Prior research has demonstrated that certain governments engage in the manipulation of air quality data through the dissemination of fake positive news, with the objective of meeting pollution control requirements38. This evidence serves to illustrate the extent of government control over environmental reporting. Conversely, countries with greater press freedom are more likely to have media organizations that operate independently, thereby providing the public with a broader and more diverse range of information. Accordingly, the objective of this pilot study is to examine the relationship between press freedom and air quality by analyzing the correlation between the Press Freedom Index and Air Quality Indices.

In order to investigate this relationship, data were collected on the Press Freedom Index, the Air Quality Index, the GDP per capita and population density from 120 countries and regions. The data pertaining to the press freedom index was derived from the World Press Freedom Index, as published by Reporters Without Borders, while the air quality data was procured from Purelogic Labs. A partial correlation analysis was conducted to examine the association between the press freedom index and the air quality index, with the inclusion of GDP per capita and population density to isolate the specific effect of press freedom.

The variables employed in the partial correlation analysis are delineated as follows:

The World Press Freedom Index (WPFI), published by the international non-governmental organization Reporters Without Borders in 2023, is an evaluation of the level of press freedom globally. It reflects the extent of freedom enjoyed by journalists, news organizations, and internet users in 180 countries and regions worldwide. A higher WPFI score indicates a greater degree of press freedom, which in turn suggests that media outlets are more likely to report independently and comprehensively.

The AQI is a numerical scale that measures the quality of the air on a daily basis, taking into account its impact on human health and the environment. AQI values that are higher than the established threshold indicate a deterioration in air quality. The AQI data for 2022 was obtained from the AQI website (https://www.aqi.in/), which is managed by Purelogic Labs. This website provides historical air quality data based on guidelines set by the U.S. Environmental Protection Agency.

The gross domestic product (GDP) per capita serves as a measure of economic development for each country or region. In this study, GDP per capita was included as a control variable in accordance with the established relationship between economic development and air quality59,60. The data for 2022 GDP per capita were obtained from the World Bank.

Population density refers to the number of individuals residing in a given area, expressed as a ratio of individuals per unit area. This indicator provides insight into the concentration of the population in a specific region. Prior research has indicated a significant correlation between population density and air quality8. The 2022 data for population density were also sourced from the World Bank.

Study 1 methods and data analysis

Building on the preliminary findings from the pilot study, which explored the relationship between the WPFI and the Air Quality Index across 120 countries, Study 1 shifts its focus to the domestic context of China and specifically concentrates on air pollution reports. This study examines the relationship between air quality and media reporting on air pollution in 33 Chinese cities during 2022, utilizing data from the WiseSearch News database. WiseSearch aggregates content from over 1200 Chinese newspapers and more than 6000 domestic and international websites, with records dating back to 1998. It is widely regarded as a reliable source for research due to its reputation for reproducibility, timeliness, and high data integrity. In order to determine which cities would be included for the purposes of data collection and subsequent analysis, we initially selected 27 provincial capitals and 4 municipalities, in accordance with the administrative divisions of China. Moreover, in light of the likelihood that more economically developed cities will receive greater news coverage, the top 30 cities ranked by GDP were also included. As a consequence of the aforementioned criteria, the final sample comprised 46 cities for the initial search of air pollution-related reports, given that there was some overlap between these cities and the provincial capitals/municipalities. The data on air pollution reporting for the selected cities was collated from the WiseSearch News database, with the period under consideration being from 1 January 2022 to 31 December 2022. The following keywords were used in the search: The following search terms were entered into the search engine in conjunction with the name of the city in question: “smog,” “mild pollution,” “moderate pollution,” “heavy pollution,” “severe pollution,” and “polluted weather.” The data were sourced exclusively from newspaper media, classified under the category of comprehensive news, and limited to the province in which the city is situated. For each city, the total number of days during which air pollution was reported in 2022 was calculated. To provide an example, data on Guangzhou were collected using the following search parameters: The search terms “smog,” “mild pollution,” “moderate pollution,” “heavy pollution,” “severe pollution,” and “polluted weather” were entered into the search engine in conjunction with the name of the city in question. The media region was specified as Guangdong Province. This resulted in the total number of days in 2022 during which air pollution was reported in Guangzhou.

