshold. For instance, for older adults who are unfamiliar with text input or complex device operations, voice interaction functions allow them to use smart devices and obtain information through speech, which facilitates their daily lives. Therefore, even digitally vulnerable older adults can benefit from the development of AI without being left behind in the digital age. Based on the above discussion, the first hypothesis proposed in this paper is as follows:

Hypothesis 1

AI development is positively correlated with the subjective well-being of older adults.

2.2. The indirect effect of AI development on the subjective well-being of older adults

2.2.1. Mechanism of older Adults’ use of elderly care services

AI development can enhance older adults’ subjective well-being by promoting the use of elderly care services. With the shrinking of family size and the separation of intergenerational living, family resources for elderly care have become relatively scarce, making the traditional elderly care model that relies solely on children for support unsustainable.²⁹ According to social support theory, social elderly care can be an effective supplement to family care, meeting the elderly care needs of older adults and thereby enhancing their subjective well-being.³⁰ However, traditional elderly care services are constrained by human resources, resulting in insufficient effective supply and demand.³¹ The development of AI has transformed the logic of elderly care services from being human-resource-driven to technology-driven, promoting the use of elderly care services by older adults from both the supply and demand sides, and thus improving their subjective well-being.

On the supply side, AI development enhances the supply capacity of elderly care services. Elderly care services under the traditional “person-to-person” model are a typical labor-intensive industry. However, care occupations generally have problems such as high labor intensity, harsh working environments, poor salary and benefits, and low social recognition. Combined with the disappearance of the demographic dividend and the rise in human capital prices, this leads to insufficient supply of elderly care services. AI development promotes the transformation of elderly care services from the “person-to-person” model to the “AI-person” model, alleviating the shortage of care personnel and improving the supply capacity of elderly care services. Specifically, first, AI reduces the work intensity of care personnel, alleviating their resistance to care work, thereby attracting more labor to the elderly care industry. Second, AI replaces part of the labor, reducing the demand for nursing personnel in the elderly care industry. The reduction in human capital investment lowers the price of elderly care services, making them affordable for more ordinary older adults and enhancing the effectiveness of elderly care service supply. On the demand side, AI can increase older adults’ willingness to use elderly care services. First, AI lowers the threshold for using elderly care services. Traditional elderly care services rely on methods such as manual registration and telephone appointments, which have cumbersome processes and restrict older adults’ use of elderly care services. In contrast, optimizations such as AI-based voice interaction and scenario-based guidance match the cognitive and action characteristics of older adults, improving perceived ease of use and enhancing their autonomy in using elderly care services. Second, AI promotes the matching of supply and demand in elderly care services. Older adults have highly personalized elderly care needs, while traditional elderly care services mostly adopt a standardized supply model, which is difficult to meet their personalized needs. With the help of AI’s ability to deeply analyze older adults’ physiological data and make adaptive decisions in complex scenarios, elderly care institutions can provide refined services and meet the personalized needs of older adults. In summary, this paper proposes the second hypothesis as follows:

Hypothesis 2

Promoting older adults’ use of elderly care services is an influencing channel.

2.2.2. Mechanism of older adults’ social participation

AI development can improve older adults’ subjective well-being by promoting their social participation. Social participation refers to activities conducted in the social environment, connected with others, and non-isolated, and it is the key to active aging.³² Activity theory points out that social participation plays an important role in helping older adults redefine themselves, maintain vitality, and meet emotional needs.³³ Intervention strategies to improve older adults’ subjective well-being through social participation have received extensive attention in academia, and sufficient studies have shown the positive effect of social participation on older adults’ subjective well-being.^34–37 However, with aging, factors such as physical function decline have greatly restricted older adults’ social participation. Empowerment theory suggests that the ability of older adults to participate in society can be enhanced through external intervention. AI can act as an external intervention and promote older adults’ social participation through two empowerment modes: external promotion and individual initiative. In terms of the external promotion mode, a series of intelligent application scenarios brought by AI development jointly build an all-round age-friendly environment, creating good external conditions for older adults’ social participation and making social participation more accessible, convenient, and safe. In terms of the individual initiative empowerment mode, AI development can promote older adults’ willingness to participate in society by enhancing their self-efficacy. Self-efficacy refers to the recognition of the possibility that an individual can effectively perform necessary actions in a certain situation,³⁸ and it plays an important role in regulating people’s psychological states and behaviors, and has a significant promoting effect on older adults’ social participation.^39,40 AI development will enhance older adults’ perception of life convenience and support, thereby improving their self-assessment of social participation ability, making them more confident and willing to actively participate in various social activities. Based on the above discussion, the third hypothesis is proposed as follows:

Hypothesis 3

Promoting older adults’ social participation is an influencing channel.

2.3. The moderating effect of older adults’ digital literacy

The association between AI development and older adults’ subjective well-being is moderated by older adults’ digital literacy. The theoretical analysis above shows that AI development brings a supportive intelligent living environment to older adults, thereby enhancing their subjective well-being. Further, the strength of this positive effect depends on whether older adults can make full use of the intelligent living environment, that is, it is influenced by older adults’ digital literacy. Digital literacy reflects the knowledge and ability to use digital technologies ⁴¹ and is crucial for adapting to the complexity of complex information environments.⁴² It is a key individual capital that transforms the macro digital environment an individual is in into personal well-being. Older adults with low digital literacy may have usage barriers and technology anxiety, making it difficult to fully convert the macro AI environment into personal actual benefits. Even in extreme cases, they may experience anxiety due to difficulties in technology adaptation, preventing them from fully enjoying the convenience and opportunities brought by AI development. In contrast, older adults with higher digital literacy have a more positive acceptance attitude and usage ability towards new technologies. They can fully and effectively utilize various intelligent facilities and services brought by AI development, thereby improving their quality of life and subjective well-being more significantly. Hence, this paper proposes the fourth hypothesis as follows:

Hypothesis 4

Digital literacy among older adults plays a positive moderating role.

3. Research design

3.1. Sample and data sources

This paper uses two sets of data for empirical research: data at the older adult individual level and data at the city level. At the older adult individual level, this study adopts two-phase data from the Chinese Longitudinal Aging Social Survey (CLASS) in 2020 and 2023. The CLASS survey is jointly carried out by the Population and Development Research Center and the Institute of Gerontology of Renmin University of China. It uses a carefully designed stratified multi-stage probability sampling method, has good national representativeness and reliability, and is authoritative data for research in fields such as gerontology. At the city level, economic data is sourced from the China Urban Statistical Yearbook and the statistical yearbooks of various provinces and cities, and artificial intelligence patent data is from the CNRDS database and the State Intellectual Property Office. This paper matches the above two sets of data according to the addresses of the interviewed older adults, and deletes samples with serious missing values and those that only participated in one phase of the survey, obtaining balanced panel data with a time span of two phases.

3.2. Variable specification

3.2.1. The dependent variable

The dependent variable in this paper is the subjective well-being of older adults. Life satisfaction is a commonly used measure of subjective happiness and is often used and recommended as a suitable overall summary indicator of subjective happiness.⁴³ In the CLASS survey, the question about subjective well-being is “Generally speaking, are you satisfied with your current life?” The answer options are “very dissatisfied”, “somewhat dissatisfied”, “neutral”, “somewhat satisfied”, and “very satisfied”. In this paper, these options are assigned values from 1 to 5 in sequence, with higher values indicating higher subjective well-being of older adults.

