Do artificial intelligence investment, energy transition, and agriculture productivity affect carbon footprint? Evidence from panel data analysis of BRICS countries Open Access

Global climate change is one of the most pressing issues of our time, with carbon emissions from human activities—such as technological advancement, health, industrial output, agricultural productivity, and energy use—playing a significant role. Technological progress has been a major driver of economic growth, but it also impacts CO₂ emissions from production processes. According to Li and Wang (2017), technological advancement can help reduce carbon intensity by improving industrial and energy structures and economic development patterns.

However, remarkable technological advancements have led to a decrease in energy consumption, resulting in lower energy prices and surpluses. This, in turn, can spur increased energy use and further accelerate energy transitions, diminishing the intended energy-saving benefits of technology—known as the “Rebound Effect” (Wei & Liu, 2017). The rise in global CO₂ emissions since the Industrial Revolution is undeniable (Li & Wang, 2017). For instance, in North Africa, Jebli and Youssef (2017) and Liu, Zhang, and Bae (2017a) suggest that while increased agricultural value-added reduces CO₂ emissions, economic growth and renewable energy consumption can raise them. This relationship between CO₂ emissions and agriculture is bidirectional, emphasising the complex links between energy use, industrial growth, and environmental impacts.

To combat these issues, both developing and developed nations have implemented policies to reduce carbon emissions, promote automation-based production processes, conserve natural resources, and accelerate energy transitions. The Paris Climate Agreement (PCCA) is a prime example, with nations pledging to lower carbon emissions and shift away from fossil fuels to renewable energy (Dong et al., 2022a). This global commitment underscores the importance of balancing energy demand with environmental sustainability.

In this context, artificial intelligence (AI), automation, and robotics are revolutionising various industries, including manufacturing, energy, agriculture, healthcare, logistics, and finance (Zavyalova, Volokhina, Troyanskaya, & Dubova, 2023). As of 2023, the global AI market is valued at $196.63 billion and is projected to expand at a 36.6% annual growth rate through 2030, highlighting AI’s widespread impact across sectors. AI offers unprecedented opportunities for environmental protection, but it also presents challenges, particularly in terms of energy consumption. As AI capabilities grow, so does the energy required for AI training—Probst (2023) notes that training a model like ChatGPT consumes 1.287 gigawatt-hours of electricity, the same amount used annually by 120 American households.

Despite these concerns, AI has the potential to improve environmental outcomes. Studies have shown that AI can enhance renewable energy grid capacity (Nair, Nair, & Thakur, 2022), estimate energy consumption in mining (El Maghraoui, Ledmaoui, Laayati, El Hadraoui, & Chebak, 2022), and optimise building energy control to reduce waste (Kaligambe, Fujita, & Keisuke, 2022). While AI’s lifecycle energy consumption (first-order effect) and productivity rebound (third-order effect) may exacerbate environmental problems (Taddeo, Tsamados, Cowls, & Floridi, 2021; Wu, Shirkey, Celik, Shao, & Chen, 2022), AI’s ability to streamline production processes and support energy transition efforts can help mitigate these challenges (Lei, Liang, & Ruan, 2023; Lyu & Liu, 2021).

AI is also considered a general-purpose technology, much like the steam engine, electricity, and computers, driving productivity and sparking complementary innovations (Brynjolfsson, Rock, & Syverson, 2019). The McKinsey Global Institute (2018) estimates that AI could boost global GDP growth by 1.2% annually by 2030, significantly outpacing the growth effects of prior technological revolutions.

In the BRICS nations—Brazil, Russia, India, China, and South Africa—representing 46.39% of global greenhouse gas (GHG) emissions, the role of AI in environmental outcomes is of critical importance. These countries are major contributors to global CO₂ emissions, with China alone responsible for 30.1% of global emissions. As the largest emitters, it is crucial for BRICS nations to adopt effective policies to combat climate change. Between 2000 and 2023, the average CO₂ emission per person in BRICS countries was 276.20 metric tonnes, with a peak of 1,476.23 metric tonnes (Mehta & Shah, 2024).

