Financial investments in AI-based technologies and carbon footprint in selected advanced industrial economies

Abstract

Artificial intelligence (AI) has rapidly expanded across multiple industries and technologies, driving economic growth and offering innovative solutions to structural challenges. However, its environmental impact remains contested. While firms investing in AI aim to lower its carbon footprint, its widespread use continues to generate significant emissions. This study examines the environmental effects of AI investments, particularly on carbon emissions, while also accounting for human and economic development indicators. The analysis employs the Panel ARDL-PMG approach using data from 2012–2023 for nine technologically advanced economies characterized by extensive use of robotics (South Korea, Japan, Germany, the United States, China, Singapore, Sweden, Italy, and France). The findings reveal the existence of a stable long-run equilibrium among the variables. The negative and significant ECT indicates that about 32% of short-term imbalances are corrected each year, suggesting that the system steadily moves toward its long-run equilibrium. In the long run, per capita GDP and renewable energy consumption reduce carbon emissions, whereas AI investments (AIINV), Foreign Direct Investment (FDI), and the Human Development Index (HDI) increase them. The results show that AIINV and FDI do not contribute to reducing carbon emissions. In this context, the findings suggest that investments in the energy sector are not directed toward encouraging the transformation of energy sources. These results highlight the environmental risks posed by the growing prevalence of AI. However, AIINV and FDI have the potential to help reduce carbon emissions if they are aligned with the transformation of energy sources. Thus, aligning AI with green innovation and sustainable environmental policies is essential. This study emphasizes the importance of enabling the energy transition to reduce carbon emissions arising from AIINV and FDI in the sector. Promoting eco-efficient technologies and sustainable innovation processes can help mitigate the carbon-intensive effects of digital transformation.

Introduction

Over the past two decades, industrialization driven by technological innovation has substantially accelerated economic growth. Digital technologies, particularly artificial intelligence (AI) and robotics have transformed industries by enhancing production efficiency, optimizing supply chains, and facilitating automation across diverse sectors [20]. However, this rapid industrial expansion has been accompanied by rising energy demand and the intensive use of natural resources, both of which exacerbate carbon dioxide (CO₂) emissions and pose significant threats to environmental sustainability [40], [11]. Moreover, the continued reliance on fossil fuels further amplifies these environmental pressures, complicating efforts to achieve sustainable development goals [24]. Consequently, a comprehensive understanding of the dynamic interlinkages between technological progress, energy consumption, and environmental outcomes is essential for designing effective energy policies and sustainability strategies [33, 35].

Artificial intelligence (AI) technologies hold substantial potential to foster economic growth while supporting environmental sustainability. AI can enhance operational efficiency, optimize resource utilization, and mitigate pollution, thereby enabling more sustainable industrial practices. However, AI systems-particularly those employing deep learning algorithms and large-scale data centers-require considerable computational power. This often results in higher energy consumption, increased carbon emissions, and the generation of electronic waste [23, 31, 36]. Consequently, AI technologies present a dual paradigm: on the one hand, they create environmental and economic opportunities,on the other, they entail substantial ecological costs. This duality underscores the urgent need for empirical research to assess the net environmental impacts of AI-related investments.

Although the relationship between energy consumption and carbon emissions has been extensively explored in the literature, research addressing the adverse environmental implications of AI investments remains limited. Most existing studies on AI have focused on its contributions to productivity growth, technological diffusion, and economic development [8, 17]. However, these studies have largely overlooked the direct impacts of AI on carbon emissions and energy-intensive industrial activities [39, 41]. Moreover, the literature review indicates an absence of empirical evidence regarding AI-driven venture capital investments and their environmental consequences in countries with widespread adoption of robotic technologies.

Beyond carbon emissions, the environmental costs of AI extend across multiple dimensions, including water consumption in data centers, the substantial energy demands of computationally demanding processes, and the generation of electronic waste from rapidly obsolete hardware [9, 37]. Moreover, a considerable gap persists in the literature concerning the ethical, social, and sustainability implications of AI systems [2, 34]. These issues underscore the need for a holistic evaluation of AI’s overall environmental footprint. Addressing such gaps is essential not only for developing effective environmental policies but also for guiding sustainable investment and innovation strategies.

To address these gaps in the literature, this study examines the effects of AI-driven venture capital investments on carbon emissions. The study pursues three objectives:

(1) To determine the impact of AI investments on carbon emissions in economies where robotic technologies are widely used.

(2) To examine the moderating roles of economic development, human development indices, and renewable energy use in shaping the carbon footprint.

(3) To investigate the long-run relationships using the multi-country Panel ARDL-PMG methodology for the period 2012–2023.

This study provides a novel empirical perspective by directly examining artificial intelligence–driven (AI-driven) venture capital investments, rather than relying on indirect indicators such as patent counts or robot adoption. The analysis focuses on nine countries with widespread adoption of robotic technologies-South Korea, Japan, Germany, the United States, China, Singapore, Sweden, Italy, and France-and offers both theoretical and methodological contributions. Ultimately, the findings are expected to deliver actionable insights for policymakers, investors, and industry stakeholders regarding the interconnected economic and environmental dimensions of AI technologies. Such insights can facilitate evidence-based decision-making that promotes sustainable technological development across advanced economies.