3.2.2. Key idependent variable

The key independent variable in this paper is the level of AI development in cities. Patents are an important symbol for measuring technological progress, which can well reflect the level of technological input and development, and are reliable indicators for evaluating the progress and maturity of AI. Therefore, the number of AI patents is often used to measure the level of AI development.^44–46 In this study, first, according to the AI technology patent classification system table in the Key Digital Technology Patent Classification System (2023) released by the State Intellectual Property Office, AI patent classification numbers are screened out. Then, AI patents are identified in the State Intellectual Property Office using these classification numbers. Finally, the identified AI patents are aggregated at the city level to obtain the number of AI patents in each city. Considering the “right-skewed characteristic” of patent data, the number of patents is further added by 1 and logarithmically processed to obtain the key independent variable of this paper, namely, the level of AI development in cities.

According to the Key Digital Technology Patent Classification System (2023), AI consists of three major branches: first, the AI hardware platform branch, covering subdivisions such as intelligent chips, GPUs, FPGAs, ASICs, brain-inspired chips, and NPUs, involving relevant classifications including G06F*, G06K*, G06N*, G06T*, G06V*, G16B*, G16C*, G16H*, H01L*, and H05K; second, the general AI technologies branch, covering machine learning (including traditional machine learning, reinforcement learning, and deep learning), knowledge graphs, brain-inspired intelligent computing, quantum intelligent computing, pattern recognition, swarm intelligence, and hybrid intelligence, involving relevant classifications such as G06F, G06K*, G06N*, G06V*, G10L*, B82Y10*, G02F2*, A61B5*, and B25J9; third, the key AI technologies branch, including natural language processing, intelligent speech, computer vision, biometric identification, AR/VR, and human-computer interaction, involving relevant classifications such as G06F, G06K*, G06N*, G06V*, G10L*, A61B5*, A63F13*, and G02B27/01.

3.2.3. Instrument variable

Following the methodology of existing studies,^46,47 this paper uses the interaction term between city-level topographic relief and the national number of AI patents in the previous year as the instrumental variable. This instrument is ideal because it satisfies both the relevance assumption and the exclusion restriction assumption:

(1) Relevance assumption. The relevance assumption requires that the instrumental variable must be correlated with AI development. Theoretically, topographic relief directly determines the construction cost of AI infrastructure, the agglomeration of AI-related industries and innovation factors, and the diffusion efficiency of AI technologies within a city, thereby affecting the city’s AI development level. The lagged national AI patent count reflects the overall trend of national AI development and drives AI innovation growth at the city level. Empirically, the relevance assumption is satisfied if the first-stage results of the 2SLS regression are significantly positive, and relevant tests reject the null hypotheses of underidentification and weak instruments.

(2) Exclusion restriction assumption. The exclusion restriction assumption requires that the instrumental variable affects older adults’ subjective well-being only through the channel of city-level AI development, with no direct effect on the dependent variable and no other indirect confounding channels. This paper verifies this assumption from two perspectives. Theoretically, topographic relief is a time-invariant natural geographic endowment formed over long geological history, which does not directly affect the subjective well-being of older adults at the micro level. Empirically, this paper adopts multiple strategies to isolate potential confounding channels. First, we control for a rich set of city-level variables, including economic scale, population size, industrial structure, medical service level, and digitalization level, which absorb the potential effects of topography on well-being through economic development, public service accessibility, and the development of other digital technologies. Second, we include individual fixed effects and year fixed effects. Individual fixed effects absorb all time-invariant unobserved heterogeneity at the individual and city levels (including the time-invariant effect of topography on urban livability), while year fixed effects absorb common national shocks to AI development and older adults’ well-being. Under the above controls, the only valid channel through which the instrumental variable affects the dependent variable is city-level AI development, thus satisfying the exclusion restriction.

In terms of indicator calculation, drawing on the methodology of a relevant study, and based on Digital Elevation Model (DEM) data, the topographic relief degree (RDLS) is calculated using the following formula:

R D L S = A L T / 1000 + {[Max (H) - Min (H)] \times [1 - P (A) / A]} / 500

(1)

In the above formula, RDLS denotes the topographic relief degree; ALT represents the average elevation (in meters) within a certain area centered on a grid cell; Max(H) and Min(H) are the maximum and minimum elevations (in meters) in the area, respectively; P(A) refers to the area of land with a slope less than or equal to 5° in the region (referred to as flat land, in square kilometers); and A is the total area of the region. In this study, a grid cell of 1 km×1 km is set as the basic evaluation unit, so the value of A is 1 km². The acquisition of ALT, Max(H), and Min(H) is realized through the Neighborhood module in the ArcGIS spatial analysis toolbox. P(A) is calculated via the Slope and Reclass commands. The final RDLS is computed using the Map Algebra module. The DEM data are derived from the GLSDEM dataset in the GLS2005 data product, jointly released by the United States Geological Survey (USGS) and the National Aeronautics and Space Administration (NASA), with a spatial resolution of 90 meters.

3.2.4. Control variables

To minimize the interference of omitted variables on the estimation results, this paper sets control variables at the individual level and the city level respectively. At the individual level, the variables include gender, age, marital status, educational attainment, chronic diseases, activities of daily living (ADL), household spending, and family size. This is to fully control the demographic characteristics of older adults. At the city-level, the control variables are set as follows: (1) Economic scale: measured by the gross regional product. (2) Population size: measured by the city’s total population. (3) Industrial structure: measured by the proportion of the tertiary industry in GDP. (4) Medical service level: First, an indicator system was constructed, including the number of hospitals and health centers per capita, the number of beds in hospitals and health centers per capita, and the number of physicians per capita Then, the entropy method was used to calculate the medical service level. (5) Digitalization level: First, an indicator system was constructed, including per capita telecommunications business volume, the number of mobile phone users per 100 people, and the Digital Inclusive Finance Index (jointly compiled by the Digital Finance Research Center of Peking University and Ant Group). Then, the entropy method was used to calculate the city’s digitalization level.

Table 1 reports the definition and descriptive statistical results of the variables.

Table 1.

Variable definition and descriptive statistics.