Moreover, data regarding the precise number of days during which pollution levels exceeded the permitted threshold in each city throughout 2022 was procured from the China Air Quality Online Monitoring and Analysis Platform, which is a well-regarded and authoritative source for air quality data in China. The platform provides daily air quality data, with the first entry dated 2 December 2013. The daily AQI is calculated as the mean of the hourly data provided by the National Environmental Monitoring Center. The total number of days during which air pollution exceeded the threshold level of 100 on the AQI scale, indicating mild pollution or worse, was calculated for each city in 2022.

Subsequently, the proportion of accurately reported pollution days for each city was calculated. This was defined as the ratio of the number of days on which pollution was reported to the number of actual days on which pollution occurred. This proportion constituted the primary outcome variable in the analysis.

$$ {Proportion\; of\; accurately\; reported\; pollution\; days}\\ =\frac{{Total\; days\; wit}h{reported\; air\; pollution}}{{Total\; days\; with\; actual\; air\; pollution}}$$ (1)

Following the completion of the database search, 11 cities were excluded on the grounds that they had no air pollution-related reports, while a further two cities were removed on the basis that they experienced no pollution throughout the year. This resulted in a final sample of 33 cities for further analysis. In the analysis, potential confounding variables were controlled for, including the 2022 per capita Gross Regional Product for each city, sourced from the National Bureau of Statistics of China. Furthermore, the number of days with reported weather conditions was taken into account, as cities with more frequent weather reporting may have increased coverage of air pollution. The data on weather reporting was gathered from the WiseSearch News database using the following search terms: “weather OR air quality AND city name.”

A Pearson partial correlation analysis was conducted to examine the relationship between the proportion of accurately reported pollution days and the 2022 AQI across the 33 cities. In order to account for potential confounding variables, this analysis controlled for 2022 per capita GDP and the number of days with reported weather conditions.

Furthermore, studies on air pollution in China have identified PM2.5 and ozone as the most critical pollutants affecting ambient air quality42,61. However, the perception of these pollutants varies considerably between individuals. The general public is more likely to observe PM2.5, a key component of smog, than ozone pollution. It can be hypothesized that public awareness of ozone pollution may be more dependent on media reporting of air quality than on direct observation of the pollutant itself. Furthermore, the temporal and spatial characteristics of air pollution in China indicate that particulate matter pollution is primarily observed from January to March and from November to December62, while ozone pollution is more prevalent from April to October63,64.

Accordingly, two distinct analyses were conducted on the 2022 data. The initial analysis sought to ascertain the relationship between the proportion of accurately reported pollution days and the AQI during the periods of elevated particulate matter pollution (January to March and November to December). This analysis employed Pearson partial correlations, with the total number of news reports during the specified months and the 2022 per capita GDP serving as control variables. The second analysis concentrated on the ozone pollution periods (April to October) and employed Pearson partial correlations, controlling for the corresponding months’ news report totals and 2022 per capita GDP.

Study 2 methods and data analysis

Building on the findings of the pilot study and Study 1, which examined the relationship between media reporting and actual air quality using publicly available data at both global and regional levels, Study 2 takes a more focused approach. In this study, we utilize an experimental priming method to investigate the impact of fictitious positive air quality reports on individuals’ behaviors and perceptions of air pollution. Prior research has demonstrated that individuals’ travel behavior has a significant impact on real air quality, as increased outdoor activities can contribute to emissions12. The objective of Study 2 is to examine the impact of fake positive air quality reports on individuals’ perceptions of air pollution and their willingness to travel. Additionally, the study will investigate whether perceptions of air pollution act as a mediator in the relationship between report type and travel intentions.

