To explore the correlation between perceived economic policy uncertainty by businesses and carbon emissions, and to verify whether this relationship exhibits non-linearity, this paper constructs OLS regression models (1) and (2).

(1)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}loa{n}_{it}+{\beta }_{3}oc{f}_{it}+{\beta }_{4}ebi{t}_{it}+{\beta }_{5}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(2)$$\begin{array}{c}\hfill lnco2d_{it}=a_0+{\beta }_{1}uncertainty+{\beta }_{2}uncertainty22+{\beta }_{3}loa{n}_{it}+{\beta }_{4}oc{f}_{it}+{\beta }_{5}ebi{t}_{it}+{\beta }_{6}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

In Eqns. 1,2, $i$ represents the region and $t$ represents the year; $lnco{2}_{it}$ is the dependent variable, representing the total amount of carbon dioxide emissions. $uncertaint{y}_{it}$ is an independent variable representing the perception of uncertainty in corporate economic policies; $uncertainty{22}_{it}$ is the square term of perceived uncertainty in corporate economic policies; $loa{n}_{it}$ represents the ratio of long-term borrowings to total assets. $oc{f}_{it}$ represents the cash flow generated from operating activities; $ebi{t}_{it}$ represents pretax profit; $ro{e}_{it}$ stands for Return on Equity. ${\beta }_1$~${\beta }_6$ represents variable coefficient; $a_0$ denotes the constant term; $\epsilon$ is the random error term.

To study the moderating effect of corporate decision-makers’ (CEOs, executives, and directors) green cognition background on the relationship between perceived economic policy uncertainty by businesses and carbon emissions, this paper constructs a moderating regression model. To eliminate multicollinearity between interaction terms and their components, thus making the model estimation more robust and interpretable, the interaction terms were centralized in this study. This approach allows for a more accurate explanation of the relationships between variables, avoiding the confounding effects of multicollinearity, and thus leading to a better understanding of the nature and impact of the moderating effect. The specific models are shown in Eqns. 3,4,5.

(3)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}epce{o}_{it}+{\beta }_{3}uncertaint{y}_{it}*epce{o}_{it}+{\beta }_{4}loa{n}_{it}+{\beta }_{5}oc{f}_{it}+{\beta }_{6}ebi{t}_{it}+{\beta }_{7}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(4)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}epbod{r}_{it}+{\beta }_{3}uncertaint{y}_{it}*epbod{r}_{it}+{\beta }_{4}loa{n}_{it}+{\beta }_{5}oc{f}_{it}+{\beta }_{6}ebi{t}_{it}+{\beta }_{7}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(5)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}epex{r}_{it}+{\beta }_{3}uncertaint{y}_{it}*epex{r}_{it}+{\beta }_{4}loa{n}_{it}+{\beta }_{5}oc{f}_{it}+{\beta }_{6}ebi{t}_{it}+{\beta }_{7}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

In Eqns. 3,4,5, $i$ represents the region and $t$ represents the year; $lnco{2}_{it}$ is the dependent variable, representing the total amount of carbon dioxide emissions. $uncertaint{y}_{it}$ is an independent variable representing the perception of uncertainty in corporate economic policies; $uncertainty{22}_{it}$ is the square term of perceived uncertainty in corporate economic policies; $epce{o}_{it}$ represents whether the CEO has a green cognitive background; $epex{r}_{it}$ represents the proportion of executives with a green cognitive background; $epbod{r}_{it}$ represents the proportion of directors with a green cognitive background; $loa{n}_{it}$ represents the ratio of long-term borrowings to total assets. $oc{f}_{it}$ represents the cash flow generated from operating activities; $ebi{t}_{it}$ represents pretax profit; $ro{e}_{it}$ stands for Return on Equity. ${\beta }_1$~${\beta }_7$ represents variable coefficient; $a_0$ denotes the constant term; $\epsilon$ is the random error term.

To investigate the mediating roles of total factor productivity and artificial intelligence in the relationship between perceived economic policy uncertainty by businesses and carbon emissions, this paper constructs a mediation effect regression model. The specific models are shown in Eqns. 6,7,8,9.