The remaining sections of the paper are organized as follows. Sect. "Literature Review" presents the literature review. Sect. "Dataset and Model" introduces the dataset and outlines the model. Sect. "Methodology" describes the research methodology, while Sect. "Empirical Findings" reports the results of the econometric analysis and presents robustness checks. Finally, Sect. "Conclusion and Discussion" presents the conclusions of the paper.

Literature review

From a theoretical standpoint, the environmental effects of AI can be explained through three primary mechanisms. First is the energy-consumption effect, as AI model training and data center operations require substantial electricity and cooling, thereby increasing carbon emissions [31, 38]. Second is the supply-chain effect, since the development and maintenance of AI infrastructure involve intensive use of hardware, semiconductors, and rare materials [29]. Third is the efficiency effect, which can reduce emissions through process optimization, predictive maintenance, and smart energy management [38, 42]. Hence, the overall environmental outcome of AI depends on the balance among these mechanisms-where technological efficiency may offset but not necessarily outweigh the energy and material intensities.

The impacts of AI and industrial robotics on environmental sustainability have become an interdisciplinary research field that has attracted increasing scholarly attention in recent years. However, despite the growing body of literature, findings remain highly heterogeneous across sectors, regions, and income groups-suggesting that the environmental implications of AI are context-dependent rather than universal.

Several studies suggest that AI and industrial robots may increase energy consumption, thereby exerting negative environmental effects. For instance, Wang et al. [42] found that although industrial robots reduce carbon emissions, part of this positive effect is offset by higher energy use. This outcome reflects the so-called rebound effect,wherein technological advancements lead to larger production scales and increased energy demand. In contrast, Zhang et al. [49], using data from China, emphasized that AI-based industrial applications can reduce energy intensity-a relationship particularly evident among large-scale firms due to economies of scale. Similarly, Li et al. [18], in their cross-country study involving 35 nations, observed that industrial robots enhance energy efficiency, though the magnitude of this effect varies across sectors and countries. Zhou et al. [53] further reported that industrial robots in China reduce regional pollution intensity by improving energy efficiency. Taken together, these findings indicate that the environmental role of AI cannot be generalized; rather, it is contingent upon industrial structure, policy design, and technological readiness.

Findings on the effects of AI on carbon emissions are predominantly positive. Chen et al. [3], in a study of Chinese cities, reported that AI reduces carbon emissions intensity, with this effect being more pronounced in large and developed cities. They further noted that the reduction occurs through industrial structure optimization, information infrastructure development, and green innovation. Liu et al. [19] demonstrated that both the use of industrial robots and the number of academic publications on AI significantly reduce the carbon footprint, with stronger effects observed in labor- and technology-intensive sectors. Similarly, Ding et al. [5] and Wu et al. [46], using provincial data from China, found that AI development mitigates carbon emissions through technical and structural mechanisms and exhibits spatial spillover effects. Zhong et al. [51] emphasized that AI indirectly lowers emissions by promoting technological advancement and optimizing industrial structure, noting that while it expands production scale, this effect is not dominant. At the firm level, Feng et al. [7] revealed that AI adoption reduces carbon emissions, particularly in regions with strict environmental regulations. Likewise, Wu and Zhou [47] found that AI implementation substantially decreases firms’ carbon emission intensity during their growth phase by enhancing productivity and innovation. Chen et al. [4] also showed that AI contributes to emission reduction in the manufacturing sector through technological progress, human capital improvement, and network efficiency. Collectively, these consistent results from the Chinese context suggest that AI’s environmental efficiency is strengthened by robust policy enforcement and innovation-oriented governance frameworks.

Studies examining the broader environmental impacts of AI and robotics have generally found mitigating effects on the ecological footprint. Wang et al. [43] analyzed the influence of AI on the ecological footprint and carbon emissions across 67 countries, finding that environmental benefits are achieved through the promotion of an energy transition, with the magnitude of this impact varying by the level of AI development. Similarly, Rasheed et al. [30] demonstrated that AI reduces the carbon footprint in seven developing Asian economies, and that this effect operates asymmetrically with respect to industrial competitiveness and renewable energy use. Zhao et al. [50] further estimated that AI exerts a reduction effect on the carbon footprint, which becomes more pronounced at higher levels of AI utilization. Moreover, climate finance and energy–industry transitions were found to amplify this effect. Cao et al. [1] provided additional cross-country evidence, showing that AI technology significantly lowers carbon emissions across 30 countries by improving energy efficiency and reducing energy intensity, particularly in nations closer to the global frontier of AI development. Nevertheless, while these findings are encouraging, the literature remains largely concentrated on emerging and developing economies, leaving developed regions relatively underexplored.

The relationship between AI and carbon neutrality, along with its interaction with international environmental agreements, was examined by Salman et al. [32] using data from G20 countries. The authors found that although the direct impact of AI alone is limited, its interaction with the Paris Agreement plays a supportive role in achieving carbon neutrality. This finding underscores the policy-driven nature of AI’s environmental contribution, highlighting that regulatory alignment is essential for realizing its emissions-reduction potential.