Prior to data collection, a G*Power analysis was conducted to determine the requisite sample size for detecting medium effect sizes (f = 0.25) with α = 0.05 and power = 0.80, as proposed by ref. 65. The analysis indicated that a minimum sample size of 210 participants would be required. A total of 316 responses were received via the Credamo platform, which was used for convenience sampling. Following the exclusion of participants who failed attention checks or whose response times fell beyond three standard deviations from the mean, 281 valid responses were available for analysis. The final sample comprised 135 males and 146 females, with a mean age of 22.06 years (SD = 4.46). Regarding educational attainment, 12 participants had completed junior high school or below, 37 had an associate degree, 194 held a bachelor’s degree, 34 had a master’s degree, and 4 had a doctoral degree. This distribution indicates that 232 participants (82.56%) had obtained a university degree or higher, while 49 participants (17.44%) had lower levels of education. Additionally, the mean subjective SES score was 4.91 (SD = 1.31). Prior to participation, all subjects provided informed consent. The priming materials comprised two sets of photographs, one depicting conditions of polluted air and the other depicting conditions of non-polluted air. Each set comprised five photographs taken in Beijing66, with corresponding images of the same locations captured under conditions of pollution and non-pollution (Fig. 2A). The photographs were accompanied by news reports that the participants observed and read. For the images depicting pollution, the news reports were either real pollution reports, accurately reflecting the polluted conditions, or fake positive reports, misleadingly suggesting good air quality despite pollution. For the non-polluted images, the reports were either real non-polluted reports, accurately depicting clean air, or fake negative reports, falsely indicating pollution.

The experimental stimuli were designed to replicate standardized air quality reports issued by China’s National Meteorological Administration and disseminated via state-affiliated or business media67. These reports—characterized by brief textual descriptions paired with visual indicators (e.g., AQI values, weather icons)—are systematically broadcast through TV, business mobile apps, and government portals, ensuring consistent public exposure. While media consumption habits vary, the key content of these reports remains uniform across platforms, aligning with the public’s primary exposure to pollution data, particularly during high-risk periods (e.g., seasonal smog). This design prioritizes ecological validity by mirroring state-mediated information structures. Therefore, information related to air pollution is widely available in the mass media, as well as from other online and offline sources in China, so Chinese citizens should have ample opportunities to acquire information68. For instance, algorithm-driven push notifications on apps like WeChat and Douyin deliver air quality updates to over 94% of Chinese internet users, regardless of their proactive engagement. This aligns with evidence that passive exposure to risk-related information—even when incidental—can shape public perceptions and behaviors. Studies on the mere exposure effect69 demonstrate that repeated encounters with standardized environmental data, whether via traditional media or digital alerts, heighten sensitivity to air quality issues. Our experimental design thus captures a critical dimension of real-world information exposure.

The study employed a between-subjects experimental design comprising two factors: photo type (polluted/non-polluted) and report type (real/fake). The photo type served to distinguish between images of a polluted and a non-polluted nature, while the report type was employed to differentiate between real and fake news reports. The experimental news reports are detailed in Supplementary Methods: Experimental Materials for Study 2.

Participants were randomly assigned to one of four experimental conditions: (1) real non-polluted reports, (2) fake negative reports, (3) real polluted reports, or (4) fake positive reports. The demographic information for each group, please see Supplementary Table 2.

The participants were first asked to provide information regarding their basic demographic characteristics, after which they were exposed to the priming materials that had been assigned to them. Prior to viewing the photographs, participants were provided with the following instructions: “Please undertake a detailed examination of the materials presented below. A series of photographs taken on 25 March 2023 (Saturday) will be presented, accompanied by the corresponding weather forecast for that day. As you view the photos and the weather forecast, it is recommended that you imagine yourself as a resident living in this location. This should entail breathing the air, hearing the sounds, and feeling the temperature. Please envisage how you would spend this Saturday.

Subsequently, participants viewed the five photographs in their respective groups, with each photograph displayed for a minimum of 20 seconds. Following the viewing, participants were required to compose a brief diary entry, with the objective of further immersing themselves in the scenario. The subsequent diary task was introduced with the following instructions: “Imagine that this is your hometown and that it is a Saturday. You are situated within this environment, experiencing the sensory stimuli of your immediate surroundings. Please compose a diary entry in which you describe how you spent this Saturday. Please describe the activities you engaged in from morning to night. Please describe your emotional state. Please provide a written account of at least 100 words.” The diary task used in the experimental design enhanced ecological validity. This method has been shown to improve scenario immersion and ecological validity in media effects research70, as it facilitates deeper engagement with the stimuli while maintaining experimental control.