Variables	Definition	Mean	SD	Min	Max
Dependent variable
Subjective well-being (SWB)	1 = very dissatisfied, 2 = somewhat dissatisfied, 3 = neutral, 4 = somewhat satisfied, 5 = very satisfied	3.78	0.835	1	5
Key independent variable
AI development	The number of authorized artificial intelligence (AI) patents in a city is added by 1, and then the logarithm is taken.	6.21	2.31	1.61	11.12
Control variables (individual level)
Gender	1 = male, 0 = female	0.50	0.50	0	1
Age	The logarithm of the age of surveyed older adults	4.25	0.07	4.09	4.56
Marital status	1 = married, 0 = otherwise (divorced, widowed, and others)	0.82	0.48	0	1
Educational attainment	0=never received education, 1=primary school, 2=junior high school, 3=high school or above	1.28	0.97	0	3
Chronic diseases	0 = no chronic diseases; 1 = have chronic diseases	0.78	0.41	0	1
Activities of daily living (ADL)	Ask older adults whether they need help with six indicators: eating, dressing, getting in and out of bed, using the toilet, walking indoors, and bathing. If the answer is “completely unable to do” or “needs some help”, it is determined that they are disabled in this item and assigned a value of 1; “does not need help from others” is assigned a value of 0.	0.15	0.72	0	6
Household spending	The logarithm of the average monthly per capita household expenditure	6.55	0.90	1.79	10.25
Family size	The logarithm of the number of family members	0.85	0.45	0	2.40
Control variables (city level)
Economic scale	The logarithm of the city’s gross domestic product	17.66	1.24	14.54	19.97
Population scale	The logarithm of the city’s population size	6.45	0.86	4.19	8.07
Industrial structure	The proportion of the added value of the tertiary industry in the city’s GDP	54.55	12.03	28.46	84.85
Medical service level	An index system including the number of hospitals and health centers per capita, hospital beds per capita, and doctors per capita.	0.56	0.17	0	1
Digitalization level	An index system including per capita telecommunications business volume, the number of mobile phone users per 100 people, and the digital inclusive finance index.	0.24	0.16	0	1

3.3. Study nature, time duration and setting

This study is an observational empirical analysis based on nationally representative microdata of older adults matched with city-level macro data. The study period spans from 2020 to 2023, using two waves of microdata from CLASS in 2020 and 2023, as well as city-level AI development and socioeconomic data corresponding to the same period.

3.4. Econometric model

This paper constructs a two-way fixed-effects (TWFE) model to examine the association between AI development and the subjective well-being of older adults. The specific setting is as follows:

S W B_{i t} = α_{0} + α_{1} A I_{c t} + \sum_{j = 1}^{n} β_{j} X_{i j t} + \sum_{z = 1}^{m} μ_{z} X_{c z t} + μ_{i} + λ_{t} + ε_{i t}

(2)

In Formula (2),

i

represents an individual,

c

represents a city, and

t

represents a year. The dependent variable

S W B_{i t}

is the subjective well-being level of individual

i

in year

t

, measured by life satisfaction. The core explanatory variable

A I_{c t}

represents the development level of artificial intelligence in city

c

in year

t

α_{1}

is the estimation coefficient concerned in this paper.

X_{i j t}

is a series of control variables for individual

i

, including demographic characteristics, economic status, family characteristics, etc.

X_{c z t}

is a series of control variables for city

c

in year

t

, including economic scale, population scale, industrial structure, etc.

μ_{i}

represents the individual fixed effect;

λ_{i}

represents the year fixed effect;

ε_{i t}

is the random disturbance term.

3.5. Statistical analysis

This study adopts standard econometric strategies to examine the association between AI development and the subjective well-being of older adults. We use the two-way fixed effects model as the benchmark estimation approach to control for individual and year fixed effects. To address potential endogeneity concerns, we use the instrumental variable approach with two-stage least squares (2SLS) estimation. We conduct multiple robustness checks including double machine learning (DML), replacing core dependent and independent variables, adding additional control variables, and adjusting the level of clustered standard errors. To explore the influencing channels, we replace the dependent variable with elderly care service use and social participation for regression analysis. We further introduce an interaction term to test the moderating role of older adults’ digital literacy, and conduct subgroup regressions to investigate the heterogeneous effects across different groups. All statistical analyses are performed using Stata 17.0.

4. Empirical results

4.1. Benchmark regression results

The results estimating the association between AI development and the subjective well-being of older adults based on the TWFE model are shown in Table 2. Column (1) only includes the key independent variable. The estimated coefficient of AI is 0.054, which is significantly positive at the 1% level. Columns (2) and (3) further incorporate individual-level and city-level control variables in turn. The estimated coefficients are 0.054 and 0.096 respectively, both significantly positive at the 1% level. Each one-standard-deviation increase in AI development is associated with an average increase of 0.27 standard deviations in the subjective well-being of older adults. The above benchmark regression results show that AI development has a significant positive statistical association with the subjective well-being of older adults, which is consistent with Hypothesis 1.

Table 2.

Benchmark regression results.

Variable	(1)	(2)	(3)
Variable	SWB	SWB	SWB
AI development	0.054***	0.054***	0.096***
AI development	(0.015)	(0.015)	(0.017)
Age		-12.961***	-13.768***
Age		(3.378)	(3.438)
Marital status		0.153**	0.146**
Marital status		(0.069)	(0.070)
Educational attainment		0.998	0.967
Educational attainment		(0.725)	(0.737)
Chronic diseases		-0.178***	-0.186***
Chronic diseases		(0.044)	(0.044)
ADL		-0.015	-0.017
ADL		(0.032)	(0.031)
Household spending		-0.054**	-0.048*
Household spending		(0.026)	(0.026)
Family size		-0.028	-0.026
Family size		(0.045)	(0.045)
Economic scale			-0.410***
Economic scale			(0.136)
Population scale			0.458***
Population scale			(0.170)
Industrial structure			-0.017***
Industrial structure			(0.004)
Medical service level			0.806***
Medical service level			(0.158)
Digitalization level			0.058
Digitalization level			(0.069)
Constant term	3.439***	57.635***	65.567***
Constant term	(0.094)	(14.353)	(15.016)
Individual fixed effect	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes
Observations	13914	11804	11594
R-squared	0.835	0.854	0.855

Note. * p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at the individual level are in parentheses.

4.2. Robustness tests

4.2.1 Double machine learning

In this section, the estimation model is further replaced for robustness testing. Chernozhukov et al. proposed the Double Machine Learning (DML) method to estimate the treatment effect.⁴⁸ This method applies machine learning twice to correct biases generated in regularization. Compared with traditional econometric models, DML has advantages in two aspects. First, it solves the problem of model misspecification. In the real world, variable relationships are usually nonlinear; the linear assumption of traditional econometric models may fail to accurately assess impact effects. DML can effectively handle nonlinear data via machine learning algorithms, avoiding model misspecification.⁴⁹ Second, it alleviates the “curse of dimensionality” of traditional econometric models. Since older adults’ subjective well-being is affected by multiple factors, this paper incorporates as many control variables as possible into the regression model. However, estimation biases may arise due to redundant control variables. DML’s regularization algorithm can automatically screen high-dimensional control variables, extracting key ones with significant predictive performance, thus alleviating biases from redundant controls. As shown in Table 3, the estimated coefficient of the key independent variable remains significantly positive, supporting the robust positive association between AI development and older adults’ subjective well-being.

Table 3.

Results of robustness tests for DML.

Variable	(1) Lasso	(2) Random forest
Variable	SWB	SWB
AI development	0.272***	13.19**
AI development	(0.028)	(6.58)
Control Variables	Yes	Yes
Individual fixed effect	Yes	Yes
Year fixed effect	Yes	Yes
Observations	12195	12195

4.2.2. Dealing with Endogeneity

Although the key independent variable in this paper is the AI development level of cities at the macro level, there is unlikely to be a serious endogeneity problem. However, to enhance the reliability of the research conclusions, this paper still uses the instrumental variable method for robustness testing. The potential endogeneity problems are as follows. First, the reverse causality problem. The reverse causality problem in this study is not serious, because the AI development level of cities is mainly affected by macro factors such as a city’s economic scale, public services, and industrial structure, while the subjective well-being of older adults at the micro level is unlikely to affect the AI development level of cities at the macro level. Second, the omitted variable problem. There are numerous factors affecting the quality of life of older adults, and many macro variables will affect the well-being of older adults while influencing the AI development level of cities. Although the baseline regression in this paper includes macro variables such as economic size, population size, and industrial structure as well as two-way fixed effects, the issue of omitted variables may still exist. Therefore, the estimation results may still be disturbed by the omitted variable problem. Based on the above discussion, this paper further adopts the instrumental variable method to alleviate potential endogeneity problems and ensure the reliability of the research conclusions.