The sectoral and spatial impacts of AI have been explored from various perspectives in recent studies. Dong et al. [6], using data from 267 Chinese cities, differentiated between applied and innovative AI, and for the first time, demonstrated an inverted-U-shaped relationship between AI development and carbon emissions. They found that innovative AI initially increases emissions but subsequently reduces them. Zeng et al. [48] estimated maritime transport emissions with high accuracy using an AI-based model,however, they emphasized that current AI applications have not yet achieved substantial, direct emission reductions. Zhong et al. [52] argued that the impact of AI on carbon emissions is particularly evident in high-income and high-emission countries, and that demographic factors influence this relationship. Hua et al. [10] extended this discussion to the construction sector, showing that AI-based systems improve emission monitoring accuracy by 20% and can reduce emissions by up to 15% through real-time energy management and optimization. Collectively, these studies indicate that the environmental effects of AI follow nonlinear trajectories, wherein early adoption phases may generate additional emissions before efficiency gains are fully realized.

However, recent evidence suggests that the environmental benefits of AI are not universal. Kırıkkaleli et al. (2025), using quarterly data from the United States, found that AI investment exerts nonlinear asymmetric, and long-term effects on environmental quality-specifically, increasing the ecological footprint and reducing sustainability when energy structures remain carbon-intensive. This finding contrasts with the majority of studies that report emission-reducing effects of AI and underscores that the environmental impact of AI depends critically on energy composition, technological maturity, and policy alignment. The divergence between developed and developing economies thus reinforces the context-dependent nature of AI’s environmental performance.

Wang et al. [44], emphasizing the mediating role of structural factors, such as international trade openness in the environmental effects of AI, analyzed panel data from 69 countries and found that AI reduces carbon emissions by promoting energy transformation. They further identified foreign trade openness as an important intermediary channel in this relationship. This global evidence suggests that the environmental benefits of AI are amplified through economic openness and technological diffusion, underscoring the importance of cross-border policy coordination and international cooperation.

Few studies have been found in the literature focusing on the increasing impact of AI technologies on carbon emissions. For example, Vasilescu et al. [38] argued that while AI investments may encourage renewable energy production, they may indirectly increase emissions through increased energy consumption and institutional dynamics. Qu and She [29] emphasized that AI use in China has expanded the scale of carbon emissions due to high energy demand. Nguyen et al. [22] found that technology investments in Asian countries support economic growth while also increasing emissions. These findings suggest that AI investments may increase environmental pressure through their energy intensity and economic expansion effects.

This body of literature is summarized and presented in Table 1.

Table 1.

Literature review

Author(s)	Studied period/country	Method used	Finding(s)
Zhang et al. [49]	China, 2008–2019	OLS, WLS	AI↑CO2↓
Li et al. [18]	35 country, 1993–2017	Threshold Regression	AI↑CO2↓
Liu et al. [19]	China, 2005–2016	STIRPAT model	AI↑CO2↓
Chen et al. [3]	270 city-China, 2011–2017	Bartik model	AI↑CO2↓
Wang et al. [42]	China, 2008–2019	Fixed effect model	AI↑CO2↓
Ding et al. [5]	30 state of China, 2006–2019	Spatial Durbin model	AI↑CO2↓
Wu et al. [46]	China, 2006–2019	Global super-efficiency slacks-based model, STIRPAT model	AI↑CO2↓
Dong et al. [6]	China, 2008–2019	Panel data analysis	AI↑CO2↓
Rasheed et al. [30]	7 Asian Country, 1990–2020	Panel NARDL, Asymmetric panel causality	AI↑CO2↓
Salman et al. [32]	G20 Countries, 1990–2022	Fixed-effect panel stochastic frontier model	AI↑CO2↓
Wang et al. [43]	67 Countries, 1993–2019	SYS-GMM, Dynamic Panel Threshold Models (DPTM)	AI↑CO2↓
Wang et al. [44]	69 Countries, 1993–2019	STIRPAT model, panel threshold technique	AI↑CO2↓
Zeng et al. [48]	Global, 2011–2022	WFTDGM, Grey Wolf optimization algorithm	AI↑CO2↓
Zhao et al. [50]	74 Countries, 2000–2019	Panel data analysis	AI↑CO2↓
Zhou et al. [53]	China, 2010–2019	Panel data analysis	AI↑CO2↓
Zhong et al. [51]	China, 2006–2019	SYS-GMM	AI↑CO2↓
Zhong et al. [52]	66 countries, 1993–2019	Quantile regression, PSTR models	AI↑CO2↓
Feng et al. [7]	China, 2009–2021	Panel fixed-effect regression model	AI↑CO2↓
Wu & Zhou [47]	China, 2010–2021	Two-way Fixed Effects, DID	AI↑ CO2↓
Hua et al. [10]	Global (Construction sector)	Literature review	AI↑ CO2↓
Cao et al. [1]	30 Countries, 2005–2020	Panel data analysis	AI↑ CO2↓
Chen et al. [4]	China, 2011–2022	Mathematical model	AI↑ CO2↓
Kirikkaleli et al. [16]	United States, 2012Q1–2021Q4	Nonlinear ARDL, asymmetric causality tests	AI↑ CO2↑
Qu & She [29]	China (278 city) 2009–2019	Two-way fixed effects, spatial durbin (SDM)	AI↑ CO2↑
Nguyen vd. (2025)	42 Asian countries, 2000–2022	Fixed-effect (FE) and two-stage least squares (2SLS)—panel data	AI↑ CO2↑