This study investigates the environmental impact of artificial intelligence (AI) within the BRICS economies, focusing on two key research questions: (1) What are the short-term and long-term effects of AI adoption and energy transitions on carbon emissions? and (2) How do economic growth and value-added contributions from the industrial and agricultural sectors influence carbon footprints? To address these questions, the study employs robust quantitative methods, including the Pooled Mean Group Autoregressive Distributed Lag (PMG/ARDL) model, Dynamic Least Squares (DOLS), and Fully Modified Least Squares (FMOLS) techniques. The analysis covers the period from 2013 to 2023, assessing the impact of AI investment, energy transition efforts, and economic expansion on carbon emissions across BRICS nations.

Additionally, the research examines the specific roles of the industrial and agricultural sectors in driving emissions. In line with McElheran (2018), the study uses the prevalence of industrial robots as a proxy for AI deployment, placing particular emphasis on the industrial sector. Furthermore, drawing on Wang, He, Wang, and Zhao (2025), it considers the positive influence of AI on China’s economic growth, especially through the lens of population-related external systems.

The paper’s objectives are to provide solid empirical evidence on the links between AI investment, energy transition, economic growth, and carbon emissions in BRICS economies. Specifically, it will show how AI investment increases carbon emissions, while energy transition efforts help reduce them.

Additionally, the paper aims to demonstrate how AI can address global environmental issues and highlight the connection between CO₂ emissions and per capita GDP, a measure of economic growth and value-added in agriculture and industry. Research by El-Aal and Mohamed (2024) indicates that the agriculture sector contributes significantly to CO₂ emissions, especially in low-income countries, while industry and economic growth also play roles in emissions. The paper provides strong empirical evidence that, while industrial growth and economic expansion raise carbon emissions, agriculture helps reduce them.

The remainder of the paper is structured as follows: Section 2 provides a literature review to identify existing research gaps. Section 3 outlines the data and methodology used in the study. Section 4 presents the empirical analysis and discusses the findings. Section 5 and Section 6 offer conclusions and policy implications respectively. Finally, Section 7 identifies limitations and directions for future research.

The interaction between artificial intelligence (AI), the energy transition, growth, the value that industry and agriculture bring, and carbon emissions is a complicated and dynamic field of research. The purpose of this literature review is to summarise the current studies examining the relationships between these variables. This section is therefore divided into two subsections: (1) empirical studies examining the relationship between AI, the energy transition, and emissions; (2) the relationship between industry, agriculture, growth, and emissions.

1. Artificial Intelligence, Energy Transition and Carbon Emission

An increasing number of studies are exploring how various factors influence carbon emissions, energy transitions, and value addition as indicators of sustainability. This growing interest is exemplified by the pioneering research in energy economics conducted by Dong et al. (2022b), with further support from the findings of Wang, Ren, and Li (2024) and Ehigiamusoe and Dogan (2022). Among the many influencing factors, technological developments are considered particularly critical. According to Wang, Hu, and Li (2024), technology plays a transformative role in shaping both social life and economic productivity, thereby reshaping national energy consumption patterns and environmental outcomes. Technological factors can significantly affect energy transitions, carbon emissions, and broader ecological impacts.

Lei et al. (2023) reaffirm the findings of Lyu and Liu (2021), who argue that artificial intelligence (AI) holds considerable potential for advancing environmental sustainability by improving production efficiency, optimising manufacturing processes, and facilitating energy transitions and conservation efforts. However, as noted by Taddeo et al. (2021) and supported by Wu et al. (2022), AI also brings environmental challenges. These include the energy consumed throughout its lifecycle (a first-order effect) and the system-wide impacts of its deployment (a third-order effect), both of which can aggravate environmental problems.

Despite these concerns, there is a prevailing consensus—originating from Bard (1986) and reinforced by Aghion, Jones, and Jones (2017)—that AI can contribute to reducing carbon emissions. Wang et al. (2025) further elaborate that low levels of AI development hinder economic progress, moderate levels significantly stimulate economic growth, and high levels tend to yield diminishing or negligible returns. Technological advancements can enhance industrial structures, shift energy frameworks, and stimulate economic growth, all of which are associated with reduced carbon emissions, as shown in studies by Li and Wang (2017, 2022).