Columns (1) and (2) of Table 4 report the estimation results of the Two-Stage Least Squares (2SLS) method. The value of the LM statistic is 120.967, which is significant at the 1% level, rejecting the null hypothesis of underidentification of the IV. The Wald F statistic is 116.161, exceeding the 10% critical value, meaning that the IV selected is not a weak instrumental variable. In terms of regression coefficients, the estimated coefficient of the instrumental variable is significant at the 1% level, indicating that the correlation between the IV and the development of AI of cities is satisfied. The estimation results of the second stage show that the estimated coefficient of the key independent variable is 0.503, which is significantly positive at the 1% level. Furthermore, to ensure the robustness of the results, this paper also uses DML based on support vector machines for estimation. As shown in Column (3) of Table 4, the estimated coefficient of AI development remains significantly positive. As shown in the above results, after excluding the interference of potential endogeneity problems through the instrumental variable method, the robust positive association between AI development and older adults’ subjective well-being still holds.

Table 4.

Results of robustness tests for the instrumental variable method.

Variable	(1)	(2)	(3)
	2SLS		DML (IV)
	AI	SWB	SWB
AI development		0.503***	0.373***
AI development		(0.150)	(0.079)
IV	-0.452***
IV	(0.043)
Control variables	Yes	Yes	Yes
Individual fixed effect	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes
LM	120.967***
CD Wald F	116.161
Observations	11326		11923

4.2.3. Replacing the dependent variable

To avoid the interference of the contingency in variable selection on conclusions, this section conducts robustness analysis by replacing the dependent variable. In the benchmark regression, this paper follows the common practice in existing research and uses life satisfaction to represent the subjective well-being level of older adults. Different from the benchmark regression, this section uses the following two indicators as dependent variables for robustness testing.

(1) Depression risk. According to China Ageing Development Report 2024: Psychological Health Status of Chinese Older Adults, 26.4% of older adults in China have different degrees of depressive symptoms. Considering the serious adverse impact of depression on older adults’ lives, this section uses the depression risk index as a measure of subjective well-being for robustness testing. Based on the characteristics of CLASS data, 9 questions related to the Center for Epidemiologic Studies Depression Scale (CES-D) are used to measure the depression risk of older adults. Specifically, they include positive questions such as “Did you feel in a good mood in the past week?” and “Did you have a lot of fun in life in the past week?”, and negative questions such as “Did you feel sad in the past week?” and “Did you feel you had nothing to do in the past week?”. The answer options include “No”, “Sometimes”, and “Often”, which are assigned 1-3 points in this paper. The positive and negative scoring method is used to calculate the depression risk of older adults, with the score range being 9-27 points. The higher the score, the higher the depression risk of older adults.

(2) Self-rated health. Self-rated health is older adults’ comprehensive subjective judgment of their physical and mental state, which can reflect their subjective feelings about their living state to a certain extent. In the CLASS survey, the item about self-rated health is “How do you think your current physical health status is?”, and the answer options are “Very unhealthy”, “Relatively unhealthy”, “Average”, “Relatively healthy”, and “Very healthy”. The above options are assigned 1-5 points in turn, and a positive indicator of older adults’ self-rated health level can be obtained.

The estimated coefficients of the key independent variable are all significant at the 1% level, indicating that AI development helps reduce the depression risk of older adults and improve their self-rated health, which verifies the positive correlation between AI development and the subjective well-being of older adults to a certain extent.

4.2.4. Replacing the key independent variable

Besides replacing the dependent variable, this paper conducts robustness tests by replacing the independent variable in two ways. First, lag the independent variable by one period. This is because the impact of AI development on older adults may be lagged, and lagging by one period also helps eliminate potential reverse causality. Second, change the construction method of the AI development variable. Text analysis is a common method to identify AI patents. This paper refers to relevant research,⁵⁰ builds a dictionary of AI-related keywords (such as “machine learning”, “intelligent elderly care”, “intelligent medical care”) to retrieve AI patents, aggregates them at the city level and performs logarithmic processing, thus reconstructing the AI development variable. Columns (3) and (4) of the Table 5 show that the replaced key independent variable is still significantly positive, verifying the robustness of the benchmark regression.

Table 5.

Results of robustness tests for replacing variables.

Variable	(1)	(2)	(3)	(4)
	Replacing the dependent variable		Replace the key independent variable
	Depression risk	Self-rated health	SWB	SWB
AI development	-0.580***	0.081***
AI development	(0.079)	(0.017)
AI development (one-period lag)			0.047***
AI development (one-period lag)			(0.017)
AI development (text analysis)				0.033***
AI development (text analysis)				(0.008)
Control variables	Yes	Yes	Yes	Yes
Individual fixed effect	Yes	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes	Yes
Obserations	10130	11676	11594	11492
R-squared	0.848	0.864	0.855	0.854

4.2.5. Adding control variables

As the 2020 CLASS questionnaire didn’t survey the income of non-retired older adults, to avoid massive missing values, the benchmark regression uses per capita household expenditure as a control variable for older adults’ economic status. In this section, older adults’ income level is added to control variables and the regression is redone. From column (1) of the Table 6, even though the sample size drops sharply from 11,594 (in the benchmark regression) to 5,038, the estimated coefficient of AI development is still significantly positive at the 1% level, further confirming the research conclusion’s robustness.

Table 6.

Results of robustness tests for adding control variables and changing the clustering level.

Variable	(1)	(2)	(3)
Variable	Adding control variables	Cluster at the city level	Cluster at the province level
AI development	0.077***	0.096***	0.096**
AI development	(0.028)	(0.036)	(0.038)
Control variables	Yes	Yes	Yes
Individual fixed effect	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes
Obserations	5038	11594	11594
R-squared	0.861	0.855	0.855

4.2.6. Changing the level of clustered standard errors

The benchmark regression sets clustered standard errors at the individual level. Considering that older adults in the same region may influence each other to have similar ideas and behaviors, this section further sets clustered standard errors at the city and province levels. From columns (2) and (3) of the Table 6, the clustering level has little impact on the estimated coefficients’ significance, proving the conclusion’s reliability.

5. Extended discussion

5.1. Mechanism tests

The theoretical analysis above shows that AI development can improve older adults’ subjective well-being by promoting their use of elderly care services and social participation. This section conducts empirical tests on these two mechanisms. Following Jiang’s suggestions on mechanism analysis,⁵¹ this paper replace the dependent variable in Equation (1) to the mechanism variables (elderly care service use and social participation) focused on in this study for regression analysis.