(6)$$\begin{array}{c}\hfill tfp{1}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}loa{n}_{it}+{\beta }_{3}oc{f}_{it}+{\beta }_{4}ebi{t}_{it}+{\beta }_{5}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(7)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}tfp{1}_{it}+{\beta }_{2}uncertaint{y}_{it}+{\beta }_{3}loa{n}_{it}+{\beta }_{4}oc{f}_{it}+{\beta }_{5}ebi{t}_{it}+{\beta }_{6}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(8)$$\begin{array}{c}\hfill ai{2}_{it}=a_0+{\beta }_{1}uncertaint{y}_{it}+{\beta }_{2}loa{n}_{it}+{\beta }_{3}oc{f}_{it}+{\beta }_{4}ebi{t}_{it}+{\beta }_{5}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

(9)$$\begin{array}{c}\hfill lnco{2}_{it}=a_0+{\beta }_{1}ai{2}_{it}+{\beta }_{2}uncertaint{y}_{it}+{\beta }_{3}loa{n}_{it}+{\beta }_{4}oc{f}_{it}+{\beta }_{5}ebi{t}_{it}+{\beta }_{6}ro{e}_{it}+{\epsilon }_{it}\hfill \end{array}$$

In Eqns. 6,7,8,9, $i$ represents the region and $t$ represents the year; $lnco{2}_{it}$ is the dependent variable, representing the total amount of carbon dioxide emissions. $uncertaint{y}_{it}$ is an independent variable representing the perception of uncertainty in corporate economic policies; $uncertainty{22}_{it}$ is the square term of perceived uncertainty in corporate economic policies; $tfp1$ stands for total factor productivity; $ai2$ represents the degree of transformation of artificial intelligence; $loa{n}_{it}$, $oc{f}_{it}$, $ebi{t}_{it}$ and $ro{e}_{it}$ are control variables. ${\beta }_1$~${\beta }_6$ represents variable coefficient; $a_0$ denotes the constant term; $\epsilon$ is the random error term.

Carbon emissions ($lnco2$) serve as the core explanatory variable in this study. Given that carbon dioxide is one of the primary greenhouse gases responsible for the rise in global temperatures, it can form a barrier in the atmosphere that prevents heat from the Earth’s surface from escaping into space, thereby leading to global warming. Thus, it is used to measure climate change. The emission of carbon dioxide primarily results from human activities such as the combustion of fossil fuels (e.g., coal, oil, and natural gas) and deforestation. Therefore, by measuring carbon emissions, scientists and policymakers can estimate the extent of human activity’s impact on the global climate and, in turn, take measures to reduce greenhouse gas emissions to mitigate the pace and impact of climate change (Cui et al, 2022).

The dependent variable in this paper is businesses’ perceived economic policy uncertainty ($uncertainty$). Drawing on the method of Zheng and Wen (2023) and Liao et al (2019), this study uses Python web-scraping software and the Jieba segmentation module to obtain firm-level data on perceived economic policy uncertainty. First, words representing “economic policy” and words representing “uncertainty” were manually summarized (due to space limitations, specific words are provided in the appendix). Next, the text of the MANAGEMENT’S DISCUSSION AND ANALYSIS (MD&A) sections of annual reports from listed companies was processed, with all punctuation marks except for the Chinese period removed, and the text was segmented into sentences using the Chinese period as the delimiter. If a sentence contains both words representing “economic policy” and words representing “uncertainty”, that sentence is considered an “economic policy uncertainty” sentence (P). Finally, using Python software and the segmentation module, sentences were converted into a series of word combinations, and the number of words ($n_s$) in “economic policy uncertainty” sentences was counted. The ratio of this count to the total number of words (N) in the MD&A section was used to measure firm-level perceived economic policy uncertainty. The larger the ratio, the higher the business’s perceived economic policy uncertainty. The specific calculation method is shown in Eqn. 10:

(10)$$uncertainty={\sum }_{s=1}^{s_{it}}{n}_s{I}_p(s)/N$$

Among them, $I_p(s$) is a demonstrative function. When S $\in$ P, I = 1; when S $\mathrm{\Delta }$ P, I = 0.