AI technologies are especially well-suited to addressing complex challenges by leveraging large datasets from diverse sources, thereby improving productivity and reducing CO₂ emissions per unit of GDP. Ding, Li, Shi, Li, and Chen (2023) support this view, showing through a global panel analysis and dynamic estimation methods that AI can significantly reduce carbon emissions. In contrast, Luan, Yang, Chen, and Regis (2022) present a more critical perspective, arguing that AI contributes to climate warming and correlates with increased air pollution across 74 countries. Notably, Wang, Zhang, Li, and Sun (2024) remain the only study to explicitly examine AI’s role in energy transitions, concluding that AI has a positive effect.

1. Agriculture, Industry, Economic Growth and Carbon Emission

In response to escalating environmental concerns, researchers are increasingly examining the factors that influence CO₂ emissions. These determinants include energy costs (He & Lin, 2011), technological advancement (Li & Wang, 2017; Churchill, Inekwe, Smyth, & Zhang, 2019), economic growth (Al-Mulali & Sab, 2012; Wang, Wu, Zhu, & Wei, 2013), and foreign direct investment (Talukdar & Meisner, 2001). Liu et al. (2017a) identify a causal relationship between economic growth and agriculture, as well as a direct link between agriculture and renewable energy. Their findings indicate that enhancing both renewable energy use and agricultural development can reduce CO₂ emissions in the ASEAN-4 countries. Similarly, Wang et al. (2013) show that in China, CO₂ emissions can be mitigated through advancements in technology, shifts in energy consumption, and engagement in international trade. Liu, Zhang, and Bae (2017b) further demonstrate that emissions are negatively affected by per capita output and renewable energy use, while positively influenced by per capita non-renewable energy consumption and agricultural activities. In the short term, they also identify a unidirectional relationship from output to non-renewable energy, and from agricultural value-added to output. Magazzino and Mele (2025) explore the impact of Russia’s economic growth on energy consumption and CO₂ emissions.

Although extensive research exists on the interrelationships among AI, economic growth, energy use, productivity, and emissions, there remains a notable gap. No existing study has specifically investigated how AI investment, energy transition, and productivity in the agricultural and industrial sectors collectively shape the relationship between AI, energy transition, and carbon emissions within the BRICS economies. This study addresses that gap by employing the novel PMG-ARDL methodology. The ultimate goal is to provide policy recommendations that support BRICS nations in achieving Sustainable Development Goal 9, which emphasises industry, innovation, and infrastructure geared toward environmentally sustainable industrialisation.

The study used annual data from 2013 to 2023 to examine the connections between AI investment, energy transition, economic development, value added from industry and agriculture, and carbon footprint in BRICS countries. If you look at a person’s “carbon footprint,” which is measured in tonnes of carbon dioxide equivalents, it shows how much carbon dioxide (CO₂), methane (CH₄), nitrous oxide (N₂O), hydrofluorocarbons (HFCs), perfluorocarbons (PFCs), and sulphur hexafluoride (SF₆) they contribute to the environment through their energy, waste, and farming. The following variables are included to examine the relationships in this study: annual private investment in artificial intelligence (AIi), defined as investments exceeding $1.5 million USD by businesses in AI-related projects; energy transition (ET), measured as the share of renewable energy consumption in total final energy consumption; and GDP per capita, expressed in constant 2021 US dollars based on purchasing power parity (PPP), representing economic growth (EG). Additionally, agricultural productivity (VAag) is measured by the value added (as a percentage of GDP) from agriculture, forestry, and fisheries, while industrial output (VAi) refers to the value added (as a percentage of GDP) from industry, including construction. All data are sourced from the World Bank’s World Development Indicators, except for AI investment data, which is obtained from Quid’s AI Index (2024), and the share of renewable energy, which is taken from the Energy Institute’s Statistical Review of World Energy (2024).