5.1.1 Mechanism of older adults’ use of elderly care services

This section tests the mechanism of elderly care service use. Based on the nature of elderly care services, this paper classifies them into two major types (Table 7): (1) Medical and health care services, mainly catering to older adults’ health promotion and integrated care needs, including home nursing, home medical treatment, rehabilitation training, rental of rehabilitation aids, free physical examinations, establishment of health records, and health lectures. (2) Daily care services, mainly meeting older adults’ basic living and instrumental needs, including home visits, elderly service hotlines, accompanying medical treatment, assistance with daily shopping, legal aid, home housework, elderly canteens or meal delivery, day care centers, and psychological counseling. For the specific variable construction method, to ensure the robustness of estimation results, this paper respectively constructs a dummy variable for whether older adults use elderly care services and a continuous variable for the degree of elderly care service use. Specifically, (1) Dummy variable: If an older adult has used any of these services in the past 12 months, it is assigned a value of 1; otherwise, 0. (2) Continuous variable: The number of elderly care service items used by older adults is summed up. The value range of the medical and health care service variable is 0-7, and that of the daily care service variable is 0-9. The higher the value, the higher the degree of older adults’ use of elderly care services.

Table 7.

Types of elderly care services and variable construction.

Service category	Core orientation	Specific service items	Variable construction method
Medical and health care services	Mainly catering to older adults’ health promotion and integrated care needs	Home nursing, home medical treatment, rehabilitation training, rental of rehabilitation aids, free physical examinations, establishment of health records, health lectures	To ensure the robustness of estimation results, this paper respectively constructs: (1) A dummy variable for whether older adults use elderly care services (2) A continuous variable for the degree of elderly care service use
Daily care services	Mainly meeting older adults’ basic living and instrumental needs	Home visits, elderly service hotlines, accompanying medical treatment, assistance with daily shopping, legal aid, home housework, elderly canteens or meal delivery, day care centers, psychological counseling

Columns (1) - (4) of the Table 8 report the test results with elderly care service use as the mechanism. Regression results show that whether it is medical and health care services or daily care services, AI development can significantly promote older adults’ use of them. Moreover, existing empirical studies find that older adults’ use of elderly care services can significantly improve their subjective well-being.^52,53 Therefore, AI development can improve their subjective well-being by promoting older adults’ use of elderly care services. Based on the above discussion, the mechanism of elderly care service use holds, and Hypothesis 2 is verified.

Table 8.

Results of the mechanism test for elderly care service use.

Variable	(1)	(2)	(3)	(4)
	Elderly care services (medical and health care)		Elderly care services (daily care)
	Dummy variable	Continuous variable	Dummy variable	Continuous variable
AI development	0.086***	0.219***	0.036***	0.075***
AI development	(0.013)	(0.024)	(0.008)	(0.019)
Control variables	Yes	Yes	Yes	Yes
Individual fixed effect	Yes	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes	Yes
Obserations	11686	11686	11686	11686
R-squared	0.800	0.756	0.801	0.715

5.1.2. Mechanism of older adults’ social participation

This part tests the mechanism of social participation. In the CLASS questionnaire, the question about older adults’ social participation is “How often have you participated in the following activities in the past year?” The specific activities include community security patrols, caring for other elderly people, environmental sanitation protection, dispute mediation, accompanying chat, voluntary services requiring professional skills, and helping to look after other people’s children.

This paper constructs a dummy variable for whether older adults participate in social activities and a continuous variable for the degree of social participation respectively. (1) Dummy variable: If an older adult participates in any of the above social activities, it is assigned a value of 1; otherwise, it is assigned a value of 0. (2) Continuous variable: The response options for each of the above social activities are “Never participated”, “Several times a year”, “At least once a month”, “At least once a week”, and “Almost every day”. This paper assigns values from 0 to 4 to the above options in turn and sums them up to obtain a continuous variable of older adults’ social participation degree with a value range of 0-28. The higher the value, the higher the participation level of older adults in the activity.

Columns (1) and (2) of the Table 9 report test results with social participation as the mechanism. Regression results show that regardless of variable construction, the estimated coefficient is significantly positive, indicating AI development promotes older adults’ social participation. Notably, this paper also conducts robustness tests by taking logarithms and constructing a high-participation variable (coded as 1 if participation > 10, and 0 otherwise) to avoid disturbances to the conclusions caused by long-tailed distributions. As shown in columns (3) and (4) of Table 9, the estimated coefficients of AI development are all significantly positive. In addition, considering the problem of excessive zero values in the continuous variable, we re-ran the regression after excluding about 5,000 zero-value samples. As shown in column (5) of Table 9, the estimated coefficient of AI development remains significantly positive, indicating that the results are highly robust. Further, combined with previous empirical studies supporting that older adults’ social participation significantly improves subjective well-being,^54,55 we can see AI development improves subjective well-being by promoting older adults’ social participation. In summary, the social participation mechanism holds, and Hypothesis 3 is verified.

Table 9.

Results of the mechanism test for social participation.

Variable	(1)	(2)	(3)	(4)	(5)
	Social participation
	Dummy variable	Continuous variable	High participation	Logarithm	Sample exclusion
AI development	0.036***	0.352***	0.018***	0.087***	0.365**
AI development	(0.010)	(0.079)	(0.005)	(0.019)	(0.152)
Control variables	Yes	Yes	Yes	Yes	Yes
Individual fixed effect	Yes	Yes	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes	Yes	Yes
Obserations	11686	11686	11686	11686	5276
R-squared	0.837	0.829	0.8319	0.8416	0.845

5.2. Moderating effect tests

The previous theoretical analysis suggests that digital literacy can enhance the positive association between AI development and older adults’ subjective well-being. This section conducts an empirical test on this relationship. Based on the characteristics of CLASS data and considering rationality and data availability, this paper constructs a comprehensive evaluation index system to measure older adults’ digital literacy level. As shown in Table 10, the index system includes 12 indicators across 8 dimensions: digital operation literacy, digital information literacy, digital entertainment literacy, digital social literacy, digital life literacy, digital learning literacy, digital business literacy, and digital health literacy. Furthermore, the equal-weight weighting method is used for objective weighting to obtain older adults’ digital literacy level, with a value range of 0-8 (a higher value means a higher digital literacy level). Table 11 reports the correlation coefficients between sub-dimensions. Overall, the correlation coefficients fall within the ideal range of 0.2 to 0.8, which not only ensures the independence of each dimension but also reflects their internal connection in jointly forming digital literacy.

Table 10.

Comprehensive evaluation indicator system for digital literacy of older adults.

Dimension	Indicator	Score
Digital operation literacy	Whether familiar with a certain Internet - accessing device	1
Digital information literacy	Whether take the Internet as the main information source	0.33
	Whether read news through the Internet	0.33
	Whether browse various articles and information except news via the Internet	0.33
Digital entertainment literacy	Whether listen to music (broadcast) or watch videos via the Internet	0.5
Digital entertainment literacy	Whether play online games	0.5
Digital social literacy	Whether make voice or video calls	0.5
Digital social literacy	Whether chat in text form	0.5
Digital life literacy	Whether use the Internet for transportation travel	1
Digital learning literacy	Whether learn or take training courses through the Internet	1
Digital business literacy	Whether conduct Internet - based investment and financial management (such as stock trading, fund purchasing, etc.)	1
Digital health literacy	Whether manage health through the Internet	1

Table 11.

Comprehensive evaluation indicator system for digital literacy of older adults.