Decision-makers’ green cognition: Referring to the measurement method for executive cognition proposed by Liu and Chen (2024), a textual analysis of listed companies’ annual reports is conducted, and relevant keywords are selected for word frequency statistics to construct the green cognition of corporate decision-makers. The original data on decision-makers’ environmental backgrounds are sourced from the personal resume information published on the Sina Finance website. If the resume includes keywords such as ‘environment’, ‘environmental protection’, ‘new energy’, ‘clean energy’, ‘ecology’, ‘low-carbon’, ‘sustainability’, ‘energy-saving’, or ‘green’, the individual is considered to have an environmental background. Based on this, the number of CEOs, executives, and directors with green cognition backgrounds in a company is calculated.

(1) CEO green cognition ($epceo$): If the company hires one or more executives with environmental backgrounds during the year, it is assigned a value of 1; otherwise, it is 0.

(2) Proportion of executives with green cognition ($epexr$): The number of executives with green cognition is calculated using the above method, and then the proportion of these executives relative to the total number of executives in the company is determined.

(3) Proportion of directors with green cognition ($epbodr$): The number of directors with green cognition is calculated using the above method, and then the proportion of these directors relative to the total number of directors in the company is determined.

(1) Total Factor Productivity ($tfp$): Total factor productivity is an efficiency indicator that measures how effectively an economy or enterprise utilizes all its production factors (such as labor, capital, technology, etc.) in the production process. It reflects the increase in output resulting from advancements in production technology and improvements in management levels, assuming other inputs remain constant. The data is sourced from the CSMAR database.

(2) Artificial Intelligence ($ai2$): This paper references the methods of Zhong et al (2022), using the frequency of AI-related terminology in the annual reports of listed companies as a proxy variable for artificial intelligence. The AI data is sourced from the CnOpenData database.

(1) Long-term debt to total assets ratio ($loan$): This measures a company’s financial leverage and its reliance on borrowing to fund assets. Higher ratios indicate greater financial risk and potentially higher pressure to manage resources cautiously. High-debt companies may be less able to invest in costly green technologies due to financial constraints, affecting their carbon emissions. Controlling for this variable helps isolate the impact of capital structure on emissions.

(2) Operating cash flow ($ocf$): This reflects the cash generated from a company’s core operations. Strong OCF enables companies to invest in long-term projects and maintain stability in response to policy changes. Companies with healthy cash flow are more likely to invest in green technologies, reducing emissions. Controlling for OCF ensures that fluctuations in cash flow do not skew the results.

(3) Earnings Before Interest and Taxes ($ebit$): This measures a company’s operational profitability, excluding the effects of financing and taxes. Controlling for EBIT helps eliminate the influence of financing costs and taxes on emission decisions, focusing instead on operational efficiency. Companies with higher EBIT may have more financial flexibility to invest in emission control measures and may respond differently to policy uncertainty.

(4) Return on equity ($roe$): This measures a company’s profitability relative to shareholder equity. ROE reflects both profitability and the efficiency of using shareholder funds. Controlling for ROE helps remove the influence of profitability differences on emissions. High-ROE companies are more likely to invest in environmental technologies and may respond differently to policy uncertainty compared to less profitable firms.

Mediation analysis assesses how a mediating variable links an independent variable (X) to a dependent variable (Y) by focusing on the indirect effects of X on Y. It tests and measures the strength of these indirect effects. The Sobel test, a common method, evaluates the significance of indirect effects using path coefficients and standard errors, with variations like Goodman1 and Goodman2 adjusting for potential bias. The bootstrap test, widely used today, estimates indirect effects through repeated sampling. These tests aim to confirm whether X influences Y indirectly through the mediator. Employing multiple methods strengthens the validity of the results, particularly across varying sample sizes and data distributions. In this study, several mediation tests were conducted, and as shown in Table 7, total factor productivity and artificial intelligence successfully passed these tests, confirming the robustness and effectiveness of the mediation analysis.