The model was built on the findings of previous studies by El-Aal and Mohamed (2024) (in the context of low- and high-income countries considering the effects of population, GDP, FDI, and value added from agriculture, industry, and services) and Wang, Li, and Li (2024) (concerning the AI impact on the environment and energy transition of 67 countries). To make estimation easier and less complicated, the study used natural logarithmic forms for all variables. This study recommends the following regression models for examining the relationship:

Where CF_it denotes carbon footprint of country i in year t; AIi_it, ET_it, EG_it,VAag_it, and VAi_it measures the AI investment, energy transition, economic growth, and value added from agriculture and industry of country i in year t; µ_it is the error term. The constant term is α₀. β₁, β₂, β₃, β₄, and β₅ represent the long-run coefficients of the explanatory variables.

The empirical study will first assess each variable’s integration order before looking at the association between explanatory factors and carbon footprint. Economic globalisation has made virtually every nation interdependent. Any financial crisis in one nation could have a big effect on others. The CSD test evaluates the cross-sectional correlations in this regard. If significant levels of positive residual cross-section dependence are present and ignored, unit root tests for panels of the first generation may produce erroneous findings as advised by Grossman and Krueger (1995) and Chudik and Pesaran (2015). So, it is only helpful to do second-generation panel unit root tests after it has been proven that residual cross-sectional dependency has a big effect on the panel. Power loss could happen if second-generation panel unit root tests that take cross-sectional dependence into account are used when cross-sectional dependence is not high enough. Therefore, it is necessary to present proof of the extent of residual cross-sectional dependency before selecting the appropriate panel unit root test. The CD test should be used even if a conventional unit root test shows cross-sectional independence (Sahoo & Sethi, 2022). We used the tests developed by Baltagi, Feng, and Kao (2012), Breusch and Pagan (1980), Pesaran (2007), and Pesaran, Ullah, and Yamagata (2008) to achieve this. Additionally, Breitung (2005) notes that if the data exhibit panel heterogeneity, presuming panel homogeneity will produce deceptive conclusions. To manage cross-sectional variation in the empirical results, Pesaran and Yamagata (2008) created slope homogeneity for panel datasets.

To make sure that all variables reveal unit roots at the same or different degrees of integration, it is crucial to check the unit root of each variable after confirming the CD in the panel data. When CSD is present in a series, the conventional test for panel unit roots is inappropriate and could provide misleading regression (Erdoğan, Yıldırım, Yıldırım, & Gedikli, 2020; Sahoo & Sethi, 2021). Since the previously suggested first-generation unit root tests (IPS and LLC) seem ineffectual or invalid, the stationarity of the data is thus verified using the second-generation unit root test (CIPS). To address this issue, we adopted Pesaran’s (2007) CIPS testing, which enables the testing and removal of potential cross-dependencies in panel data. The non-stationary null hypothesis is employed in the CIPS test. After that, an autoregressive distributed lag model (ARDL), which was first suggested by Pesaran, Shin, and Smith (1999) and Pesaran and Smith (1995), was used to look at the long- and short-term estimates of variables in this study. The ARDL model is better than traditional cointegration techniques because it accurately shows how the data is generated, can be used with series that are level or first-order differentially stationary, and reduces endogeneity and autocorrelation problems by including the right number of lags as determined by the Akaike, Schwarz, and Hannan-Quinn Information Criterion (we won’t go into detail because of space constraints).This study uses a series that combines first-order differential stationarity with level stationarity. It demonstrates that since the variables are not stationary in the same order, we are unable to apply the Pedroni Cointegration model in this case. Because of this, the ARDL method is always the best choice, even if the study regression number shows I (0), I (1), or a mix of the two (Pesaran and Smith, 1995). An 11-year data cycle informs the application of the panel data approach.

To examine the long- and short-term relationships between AI investment, energy transition, economic growth, value added from industry and agriculture, and carbon footprint, the researchers used pooled mean group (PMG) in light of the extensive literature on dynamic panel data. PMG restricts long-term equilibria to be homogeneous across countries while allowing for heterogeneity in short-term relationships. The short-term link shows that each country is different, which could explain why different countries react differently to stability policies, outside factors, or financial crises. The data obtained from the BRICS economies in this study indicate a consistent pattern in long-term economic growth. Given the country-specific differences, short-run behaviour exhibits heterogeneity, making the PMG estimator a superior model. Therefore, we can reform Equation (1) into the primary model of the ARDL form and prescribe it as follows.