	Digital operation literacy	Digital information literacy	Digital entertainment literacy	Digital social literacy	Digital life literacy	Digital learning literacy	Digital business literacy	Digital health literacy
Digital operation literacy	1
Digital information literacy	0.6974	1
Digital entertainment literacy	0.5266	0.6681	1
Digital social literacy	0.7241	0.7759	0.6592	1
Digital life literacy	0.0922	0.1328	0.1334	0.1001	1
Digital learning literacy	0.2220	0.2483	0.1535	0.2078	0.2030	1
Digital business literacy	0.3991	0.4395	0.3852	0.3827	0.1892	0.3036	1
Digital health literacy	0.2890	0.3190	0.2396	0.2543	0.2356	0.2817	0.4181	1

The interaction term between digital literacy and AI is added to Equation (1) for regression. If the estimated coefficient of the interaction term is significant, the moderating effect test is passed. From the estimation results reported in Table 12 (Columns 1 and 2), the estimated coefficient of the interaction term is always significantly positive at the 1% level, indicating that the higher older adults’ digital literacy is, the greater the improvement effect of AI development on their subjective well-being. In addition, to more intuitively reflect the moderating effect of older adults’ digital literacy, this paper also draws a moderating effect diagram and a marginal effect diagram. As shown in Figures 1 and 2, when older adults’ digital literacy is at a relatively high level, the improvement effect of AI development on their subjective well-being is stronger. Furthermore, to enhance the robustness of the conclusions, this paper also constructs an alternative indicator of digital literacy. Specifically, following the principle of health as the core, whether older adults use digital health devices is taken as the proxy variable for their digital literacy level. If older adults use intelligent health devices (including intelligent wheelchairs, electronic blood pressure monitors, blood lipid detectors, smart bracelets, infrared cameras, smart all-in-one machines, smart sleep detectors, and audiobooks), it is assigned a value of 1; otherwise, it is 0. As shown in Table 12 (Column 3), the coefficient of the interaction term is significantly positive at the 1% level, indicating that the moderating effect of digital literacy is robust. In summary, digital literacy can strengthen the positive association between AI development and older adults’ subjective well-being, and Hypothesis 4 is supported.

Table 12.

Results of the moderating effect test.

Variable	(1)	(2)	(3)
Variable	SWB	SWB	SWB
AI development	0.043***	0.085***	0.102***
AI development	(0.015)	(0.016)	(0.017)
Interaction term	0.011**	0.010**	0.054***
Interaction term	(0.004)	(0.005)	(0.017)
Digital literacy	0.046***	0.060***	-0.039
Digital literacy	(0.011)	(0.013)	(0.038)
Control variables	No	Yes	Yes
Individual fixed effect	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes
Obserations	13914	11594	11594
R-squared	0.837	0.857	0.856

Figure 1.

The moderating effect of digital literacy.

Figure 2.

Marginal effect.

5.3. Heterogeneity analysis

To further examine whether the association between AI development and subjective well-being varies across groups of older adults with different characteristics, this paper conducts a heterogeneity analysis based on their gender, age, education level, marital status, and residence type. Among them, those over 70 years old are classified as the old-old group, and those below 70 years old are classified as the young-old group; those with an education level above primary school are classified as the highly educated group. As shown in Table 13, the positive association between AI development and the subjective well-being of older adults is significant across all subsamples. The p-values of the between-group difference tests for gender, education level, marital status, and urban-rural dimensions are all greater than 0.05, while the p-value for the age dimension is 0.016 (less than 0.05), indicating that the estimated coefficient for the old-old group (0.163) is higher than that for the young-old group (0.065). These results suggest that the positive association between AI development and the subjective well-being of older adults exhibits significant age heterogeneity: the old-old group benefits more prominently, while the differences in other characteristic dimensions are not statistically significant. Compared with young-old older adults, old-old older adults have weaker self-care ability and social adaptability, and are more dependent on external support. Therefore, the positive association between AI development and subjective well-being is stronger for this group.

Table 13.

Results of heterogeneity analysis.

Variable	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Variable	Female	Male	Young-old	Old-old	Low-educated	High-educated	Unmarried	Married	Rural	Urban
AI development	0.116***	0.076***	0.065**	0.163***	0.107***	0.085***	0.107***	0.090***	0.103***	0.078***
AI development	(0.025)	(0.022)	(0.026)	(0.034)	(0.023)	(0.023)	(0.037)	(0.019)	(0.028)	(0.024)
Constant term	80.793***	53.985***	27.472	164.155***	83.920***	37.950*	68.368**	62.348***	77.176***	56.597***
Constant term	(21.997)	(20.361)	(39.868)	(46.579)	(20.480)	(21.913)	(28.054)	(16.600)	(24.637)	(19.388)
Individual fixed effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year fixed effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Observations	5818	5776	5276	3480	6910	4684	1784	9234	6254	4598
Between-group difference test	0.170		0.016		0.258		0.452		0.294
R-squared	0.854	0.858	0.868	0.824	0.842	0.877	0.908	0.865	0.864	0.848

6. Limitations

This study has several limitations that need to be explicitly noted.

First, there are limitations regarding the measurement of the core explanatory variable. This paper employs city-level AI patent counts to measure regional AI development, which is a well-established and widely adopted approach in the existing literature and can reliably reflect regional technological innovation and supply capacity. Nevertheless, constrained by data availability, this indicator still cannot fully capture the AI application environment that older adults truly perceive, access, and use in daily life. Specifically, patent counts mainly reflect the R&D reserves and innovation vitality of urban AI technologies, and may not fully reflect the actual penetration of AI technologies in scenarios closely related to older adults, such as elderly care, medical care, transportation, and daily life services. The AI technology layout in some cities may be more concentrated in industrial and commercial fields rather than daily life scenarios directly accessible to older adults. Meanwhile, patent authorization only represents the output of technological innovation and is not equivalent to actual technology application. Some patents have not been transformed into age-friendly products and services that older adults can conveniently use, so there is still room for improvement in accurately capturing the AI environment truly accessible to older adults. In addition, patent counts cannot reflect the age-friendly design, operational convenience, and actual user experience of AI applications. Even if relevant technologies are implemented, they may not be effectively used by older adults if they fail to fully adapt to the cognitive and behavioral characteristics of older adults. Overall, under existing data constraints, city-level AI patent counts remain a relatively reasonable and reliable indicator for measuring the macro AI development environment. This paper also mitigates potential measurement bias through robustness checks such as replacing the indicator construction method and reconstructing AI indicators via text mining.

Second, this study only verifies the mediating roles of elderly care service use and social participation, as well as the heterogeneity of individual demographic characteristics. Other potential influencing mechanisms and heterogeneous effects may still exist and need to be further explored in future research.

7. Conclusion and suggestions

Based on the micro data of older individuals from the China Longitudinal Aging Social Survey (CLASS) in 2020 and 2023, combined with the AI development levels of Chinese cities measured by patent data, this paper empirically analyzes the association between AI development and older adults’ subjective well-being. The conclusions of this study are as follows. First, the baseline regression results show that AI development is positively correlated with the subjective well-being of older adults. Second, the mechanism analysis suggests that the use of elderly care services and social participation among older adults may serve as potential channels through which AI development is associated with their subjective well-being. Third, moderating effect analysis shows that the digital literacy of older adults plays a crucial moderating role in the process of AI development influencing their subjective well-being. Improving the digital literacy of older adults helps to release the improvement effect of AI development on their subjective well-being. Fourth, the heterogeneity analysis shows that the positive association of AI development is stronger for the old-old group than for the young-old group. In addition, this paper uses double machine learning, the instrumental variable method, replacement of variable construction methods, and change of clustering levels to conduct robustness tests, thereby confirming the reliability of the estimation results in this paper. Based on the above conclusions, in order to better exert the positive role of AI development on the subjective well-being of older adults and help them share the development achievements of the digital era, this paper puts forward the following policy suggestions.