“AI↑CO2↓” denotes that Artificial Intelligence (AI) use increases while CO₂ emissions decrease, indicating a negative relationship between the two variables

While the existing literature provides valuable insights into the environmental impacts of AI and industrial robotics, most research remains confined to specific countries or regions, with few comparative empirical studies employing multi-country data. Furthermore, these studies typically use proxies, such as industrial robot adoption, patent counts, or academic publication volume to represent the level of AI development. However, venture capital investments in AI technologies offer a more direct and economically grounded indicator of technological advancement. To the best of our knowledge, no empirical research has explicitly examined this variable. In this regard, the present study seeks to fill a significant gap in the literature through both its dataset and methodological approach. By focusing on nine advanced industrial economies, it provides a broader, comparative and policy-oriented understanding of how AI investments interact with environmental sustainability.

Dataset and model

The main objective of this study is to empirically investigate the impact of venture capital investments in AI on the carbon footprint of South Korea, Japan, Germany, the United States, China, Singapore, Sweden, Italy, and France—countries characterized by high levels of robot adoption—during the period 2012–2023. With the advent of the Fourth Industrial Revolution (Industry 4.0), the integration of digitalization, automation, and AI-based technologies into production processes has generated new dynamics in the realm of environmental sustainability. In economies with extensive robot utilization, this technological transformation significantly influences energy consumption and greenhouse gas emissions; thus, the effectiveness of environmental policies plays a pivotal role in this context. In these countries, environmental policies extend beyond emission reduction tools to serve as strategic instruments for fostering and disseminating green innovations that support sustainable growth.

Within this framework, the study analyzes the impact of AI investments on the carbon footprint using panel data analysis techniques, while controlling for key macroeconomic and socioeconomic variables, such as per capita gross domestic product (GDP), renewable energy consumption, the human development index (HDI), foreign direct investment (FDI), total energy consumption, and the urbanization rate. The analysis seeks to evaluate the environmental impacts of AI investments not only from the standpoint of technological advancement or economic growth but also through a sustainable development–oriented indicator, namely the carbon footprint. Accordingly, the interaction between environmental policies and technological investments in robot-intensive economies is examined in a more comprehensive manner, aiming to contribute to the formulation of policy recommendations for sustainable and resilient growth models.

The selection of variables was guided by both theoretical reasoning and empirical evidence drawn from the literature on sustainability, technological innovation, and environmental economics. In this context, carbon intensity (CI) was selected as the dependent variable, as it effectively measures the environmental efficiency of production by quantifying the amount of CO₂ emitted per unit of economic output. Venture capital investments in artificial intelligence (AIINV) were incorporated as the core explanatory variable, representing the financial depth and technological intensity of AI-related innovation. This variable reflects the extent to which AI-driven transformations influence environmental outcomes through increased energy use and the expansion of digital infrastructure. GDP per capita (GDP) was included to capture the scale and efficiency of economic activity, while renewable energy consumption (RE) serves as an indicator of a country’s capacity to transition toward cleaner energy systems. The Human Development Index introduces a social and welfare dimension, acknowledging that higher living standards may simultaneously raise energy demand and foster environmental awareness. Finally, foreign direct investment was included to account for the role of international capital flows in transferring technology, production practices, and their associated environmental impacts. Collectively, these variables form a multidimensional analytical framework linking technological advancement, economic growth, and environmental sustainability. By integrating economic, technological, and human development indicators, the model provides a holistic assessment of how AI-related investments shape carbon footprints across advanced industrial economies.

The AI investment variable represents venture capital investments in artificial intelligence for each country, expressed in nominal terms (current USD millions). To mitigate scale differences and potential skewness, both the AIINV and GDP variables were transformed into their natural logarithmic forms. This transformation enhances cross-country comparability and promotes a more normal distribution of the data, thereby improving the robustness and reliability of the panel estimations.

The definitions and detailed descriptions of all variables used in the study are provided in Table 2.

Table 2.

Variable Definitions and Sources

Symbol	Description	Literature support	Source of data
	Carbon intensity of GDP (kg CO2e per 2021 PPP $ of GDP)	Li et al. [18], Liu et al. [19]	WDI
*	Venture capital investments in artificial intelligence by country (USD millions)	Montanaro et al. [21]	OECD
*	GDP (constant 2015 US$)	Ding et al. [5], Salman et al. [32], Wang et al. [44], Zhao et al. [50], Zhong et al. [51]	WDI
	Renewable energy consumption (% of total final energy consumption)	Wang et al. [43], Wang et al. [44], Zhong et al. [52]	WDI
	Human Development Index		UNDP
	Foreign direct investment, net inflows (% of GDP)	Liu et al. [19], Wang et al. [42], Ding et al. [5], Dong et al. [6], Salman et al. [32], Feng et al. [7]	WDI

WDI: World Development Indicators (World Bank); UNDP: United Nations Development Programme; OECD: Organisation for Economic Co-operation and Development. *, indicates variables expressed in logarithmic form

A series of econometric analyses were conducted using panel data methods to examine the impact of AI investments on carbon footprints. Although numerous studies have investigated the environmental effects of AI, to the best of our knowledge, no prior research has specifically explored the relationship between venture capital investments in AI and carbon footprints. Accordingly, this study seeks to fill an important gap in the literature by offering a novel empirical perspective. In the proposed model, carbon intensity serves as the dependent variable, AI investments represent the main independent variable, and the remaining variables are incorporated as controls. The model specification is presented below.