Where the error term is ε_it and the intercept is α₁. The long-run relationship and error correction dynamics are represented by the first and second portions of the equation, respectively.

The study also used fully modified ordinary least squares (FMOLS), which was first suggested by Pedroni (2001), and dynamic ordinary least squares (DOLS), which was first suggested by Mark and Sul (2003). The pooled method employed both techniques to ensure a robust analysis. The FMOLS and DOLS methods use different approaches to deal with endogeneity bias and serial correlation when figuring out long-term relationships in panels that are not all the same. DOLS employs parametric methodologies. DOLS indicates the presence of cointegration relationships among all variables. DOLS addresses contemporaneity by identifying the values of leading and lagging variables. The FMOLS method employs a non-parametric approach to address endogeneity and serial correlation concerns.

Once we confirm the long-term relationship between the variables, we use ECM to estimate the short-run coefficients and error correction term (ECT). The following symbol represents the ECM:

This study employs multiple diagnostic techniques to detect and address outliers, thereby enhancing the robustness of the PMG/ARDL (panel) estimation. To assess the presence of cross-sectional dependence, the analysis incorporates the Pesaran scaled LM test, the bias-corrected scaled LM test, Pesaran’s CD test (PCD), and the Breusch-Pagan LM test. Stationarity is evaluated using the Cross-sectionally Augmented IPS (CIPS) test, which accounts for both cross-sectional dependence and potential outliers. These diagnostic measures collectively strengthen the validity and reliability of the model estimates by mitigating the effects of outliers and data irregularities.

The panel estimators discussed above do not give enough details about the causal connections between the variables. The Dumitrescu and Hurlin (2012) method for panel causality testing is utilised to assess the causal relationships among AI investment, energy transition, economic growth, agricultural productivity, industry output, and carbon footprint. We recognise the test as an enhanced iteration of Granger causality analysis. This test’s use lies in its ability to analyse imbalances in panel data, taking into account both the time dimension and the size of the cross-section. It performs effectively in contexts characterised by slope heterogeneity and cross-sectional dependence among nations (Bhujabal, Sethi, & Padhan, 2021; Dumitrescu and Hurlin, 2012).

The descriptive statistics presented in Table 1 indicate a lack of stability in the variables throughout the study period. The maximum and minimum values of carbon footprint and AI investment exhibit considerable variability from their mean values. Regarding AI investment, the average is 2,711 million USD, with a maximum of 26,600 million USD and a minimum of 0.89 million USD, indicating significant instability throughout the study period. The elevated standard deviation values (CF 4.7754, AIi 5396.21, ET 16.45, EG10085.59, VAag 5.24, and VAi 7.00) further indicate instability. During the study period, the energy transition, economic growth, and value added by agriculture and industry all had high maximum values that varied a lot from their mean. The significant disparities between maximum and minimum values indicate that many variables exhibit considerable instability during this period.

Table 2 shows the results of all cross-section dependency (CSD) tests, such as BPLM (Breusch-Pagan LM), PSLM (Pesaran scaled LM), BCSLM (bias-corrected scaled LM), and PCD (Pesaran CD). These tests are used instead of the usual unit root test to avoid false regression. The results strongly disagree with the null hypothesis, which says that there is no cross-sectional dependence between nations. They also confirmed cross-sectional dependence for a group of BRICS nations. According to this research, a shock that strikes one developing nation could spread to others. In the bottom part of Table 2, the “slope homogeneity test” of Pesaran and Yamagata (2008) was conducted, and the result indicated that country-specific heterogeneity is confirmed for rising nations when the null hypothesis of slope homogeneity is thrown out.

The present study employed the CIPS unit root test, proposed by Pesaran (2007), after validating the cross-sectional dependency in the panel data since it yields consistent findings in cross-sectional dependence data (Bhujabal et al., 2021; Khan, Yu, Belhadi, & Mardani, 2020; Khattak, Ahmad, Khan, & Khan, 2020). The unit root analysis was done to a level, and at first, the difference of each series. The CIPS unit root test in Table 3 demonstrated that the variables of interest possess both stationary and non-stationary characteristics. At order I (1), lnCF, lnET, and lnVAag are stationary, whereas at order I (0), lnAIi, lnEG, and lnVAi are stationary at the 5% significance level. As a result, the variables under investigation show a combination of stationarity traits, including level and first-order difference stationarity. It confirms that the variables are not stationary in the same order; hence, the study is unable to employ the Pedroni Cointegration model.