To examine how total factor productivity (TFP) mediates the effect of firms’ perception of economic policy uncertainty on carbon emissions, we conducted a mediation regression analysis (results in Table 8). Economic policy uncertainty boosts TFP, with a coefficient of 0.566 (significant at the 1% level), as firms maximize resources like energy, labor, and capital to stay competitive in uncertain times, leading to short-term productivity gains. However, higher TFP increases carbon emissions, with a coefficient of 1.445 (significant at the 1% level), as firms relying on cheaper, traditional energy sources like coal and oil to quickly enhance productivity also drive up emissions. For instance, industrial firms anticipating tax reforms or stricter environmental rules may ramp up production to secure profits before new regulations take effect, causing a short-term surge in emissions.

During this process, increased energy consumption leads to higher carbon emissions, especially as firms invest in high-emission projects ahead of regulatory changes. According to the short-sighted theory, firms facing high policy uncertainty focus on immediate market responses rather than long-term strategic planning. In such contexts, firms may also relax their adherence to environmental standards to save costs, temporarily improving TFP but exacerbating environmental degradation over time. Instead of investing in sustainable technologies, they opt for cost-effective, high-carbon production methods, validating hypothesis 4 of this study. Thus, policy uncertainty drives TFP growth but at the expense of higher carbon emissions and long-term environmental health.

To explore how artificial intelligence (AI) mediates the effect of firms’ perception of economic policy uncertainty on carbon emissions, we conducted a mediation regression analysis (results in Table 9). First, policy uncertainty promotes AI adoption, with a coefficient of 0.024 (significant at 5%). In uncertain environments, firms invest in AI technologies such as automation, smart supply chains, and data analytics to quickly enhance productivity and cope with external challenges, helping them remain competitive. Second, increased AI usage is associated with higher carbon emissions, with a coefficient of 0.795 (significant at 1%). Despite its advanced capabilities, AI’s high energy consumption mirrors the carbon-intensive nature of traditional industries. As firms deploy AI to counter policy uncertainty, their growing reliance on these technologies raises carbon emissions.

From a Marxist perspective, where the primary goal of a capitalist economy is profit maximization, firms turn to AI to boost efficiency and productivity, particularly through automation that reduces labor costs. However, the development and operation of AI require significant energy, contributing to higher emissions, especially in data centers. This dynamic reflects the “efficiency paradox” (Jevons Paradox), where AI-driven productivity gains lead to increased energy demand and higher carbon emissions, validating hypothesis 4. Thus, while AI boosts productivity, it also significantly increases emissions.

To test the robustness of the regression results, this study employed methods such as replacing the main regression variables and adding control variables for robustness checks, with the results shown in Table 10. Firstly, the text used the proportion of words related to firms’ economic policy uncertainty ($uncertainty1$) and the proportion of uncertain sentences ($uncertainty2$) to replace firms’ perception of economic policy uncertainty ($uncertainty$). The regression results are largely consistent with the baseline regression results, as can be seen in Table 10 (1) and (2). Secondly, by gradually adding control variables, cash flow ratio and Tobin’s Q were included, and the regression results remained consistent with the baseline regression results, as seen in Table 10 (3) and (4). Therefore, the robustness of the regression has been validated.

To address endogeneity, this study employs both “Two-Stage Least Squares (2SLS)” and the “Generalized Method of Moments (GMM)”. Endogeneity occurs when explanatory variables are correlated with the error term, often due to omitted variables, measurement errors, or reverse causality. 2SLS tackles this issue by using instrumental variables that are strongly correlated with the endogenous variables but uncorrelated with the error term, ensuring consistent and unbiased estimates. The validity of the instruments is confirmed through tests for correlation and exogeneity, enhancing the reliability of the analysis. GMM, on the other hand, offers greater flexibility and robustness, particularly in handling endogeneity in the presence of heteroskedasticity or autocorrelation. By using multiple instruments, GMM accounts for the correlation of error terms and provides consistent estimates, making it a suitable method for analyzing the impact of economic policy uncertainty on carbon emissions. Both 2SLS and GMM help mitigate the biases associated with endogeneity, ensuring more accurate results.

As shown in Table 11, in both the two-stage least squares (2SLS) and generalized method of moments (GMM) regression results, firms’ perception of economic policy uncertainty has a positive effect on carbon emissions, with coefficients of 2.184 and 2.991, significant at the 1% and 10% levels, respectively. The regression results are consistent with the baseline regression results, indicating that the results are robust and reliable.