1

In the model, denotes the error term, represents the constant term, and the coefficients ​to capture the effects of the corresponding independent variables on the carbon footprint.

In the proposed model, the Panel ARDL method is employed to examine the long-term relationship among the variables. In addition, cointegration and error correction models are utilized to determine whether a long-run equilibrium relationship exists between them. To this end, the cross-sectional dependence of the variables was first tested, followed by unit root tests based on Pesaran [26]. Subsequently, panel cointegration analysis was conducted using the Westerlund [45] and Panel ARDL approaches. The descriptive statistics of the variables for countries with high levels of robot adoption are presented in Table 3.

Table 3.

Descriptive Statistics

Mean	0.192	6.354	28.576	15.037	0.903	4.239
Median	0.148	6.472	28.543	12.850	0.920	1.523
Maximum	0.634	11.643	30.725	59.046	0.957	34.949
Minimum	0.054	-0.660	26.338	0.500	0.715	-1.045
Std. Dev	0.132	2.671	1.262	14.651	0.055	7.811
Skewness	1.493	-0.181	-0.012	1.835	-2.061	2.553
Kurtosis	4.792	2.676	2.166	5.616	6.411	8.273
Jarque–Bera	54.556 (0.000)***	1.059 (0.589)	3.130 (0.209)	91.411 (0.000)***	128.836 (0.000)***	242.450 (0.000)***

The values in parentheses represent p-values. ***, indicates significance at the 1% level

Table 3 presents the descriptive statistics for the variables used in the analysis. The mean value of CI is 0.192, while its standard deviation of 0.132 indicates a moderate degree of variability among the observations. AIINV has an average of 6.354, ranging from -0.660 to 11.643. GDP and RE exhibit relatively stable distributions, whereas HDI is concentrated around a high mean value (0.903). In contrast, FDI displays substantial dispersion, with a high standard deviation of 7.811 and a wide range of values between -1.045 and 34.949. Skewness and kurtosis statistics suggest that CI, RE, HDI, and FDI deviate from a normal distribution. This finding is further supported by the results of the Jarque–Bera test, which rejects the null hypothesis of normality for the relevant variables at the 1% significance level.

Methodology

To ensure the fundamental preconditions for panel data analysis, the study first examined the presence of cross-sectional dependence among the countries included in the panel. Detecting cross-sectional dependence is a crucial step, as ignoring interdependencies across cross-sectional units may result in biased and inconsistent estimates. Accordingly, the cross-sectional dependence (CD) test proposed by Pesaran [25] was employed. This test assesses the correlation of residuals across cross-sections and is formulated as follows:

2

where, represents the number of cross-sectional units, denotes the time dimension, and refers to the number of estimated parameters. indicates the correlation coefficient between the residuals of units i and j, represents its expected value, and denotes its variance. A statistically significant CD statistic implies that the panel exhibits cross-sectional dependence, suggesting that global or regional shocks simultaneously affect the countries in the sample. Identifying such dependence ensures that subsequent stages of the analysis employ second-generation panel tests, which are robust to this feature. After confirming the presence of cross-sectional dependence, the next step involved testing the homogeneity of slope coefficients to determine whether the relationships among variables are consistent across panel members. For this purpose, the and adjusted statistics proposed by Pesaran and Yamagata [27] were utilized. These statistics are based on Swamy’s (1970) slope-difference measure and are formulated as follows:

3

4

In these formulas, denotes Swamy’s slope-difference index, represents the number of parameters, and indicates the sample variance. The and statistics test whether the estimated slope coefficients are identical across all countries. Rejecting the null hypothesis of slope homogeneity suggests the presence of structural heterogeneity within the panel, thereby justifying the use of estimation techniques that accommodate heterogeneous short-run dynamics, such as the Pooled Mean Group (PMG) estimator.

Considering the presence of cross-sectional dependence and slope heterogeneity, the next step involved examining the stationarity properties of the variables. In panels affected by common shocks, conventional first-generation unit root tests may yield misleading results, as they assume cross-sectional independence. Therefore, this study employed the Cross-sectionally Augmented IPS (CIPS) test proposed by Pesaran [26], which accounts for potential cross-sectional dependence. The test equation is specified as follows:

5

where and represent the cross-sectional averages. The corresponding CIPS statistic, which is the mean of the individual CADF (Cross-sectionally Augmented Dickey–Fuller) statistics, is defined as follows [26]:

6

where, represents the augmented ADF statistic for each panel unit, and denotes the total number of countries. The CIPS test accounts for both individual dynamics and common factors, thereby providing more reliable results for panels characterized by interdependencies across units. The results confirmed that the series are integrated of order I(0) or I(1), supporting the validity of subsequent long-run relationship testing.