First, precisely unlock the positive effect of AI development on the subjective well-being of older adults. The baseline regression confirms a significant positive correlation between AI development and older adults’ subjective well-being. Therefore, AI-enabled elderly care should be integrated into the core policy framework for actively responding to population aging. Guided by the real needs of older adults, we will promote the in-depth integration and adaptive transformation of AI technologies in high-frequency life scenarios such as elderly care services, health management, and social participation. By lowering technical barriers through age-friendly design, we can ensure that the dividends of the digital era are more fairly and accessibly delivered to older adults, effectively enhancing their subjective well-being.

Second, strengthen the well-being transmission of AI through dual channels of elderly care services and social participation. Mechanism analysis shows that the use of elderly care services and social participation are critical mediating paths through which AI improves older adults’ well-being. We should focus on building a dual-support system of “AI + elderly care services” and “AI + social participation”: on one hand, optimize the supply of elderly care services with intelligent technologies and build an integrated online-offline smart elderly care platform to precisely match essential services such as meal assistance, bathing assistance, and health monitoring; on the other hand, expand channels for older adults’ social interaction, public welfare participation, and interest-based learning through AI-driven digital platforms.

Third, take the improvement of digital literacy as a core lever to amplify the positive moderating effect of AI. Moderating effect analysis reveals that older adults’ digital literacy plays a significantly positive moderating role in the process of AI affecting their subjective well-being. Therefore, enhancing the digital literacy of older adults should be a key breakthrough to unlock the well-being potential of AI. We will establish a hierarchical and classified digital skills training system: for groups with weak basic abilities, we offer introductory training on device operation and information acquisition; for those with basic competence, we focus on advanced content such as risk prevention, health management, and online socialization. Meanwhile, we will improve the community “digital assistance for the elderly” mechanism, providing one-on-one guidance from volunteers and practical courses in senior universities to help older adults cross the digital divide, master AI-related technologies and services, and fully enjoy the convenience and well-being brought by technological progress.

Fourth, prioritize the AI well-being needs of the old-old group. Heterogeneity analysis finds that the well-being improvement effect of AI development is significantly stronger for the old-old group than for the young-old group. Thus, in policy design and resource allocation, we should tilt toward vulnerable older groups such as the old-old, disabled, and empty-nest elderly. We will prioritize the R&D of easy-to-operate smart health monitoring devices, emergency call systems, and age-friendly smart home products tailored for the old-old, and optimize AI-enabled elderly care services and social participation scenarios for this group. This will ensure that technological progress more accurately benefits the most vulnerable elderly groups.

Footnotes

Acknowledgements

Thanks to the Institute of Gerontology at Renmin University of China for providing the CLASS data.

ORCID iD

Fulei Jin

Author contributions

F. J. led the research design, data collection, and analysis, independently wrote the first draft of the paper, and was responsible for the overall revision and optimization of the manuscript. S. X. assisted in the research design, undertook literature research and data processing support, participated in the review and revision of the paper, and jointly ensured that the research results met the publication standards.

Funding

This work was supported by Shandong Provincial Natural Science Foundation (ZR2025QC1268).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets used during the current study are available from the Institute of Gerontology at Renmin University of China.*

References

Diener

Oishi

Lucas

. Personality, culture, and subjective well-being: Emotional and cognitive evaluations of life. Annu Rev Psychol 2003; 54(1): 403–425. https://doi.org/10.1146/annurev.psych.54.101601.145056

Kumar

. Biases in artificial intelligence applications affecting human life: A review. Int J Recent Technol Eng 2021; 2021(1): 54–55.

Watkins

Koch

Sherwood

, et al. Association of anxiety and depression with all-cause mortality in individuals with coronary heart disease. J Am Heart Assoc 2013; 2(2): e000068. https://doi.org/10.1161/JAHA.112.000068

Hessami

. The size and composition of government spending in Europe and its impact on well-being. Kyklos 2010; 63(3): 346–382. https://doi.org/10.1111/j.1467-6435.2010.00478.x

Papastergiou

Latinopoulos

Evdou

, et al. Exploring associations between subjective well-being and non-market values when used in the evaluation of urban green spaces: A scoping review. Land 2023; 12(3): 700. https://doi.org/10.3390/land12030700

Chyi

Mao

. The determinants of happiness of China's elderly population. J Happiness Stud 2012; 13(1): 167–185. https://doi.org/10.1007/s10902-011-9256-8

Steptoe

Deaton

Stone

. Subjective wellbeing, health, and ageing. Lancet 2015; 385(9968): 640–648. https://doi.org/10.1016/S0140-6736(13)61489-0

Jin

Chen

. Community nursing on subjective well-being of the elderly: Evidence from CLASS data. BMC Nurs 2025; 24(1): 757. https://doi.org/10.1186/s12912-025-03205-7

Kandilov

. Does mobile Internet use affect the subjective well-being of older Chinese adults? An instrumental variable quantile analysis. J Happiness Stud 2021; 22(7): 3137–3156. https://doi.org/10.1007/s10902-021-00365-6

10.

Chen

Sun

. The impact of mobile payment on the subjective quality of life of rural middle aged and elderly residents and its implications for ethnic regions. Guizhou Ethn Stud 2024; 45(4): 62–69.

11.

Yan

. The impact of internet application usage on the mental health of older adults: Taking WeChat, short videos, audio-visual, and games as examples. Popul J 2023; 45(3): 78–89.

12.

Duan

. Digital villages construction and subjective well-being of the elderly: Evidence from China longitudinal aging social survey. Appl Res Qual Life 2026; 21: 225–244. https://doi.org/10.1007/s11482-024-10362-5

13.

Duan

. Digital divide or digital dividend: The impact of digital economy development on the elderly. Behav Inf Technol 2025.

14.

Zeng

Han

Liu

, et al. The impact of digital finance on elderly physical health: Evidence from China. Digit Health 2025; 11: 20552076251353318. https://doi.org/10.1177/20552076251353318

15.

Shen

Shi

. The impact of urban artificial intelligence development on individual labor participation and working hours: Findings based on artificial intelligence technology density. Chin J Popul Sci 2025; 39(2): 45–61.

16.

Tang

Shao

, et al. Smart wearable devices and elderly health: Evidence from smart bracelets. Popul J 2023; 45(6): 50–67.

17.

Zhang

Jiang

Liu

. Smart wearables and subjective well-being among Chinese older adults. Intell Soc Res 2025; 2025(1): 132–151+243.

18.

Neves

Franz

Judges

, et al. Can digital technology enhance social connectedness among older adults? A feasibility study. J Appl Gerontol 2019; 38(1): 49–72. https://doi.org/10.1177/0733464817741369

19.