Before performing PMG/ARDL test, it is necessary to determine the suitable lag length, as the autoregressive model is sensitive to lag selection. Table 4 displays the appropriate lag length determined by the widely accepted parameters, AIC, SIC, and HQC. The ideal lag length for this model is 1, as recommended by all three criteria.

This study employed a panel ARDL model to describe the short- and long-term connections without cointegration. The panel ARDL technique is always the best choice, according to Pesaran and Smith (1995), whether the research regression quantity shows I (0), I (1), or both. The study used pooled mean group (PMG) to look into the short- and long-term effects of AIi, ET, EG, VAag, and VAi on CF. This was done because there was a lot of research on dynamic panel data. Table 5 showed long-run estimation under all three methods adapted.

Contrary to the FMOLS and DOLS methods, the PMG/ARDL result showed a significant improvement in carbon footprint with AI investment, which aligns with previous research by Higón, Gholami, and Shirazi (2017) and Luan et al. (2022). In the long term, all three estimates agree with what Liu et al. (2017a, b) and Ehigiamusoe and Dogan (2022) found: the energy transition has a statistically significant absorbed carbon footprint because it uses less fossil fuel energy. All three overtime estimations revealed that economic growth significantly increases carbon footprints, on par with earlier research by Ang (2008) and Cheng, Ren, and Wang (2019). Agricultural yield statistically impedes carbon footprint, as confirmed by all three estimators. These findings are also similar to those of Liu et al. (2017a) but in contradiction with El-Aal and Mohamed’s (2024) findings. Though FMOLS and DOLS showed industry output is insignificant, PMG/ARDL confirms it has a high rate of influence on carbon footprint expansion. It indicated that a 1% increase in the industry would lead to a 44% surge in emissions.

The top part of Table 6 shows a short-run effect: none of the variables have a big effect on carbon footprint, which is supported by the PMG/ARDL model. The only thing that does is economic growth, which causes emissions to rise.

Lastly, the Dumitrescu-Hurlin test for panel causality in the lower part of Table 6 confirmed agricultural productivity causes economic growth similar to Bashir, Thamrin, Farhan, and Atiyatna (2019) and industry output. The energy transition is also causally associated with industry output. Furthermore, rising economic growth leads to improved AI investment in BRICS economies.

The diagnostic results presented in the upper-central section of Table 6 support the overall adequacy of the model. The Root Mean Squared Error (RMSE) is low (0.0113), indicating minimal prediction error. Additionally, the Akaike Information Criterion (AIC), Schwarz Criterion (SC), and Hannan–Quinn (HQ) values are relatively low, while the log-likelihood score is high (174.03), all of which suggest a good model fit.

This section discusses the results and their applications. The findings show that industry value addition, economic growth, and AI investment are key factors in increasing the carbon footprint of the BRICS nations. The empirical study indicates that AI investments contribute to higher CO2 emissions, likely due to the production of machinery and equipment. This aligns with Higón et al. (2017), who found that ICT can have both positive and negative environmental impacts, with developed nations seeing a decline in emissions as ICT advances. Luan et al. (2022) argued that AI exacerbates climate change and air pollution in 74 countries. Huang, Haseeb, Usman, and Ozturk (2022) observed that while G-7 nations reduced their ecological footprint, E−7 nations saw an increase due to ICT growth.

However, studies such as Wang, Li, and Li (2024), Zhong, Zhong, Han, Yang, and Zhang (2023), and Li, Zhang, Pan, Han, and Veglianti (2022) suggest that AI can reduce carbon emissions. The study confirms that higher per capita GDP and industrial growth are linked to increased emissions, echoing findings from Cheng et al. (2019) related to OECD countries and Ang (2008) in Malaysia. It also suggests that energy transition and renewable energy use can mitigate emissions, particularly in the industrial sector. El-Aal and Mohamed (2024) noted similar findings, where industry sectors in high- and low-income countries contribute significantly to emissions, but energy transition can reduce reliance on fossil fuels.