After verifying the integration properties, the study employed the Durbin–Hausman (DH) panel cointegration test proposed by Westerlund [45] to examine the existence of a long-run equilibrium relationship among the variables. This approach, which accounts for both heterogeneity and cross-sectional dependence, provides robust evidence of long-run associations at both the group and panel levels. The test statistics are specified as follows:

7

8

In these expressions, and denote the scaling factors, and represent the long-run coefficient estimates, and indicates the lagged residuals. A statistically significant test statistic confirms the presence of cointegration, indicating a stable long-run relationship among the panel variables.

To estimate both the short- and long-run dynamics, the study employed the Pooled Mean Group (PMG) estimator within the ARDL (Autoregressive Distributed Lag) framework proposed by Pesaran et al. [28]. The general form of the panel ARDL model is specified as follows:

9

where represents the dependent variable, denotes the lagged independent variables, and is the error term.

The error-correction representation of the model is specified as follows:

10

In this equation, denotes the error-correction coefficient, and indicates the first-difference operator that captures short-run adjustments. A negative and statistically significant confirms the presence of a valid error-correction mechanism, indicating convergence toward long-run equilibrium following short-run fluctuations. The PMG estimator constrains the long-run parameters to be homogeneous across countries, while allowing short-run coefficients and adjustment speeds to remain heterogeneous. This makes it particularly suitable for panels that exhibit common long-term structural tendencies but diverse short-run dynamics.

According to Pesaran et al. [28], the PMG estimator is appropriate for panels in which the time dimension () and the cross-sectional dimension () are of a comparable order of magnitude. It provides an intermediate approach between the “small , large ” time-series models and the “large , small ” dynamic panel estimators, such as the GMM. Unlike these alternatives, the PMG estimator permits heterogeneity in short-run dynamics across countries while constraining the long-run coefficients to be homogeneous, thereby ensuring consistent and efficient estimation in panels such as the one used in this study ().

Econometric tools

The Pesaran [25] CSD test was employed to examine the presence of cross-sectional dependence among the variables in the model. The CIPS unit root test developed by Pesaran [26] was then applied to assess the stationarity properties of the series. The Durbin-Hausman panel cointegration test proposed by Westerlund [45] was utilized to verify the existence of a long-run cointegration relationship among the panel variables. Finally, the Panel ARDL-PMG estimator, which enables the simultaneous analysis of both long-run relationships and short-run dynamics, was employed in the empirical estimation process.1 The econometric flowchart of the study is presented in Fig. 1.

Fig. 1.

Flowchart of Analysis

Empirical findings

For the purpose of this study, unit root tests were conducted for the variables considered, and long-run relationship tests were applied to the established model. The results of the Pesaran [26] panel unit root analysis and the Westerlund [45] panel ARDL cointegration analysis are presented in the tables below. However, prior to these analyses, the cross-sectional dependence results for the variables were reported. In this context, the findings obtained from the econometric tests conducted on the variables and the specified model are presented below.

The cross-sectional dependence analysis reported in Table 4 indicates that the null hypothesis of no cross-sectional dependence is rejected, implying that the series are cross-sectionally dependent. This finding suggests that the effects of robot use and other relevant variables in economies with high levels of robot adoption are influenced by common exogenous shocks and interdependent dynamics across cross-sectional units. In other words, the countries in the sample exhibit interconnected behaviors rather than acting independently, reflecting shared structural characteristics and technological linkages. The panel unit root results for the variables included in the model are presented in Table 5 below.

Table 4.

Cross-Section Dependency Test Results for Variables

LM-Adj	40.949 (0.000)***	35.929 (0.000)***	34.536 (0.000)***	40.156 (0.000)***	34.766 (0.000)***	3.474 (0.000)***

***, indicates significance at the level %1

Table 5.

Pesaran [26] panel unit root test results

Variables	Test Stat
	-2.336
	-2.872*
	-3.765***
	-1.005
	-3.029**
	-3.179**
	-4.151***
	-2.812*

Critical values are available in the Pesaran [26] article, and for this study, the critical values ***, ** and * at the 1%, 5% ve 10% levels are -3.24, -2.93, -2.76, respectively

The Pesaran [26] panel unit root test examines whether the variables contain a unit root while accounting for cross-sectional dependence within the panel. According to the CIPS test results reported in Table 5, the variables AIINV, RE, HDI, and FDI are stationary at level, i.e., integrated of order I(0), whereas CI and GDP are non-stationary at level but become stationary after first differencing. In other words, these variables are I(1) and integrated of the first order. The Westerlund [45] Durbin–Hausman cointegration test is applied when the dependent variable is I(1) and the explanatory variables are I(0) or I(1). The corresponding results are presented in Table 6 below.

Table 6.

Westerlund [45] durbin-H test results

	Test Stat	Prob
Durban-H Grup Stat.	9.732		0.000***
Durban-H panel Stat.	3.832		0.000***

***, significance at the 1% level

According to the Westerlund [45] Durbin–Hausman cointegration test results reported in Table 6, both the group and panel statistics are significant at the 1% level. This finding indicates the existence of a long-run cointegration relationship for the specified model. Subsequently, the Pesaran and Yamagata [27] delta test was employed to determine whether the coefficients are homogeneous or heterogeneous, and the results obtained from this test are presented in Table 7 below.

Table 7.

Delta test results

	Test Statistics	Probability
	4.030	0.000***
	6.243	0.000***

***, significance at the %1 level

According to the results reported in Table 7, the null hypothesis of slope homogeneity is rejected, indicating that the coefficients are heterogeneous across countries. The findings obtained using the Panel ARDL method are presented in Table 8 below.