Zhong

Cai

. The impact of smart senior care products on mental health among older adults: Evidence from the China longitudinal aging social survey. Popul Dev 2025; 31(2): 116–126.

20.

Bronfenbrenner

. Toward an experimental ecology of human development. Am Psychol 1997; 32(7): 513–531. https://doi.org/10.1037/0003-066x.32.7.513

21.

Lawton

Simon

. The ecology of social relationships in housing for the elderly. Gerontologist 1968; 8(2): 108–115. https://doi.org/10.1093/geront/8.2.108

22.

Lien

Steggell

Iwarsson

. Adaptive strategies and person-environment fit among functionally limited older adults aging in place: A mixed methods approach. Int J Environ Res Public Health 2015; 12(9): 11954–11974. https://doi.org/10.3390/ijerph120911954

23.

Battarra

Zucaro

Tremiterra

. Smart mobility and elderly people. Can ICT make the city more accessible for everybody? TeMA-J Land Use Mob. Environ 2018; 23: 23–42.

24.

Bokolo

. Inclusive and safe mobility needs of senior citizens: Implications for age-friendly cities and communities. Urban Sci 2023; 7(4): 103. https://doi.org/10.3390/urbansci7040103

25.

Yan

Jiang

Huang

, et al. Intelligent urbanism with artificial intelligence in shaping tomorrow’s smart cities: Current developments, trends, and future directions. J Cloud Comput 2023; 12: 179. https://doi.org/10.1186/s13677-023-00569-6

26.

Wolniak

Stecuła

. Artificial Intelligence in Smart Cities-Applications, Barriers, and Future Directions: A Review. Smart Cities 2024; 7(3): 1346–1389. https://doi.org/10.3390/smartcities7030057

27.

Lohmann

. Information technologies and subjective well-being: Does the Internet raise material aspirations? Oxf Econ Pap 2015; 67(3): 740–759. https://doi.org/10.1093/oep/gpv032

28.

Cai

Liu

. Urban digital transformation and residents' subjective well-being. Econ Theory Bus Manag 2023; 43(12): 107–120.

29.

Zeng

. Effects of Community Pension Service Facilities on the Subjective Welfare of the Elderly in Urban and Rural Areas. Popul Dev 2022; 28(6): 148–160.

30.

. The impact of community elderly care services on the subjective well-being of older adults: Evidence from the Weifang elderly health survey. Dong Yue Trib 2023; 44(12): 145–154.

31.

Zhang

. Human resource dilemma and intelligence choice of elderly care service in China. J Xi'an Jiaotong Univ (Soc Sci) 2023; 43(6): 152–163.

32.

Löfgren

Nyman

Larsson

, et al. Fostering social participation among older adults: Perspectives of stakeholders. Scand J Occup Ther 2024; 31(1): 2384405. https://doi.org/10.1080/11038128.2024.2384405

33.

Havighurst

Albrecht

. Older People. Longmans Green, 1953.

34.

Dury

Stas

Switsers

, et al. Gender-related differences in the relationship between social and activity participation and health and subjective well-being in later life. Soc Sci Med 2021; 270: 113668. https://doi.org/10.1016/j.socscimed.2020.113668

35.

Mackenzie

Abdulrazaq

. Social engagement mediates the relationship between participation in social activities and psychological distress among older adults. Aging Ment Health 2021; 25(2): 299–305. https://doi.org/10.1080/13607863.2019.1697200

36.

Cai

. Does social participation improve cognitive abilities of the elderly? J Popul Econ 2022; 35: 591–619. https://doi.org/10.1007/s00148-020-00817-y

37.

Shimoda

Tomida

Nakajima

, et al. Association between the level of social participation and depressive symptoms among older Japanese adults: A cross-sectional survey. Psychogeriatrics 2024; 24(5): 1095–1102. https://doi.org/10.1111/psyg.13163

38.

Bandura

. Self-efficacy: Toward a unifying theory of behavioral change. Psychol Rev 1977; 84(2): 191–215. https://doi.org/10.1037//0033-295x.84.2.191

39.

Tadaka

. Development of the Self-efficacy for Social Participation scale (SOSA) for community-dwelling older adults. BMC Public Health 2023; 23(1): 2294. https://doi.org/10.1186/s12889-023-16774-6

40.

Xie

, et al. Association between exercise self-efficacy and physical activity in elderly individuals: A systematic review and meta-analysis. Front Psychol 2025; 16: 1525277. https://doi.org/10.3389/fpsyg.2025.1525277

41.

Gilster

. Digital literacy. Wiley Computer Pub, 1997.

42.

Bawden

. Information and digital literacies: A review of concepts. J Doc 2001; 57(2): 218–259. https://doi.org/10.1108/eum0000000007083

43.

Cheng

Yang

. Impact of informal caregiving on caregivers' subjective well-being in China: A longitudinal study. Arch Public Health 2023; 81(1): 209. https://doi.org/10.1186/s13690-023-01220-1

44.

Luo

Feng

. Exploring artificial intelligence and urban pollution emissions: “Speed bump” or “accelerator” for sustainable development? J Clean Prod 2024; 463: 142739. https://doi.org/10.1016/j.jclepro.2024.142739

45.

Chen

Zhang

. Research on the impact of artificial intelligence technology on urban public health resilience. Front Public Health 2025; 12: 1506930. https://doi.org/10.3389/fpubh.2024.1506930

46.

Qiao

Qian

. Artificial intelligence technology, entrepreneurial ecosystem, and new firm entry. Quant Tech Econ 2025; 42(8): 110–130.

47.

. The impact of market-based allocation of factors in China on the new trend of urban-rural income gap. People's Publishing House, 2025, pp. 188–336.

48.

Chernozhukov

Chetverikov

Demirer

, et al. Double/debiased machine learning for treatment and structural parameters. Econometrics J 2018; 21(1): C1–C68. https://doi.org/10.1111/ectj.12097

49.

Yang

Ren

, et al. Does the development of the Internet contribute to air pollution control in China? Mechanism discussion and empirical test. Struct Chang Econ Dyn 2021; 56: 207–224. https://doi.org/10.1016/j.strueco.2020.12.001

50.

Yao

Zhang

Guo

, et al. How does artificial intelligence improve firm productivity? based on the perspective of labor skill structure adjustment. Manag World 2024; 40(2): 101–122.

51.

Jiang

. Mediating effects and moderating effects in causal inference. China Ind Econ 2022; 40(5): 100–120.

52.

Zhang

Mao

Shui

, et al. Do community home-based elderly care services improve life satisfaction of Chinese older adults? An empirical analysis based on the 2018 CLHLS dataset. Int J Environ Res Public Health 2022; 19: 15462. https://doi.org/10.3390/ijerph192315462

53.

Wei

. The impact of using community home-based elderly care services on older adults' self-reported health: Fresh evidence from China. Front Public Health 2023; 11: 1257463. https://doi.org/10.3389/fpubh.2023.1257463

54.

Zhang

. Social participation and subjective well-being among retirees in China. Soc Indic Res 2015; 123(1): 143–160. https://doi.org/10.1007/s11205-014-0728-1

55.

Shek

DTL

, et al. The relationship between social participation and subjective well-being among older people in the Chinese culture context: The mediating effect of reciprocity beliefs. Int J Environ Res Public Health 2022; 19(23): 16367. https://doi.org/10.3390/ijerph192316367