The results also highlight that energy transition, including greater use of renewable energy, and increased agricultural productivity substantially reduce emissions. This aligns with Ehigiamusoe and Dogan (2022), who argued that renewable energy reduces emissions in low-income countries. Liu et al. (2017a, b) found similar effects in ASEAN-4 and BRICS nations. The study suggests that the BRICS economies should prioritise renewable energy to improve environmental quality, as these countries are major contributors to global CO2 emissions.

Additionally, agricultural productivity in the BRICS nations plays a significant role in absorbing emissions, though this finding contrasts with some previous studies. Overall, the BRICS nations are transitioning from fossil fuels to renewable energy, supporting the agriculture-industry growth-led hypothesis and reinforcing the potential for a cleaner future through enhanced renewable energy and agricultural productivity.

This study contributes to the literature by examining the relationships between AI investment, energy transition, economic growth, agricultural productivity, industrial output, and carbon footprint in BRICS countries from 2013 to 2023. It accounts for slope heterogeneity and cross-sectional dependence, which previous research did not fully address. To test stationarity with cross-sectional dependence, we used the CIPS test. The short- and long-term effects of AI investment, energy transition, economic growth, and industry/agriculture value added on the carbon footprint were analysed using the PMG/ARDL technique. FMOLS and DOLS methods ensured accuracy. The Dumitrescu-Hurlin test was used to assess causality between variables. Key findings include: (1) The PMG/ARDL model shows that AI investment significantly increases the carbon footprint, while FMOLS and DOLS show insignificant effects. (2) All models indicate that energy transition significantly reduces the carbon footprint in the long run. (3) All three estimators confirm that agricultural productivity reduces emissions. (4) PMG/ARDL shows that industry output has a substantial impact on emissions, with a 1% increase in industry leading to a 44% rise in emissions. (5) Economic growth significantly increases emissions in the long run. (6) In the short term, economic expansion is the only variable that significantly influences carbon emissions; other factors exhibit minimal or statistically insignificant effects. (7) The Dumitrescu-Hurlin panel causality test reveals that agricultural productivity causally affects both industrial output and economic growth, while the energy transition has a direct causal impact on the industrial sector. Additionally, investment in AI is driven by economic expansion.

The study offer following policy implications:

1. The development of AI in BRICS requires extensive machinery production, leading to higher CO2 emissions. Policymakers should promote energy conservation and the use of renewable energy to offset these emissions.

2. Increasing renewable energy usage can absorb emissions and boost industry output. Governments should invest in renewable energy R&D and promote clean energy technologies like nuclear, wind, and hydro.

3. Policies should encourage the adoption of advanced energy transition technologies by both industries and the agricultural sector to enhance sustainability and, in turn, stimulate short-term economic growth.

4. BRICS nations should create flexible environmental policies to ease the adoption of renewable energy technologies, focusing on forest resource production.

5. Although AI may initially have adverse environmental effects, it holds significant potential to drive long-term economic growth. Policymakers should therefore support the integration of AI in value-added sectors to foster sustainable and inclusive economic development.

This study focuses solely on BRICS countries and does not consider other developed or developing nations. Future research could compare BRICS with other countries and use time-series data to analyse these variables individually. Additional factors like population density, government spending, governance quality, and taxation were not included, and their effects warrant further investigation. Future studies could provide more specific policy recommendations for individual BRICS nations and explore the impact of these factors on environmental quality.

The authors would like to express their gratitude to the anonymous journal referees for their helpful advice on improving the article. The authors are also grateful to his family, friends, and colleagues and special thanks to the Editors of the Journal for taking submission of this article.

• https://edgar.jrc.ec.europa.eu/index.php

• https://www.iea.org/publications/freepublications/publication/kwes.pdf

• http://www.worldenergyoutlook.org/publications/weo-2012/#d.en.26099

• https://ourworldindata.org/grapher/private-investment-in-artificial-intelligence

• https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-ai-frontier-modeling-the-impact-of-ai-on-the-world-economy

The supplementary material for this article can be found online.