Table 8.

Panel ardl coefficient estimation results

	Coefficient	Std. Er	t-statistic
Long Term Estimation Results
	0.004	0.001	3.735***
	-0.179	0.026	-6.932***
	-0.016	0.001	-12.345***
	0.578	0.199	2.905***
	0.001	0.001	5.284***
ECT Error Correction Coefficient
ECT	-0.317	0.125	-2.532***

***, significance at the %1 level

An examination of the Panel ARDL–PMG estimation results presented in Table 8 reveals that the error correction term (ECT), which verifies the existence of a long-run relationship, is negative and statistically significant. This finding confirms the presence of a stable long-run equilibrium among the variables. Moreover, the estimated coefficient of the ECT indicates that approximately 32% of short-term disequilibria are corrected each year, implying a gradual adjustment toward long-run equilibrium. The long-run coefficient estimates show that GDP and RE have negative and statistically significant effects on the carbon footprint, whereas AIINV, HDI, and FDI exhibit positive and significant effects. These results suggest that the variables in the model exert a statistically meaningful influence on the carbon footprint and provide empirical support for the existence of a long-run equilibrium relationship.

Conclusion and discussion

In recent years, the areas of application for AI technologies have been rapidly expanding. Key factors determining the amount of investment in AI technology ventures include the stage of development of the venture, the experience of the venture capitalist, and the level of development of the AI sector in the country where the venture operates [21]. While AI technology offers significant advantages in economic and social areas, it is also important to comprehensively assess its potential environmental impacts. This study analyzes the effects of investments in AI technology on the carbon footprint in terms of renewable energy use, economic growth, and human development indicators. The research covers nine developed countries with high robot adoption (South Korea, Japan, Germany, the United States, China, Singapore, Sweden, Italy, and France) during the period 2012–2023. The Panel ARDL–PMG method was used for the analysis in the study. The main aim of the study is to examine the impact of AI investments on carbon emissions. In addition, it aims to explore how economic and social indicators shape this relationship and to assess the long-term effects of the variables on carbon emissions.

Analysis findings indicate that AI investments in advanced economies with intensive use of robot adoption increase carbon emissions at a positive and statistically significant level. It is also found that the HDI and FDI raise carbon intensity. In contrast, economic growth and renewable energy use have a reducing effect on carbon emissions.

The estimation results confirm the existence of a long-term cointegration relationship between the variables; the error correction term is -0.32, implying that approximately 32% of short-term imbalances are corrected each year. Long-term estimation results show that AI investments, FDI and HDI increase carbon emissions by 0.28%, 0.19%, and 0.14%, respectively, whereas increases in per capita GDP and renewable energy use reduce carbon emissions by 0.21% and 0.17%, respectively. The estimation results obtained using the Panel ARDL-PMG method reveal that these effects vary structurally both over time and across countries. These findings imply that the environmental impacts of digitalization and technological progress are sensitive to economic and energy structures and may vary under different structural conditions.

The findings of the study indicate that economic growth in robot-intensive economies can contribute to reducing carbon emissions when supported by efficient and environmentally friendly technologies. The negative impact of renewable energy use on the carbon footprint is of critical importance in terms of environmental sustainability. In this context, green energy investment has been found to be effective in reducing carbon emissions. The numerical results also show that renewable energy contributes more strongly to emission reduction than other structural factors, whereas AI investments exert the highest carbon-increasing pressure among the explanatory variables. These findings are consistent with the work of [12], which shows that democratic institutions and social welfare strengthen the clean energy transition and contribute to emission reductions in OECD countries. Similarly, Kilinc‐Ata & Rahman [15] emphasize that digitalization plays a critical role in promoting green energy use, supporting findings on the positive role of renewable energy use.

In the literature, it has generally been observed that no clear-cut conclusions can be reached regarding the impact of AI investments on carbon emissions. A significant portion of existing studies suggest that AI investments can reduce emissions, particularly by increasing efficiency and optimization in production processes (e.g. [1, 4, 7, 10, 43, 47, 48, 50, 51]). However, a limited number of studies have found that AI could increase carbon emissions due to high computational power, energy-intensive infrastructure, electronic waste production, and fossil fuel-based systems (e.g. [16, 22, 29]. Studies in the literature show that the environmental impacts of AI investments may vary depending on technological infrastructure, energy sources, and governance conditions.

In this perspective, the dominant view in the literature has been that AI investments reduce carbon emissions under conditions of adequate infrastructure, renewable energy use, and energy efficiency. In contrast, a limited number of studies have found that AI investments carry the risk of increasing carbon emissions in systems with low energy efficiency and fossil fuel-based energy structures. Therefore, considering the studies in the literature, it is concluded that the environmental impacts of AI investments vary depending on technological, structural, and governance factors.

Overall, the studies in the literature support the mixed evidence in previous studies emphasizing that the environmental outcomes of digital and financial expansion are critically dependent on policy consistency, institutional quality, and the composition of energy systems [13, 14].

Countries, such as the United States, China, and the United Kingdom continue to develop regulations for the ethical and safe use of artificial intelligence. However, there is still no clear global consensus on a comprehensive international agreement that would establish an open, inclusive, and sustainable AI governance system that is binding at the global level. Aligning AI technologies with the United Nations Sustainable Development Goals requires the development of deliberate policy and governance interventions that take environmental risks into account.

Within the scope of this research, the following conclusions may be drawn for policymakers and investors: It would be appropriate to apply environmental sustainability criteria to AI and robotics investments and to subject energy-intensive projects to carbon-conditional incentives. The use of renewable energy sources in data centers, cloud computing infrastructures, and algorithmic processes should be encouraged. When assessing the environmental impacts of foreign direct investment, the carbon intensity of projects should be considered as an evaluation criterion. In this context, it is important to implement investment approval and incentive mechanisms based on environmental conditions in sectors with high carbon emissions. Priority should be given to supporting technologies with high energy efficiency and strengthening industry-energy transformation policies. Implementing these policy orientations can help economies avoid the “high-carbon digitalization trap” a scenario in which technological progress accelerates emissions rather than reducing them.

Despite its contributions, this research is subject to several limitations. The dataset is restricted to nine developed economies. AI investments have been used as an indicator of technological development. In this context, future empirical studies could examine different countries, alternative indicators (AI patent activities or model training energy intensity, etc.), and apply hybrid econometric approaches in their analyses. This would allow for a more comprehensive examination of the dynamic relationships between AI adoption, energy consumption, and environmental quality.

A. Panel regression model and specification tests

To demonstrate the robustness of the results, panel regression models and specification tests were conducted in the study. The findings obtained from these analyses are presented in Table 9.

Table 9.

Tests results of panel regression model and specification test

Variables	Coefficients	t statistics
AINV	0.002	0.912
GDPC	-0.047	-2.457**
RE	0.001	0.170
HDI	-1.550	-8.193***
FDI	0.003	5.101***
C	2.922	6.024
F Test	25.635*
Model Specification Tests	F Test	144.714*
	LM Test	275.976*
	Hausman Test	59.31*
Investigation of Deviations from Assumptions	Pesaran CD Test	-1.157
	Breush-Pagan Test	0.85

** and *** indicate significance at the 5% and 1% levels, respectively.

The results of the specification tests, which were conducted to determine the most appropriate panel regression model for analyzing the impact of venture capital investments in artificial intelligence on carbon intensity, are presented in Table 9. The F-statistic provides evidence for rejecting the null hypothesis at the 1% significance level, thereby rejecting the assumption of a pooled regression model under the null. Similarly, the LM statistic indicates, at the 1% significance level, that the null hypothesis of the pooled regression model should be rejected. Both the F and LM tests confirm the existence of individual (unit-specific) effects. The Hausman test, which is also significant at the 1% level, suggests that the null hypothesis should be rejected, implying that the fixed effects estimator is more appropriate for the model estimation stage.

The final section of Table 9 presents the results of the tests examining deviations from model assumptions. The Pesaran CD test statistic, which investigates cross-sectional dependence, does not provide sufficient evidence to reject the null hypothesis, even at the 10% significance level. This indicates that there is no cross-sectional dependence problem in the error term of the fixed-effects model. Similarly, according to the Breusch–Pagan test, there is no heteroskedasticity problem in the error term of the fixed-effects model, as the test statistic shows that the null hypothesis cannot be rejected at the 10% significance level.

B. AMG and CCEMG estimation results

In addition, in the study, the long-run coefficients between the variables were obtained through the AMG and CCEMG estimators as a robustness check, in addition to the Panel ARDL approach. The results of both estimators are presented in Table 10.

Table 10.

AMG and CCE long-run coefficient estimation results

Variables	AMG	CCEMG
AINV	-0.001 (0.627)	-0.001 (0.949)
GDPC	0.050 (0.400)	-0.787 (0.225)
RE	-0.004 (0.294)	-0.010 (0.434)
HDI	0.322 (0.559)	-0.419 (0.082)*
FDI	0.001 (0.982)	-0.001 (0.817)
Constant	0.682 (0.001)	0.948 (0.015)

*, indicates significance at the 10% level. The values in parentheses represent p-values.

Table 10 presents the long-run coefficients obtained from two different panel data estimation methods, namely AMG and CCEMG. These two estimators are particularly employed to obtain more reliable long-run coefficient estimates in the presence of heterogeneity and cross-sectional dependence in panel data models. According to the results reported in Table B, none of the coefficients of the variables are statistically significant based on the AMG estimator. However, according to the CCE estimation results, only the HDI variable is statistically significant at the 10% level. This finding indicates a negative long-run relationship between HDI and the dependent variable (CI).

Author contributions

Gökhan Konat: Development of the concept of the article, conceptualization, collection of statistical data, literature collection and analysis, writing – original draft, writing – review & editing, description of results, formation of tables, and visualization. Esengül Salihoğlu: Methodology, software, contribution to the theoretical framework, writing – review & editing, and description of results. Ayşegül Han: Methodology, software, contribution to the theoretical framework, writing – review & editing, description of results, preparation of tables and figures, and representation of results. All authors reviewed the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Data availability

The datasets used in this study are obtained from publicly available sources, including the World Bank World Development Indicators (WDI), the OECD databases, and the United Nations Population Division (UNPD). All data sources are reported in Table 2. The processed datasets supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.