Once the diary had been completed, participants were invited to take part in a series of questionnaires designed to assess their perceptions of air pollution and their willingness to travel, as well as environmental efficacy, pro-environmental intention and demographic variables like the highest education level and subjective socioeconomic status of participants. Further details can be found in Supplementary Methods: Questionnaires Used in Studies 2 and 3. Prior to completing each questionnaire, participants were instructed to envisage themselves in the scenario depicted in the accompanying photographs and reports.

The data were analyzed using the statistical software package SPSS 22. The initial travel willingness, gender, age, highest education level, SES, environmental efficacy and concerns about environmental pollution of the participants were included as control variables, with travel willingness itself serving as the dependent variable. To investigate the impact of fake positive news reports on individual travel willingness, a 2 (photo type: polluted vs. non-polluted) × 2 (report type: real vs. fake) factorial ANOVA was conducted on individual travel willingness. Furthermore, the PROCESS macro was employed to investigate the potential mediating role of perceived air pollution in this relationship.

Study 3 methods and data analysis

Building on the findings from Study 2, which employed priming materials to explore the influence of fake positive air quality reports on travel behavior and the mediating role of air pollution perception, Study 3 was designed to address the concerns regarding the ecological validity of the prior research. In particular, Study 2 employed imagined scenarios that did not take into account the actual environmental conditions of the participants. To address this limitation, Study 3 employs a field experiment approach, focusing exclusively on environments with actual air pollution, to further examine the influence of fake positive air quality reports on travel behavior and the underlying psychological processes.

The field experiments were conducted in cities experiencing moderate to severe air pollution on designated weekends and weekdays. The participants were residents of the aforementioned cities who were recruited to complete surveys regarding their perceptions of air pollution and travel intentions. The participants were randomly assigned to one of two conditions: a fake positive air quality report or a real pollution report (single-factor design). To ensure the robustness of our findings, field experiments were conducted both on weekends and on a weekday in cities experiencing moderate to severe air pollution. In the weekend experiments, participants were recruited in Xi’an, Shanghai, and Harbin, with the corresponding AQI values recorded on 13 January 2024 (Xi’an: AQI = 241; Shanghai: AQI = 159) and on 27 January 2024 (Harbin: AQI = 328). In parallel, a weekday experiment was conducted on Tuesday, 30 January 2024, in Xi’an (AQI = 202) and Zhengzhou (AQI = 211). The same experimental materials, including the corresponding local photographs, weather forecasts, and priming instructions, were used in both sets of experiments. Further details of experimental materials can be found in Supplementary Methods: Experimental Materials for Study 3. Participants were randomly assigned to one of two conditions: a fake positive air quality report or a real pollution report.

For the weekend experiments, a total of 252 participants completed the study, with the final analysis sample comprising 199 individuals after excluding those who failed attention checks or had inconsistent IP addresses (n fake positive reports = 101, n real reports = 98; 76 males and 123 females, with M age ± SD = 32.04 ± 6.954). The sample demographics for Xi’an, Shanghai, and Harbin are presented in Supplementary Table 3. In the weekday experiment, 120 participants completed the study, with 100 participants retained for analysis (n fake positive reports = 48, n real reports = 52; 33 males and 67 females, with M age ± SD = 28.16 ± 7.96). The sample demographics for Xi’an and Zhengzhou are presented in Supplementary Table 4. The demographic characteristics across both experiments were comparable, ensuring the consistency of our findings. Prior to their participation, all subjects provided informed consent.

The methodology employed in the weekday and weekend experiment was largely consistent with that utilized in Study 2. The initial stage of the procedure involved the provision of demographic data by the participants, prior to their observation of the weather forecast corresponding to their assigned condition. The instructions were as follows: “The weather forecast for the city of Xi’an (or Shanghai/Harbin/Zhengzhou) will now be displayed. Please observe the forecast meticulously, as memory checks will be conducted subsequently. As you observe the presentation, consider the atmospheric conditions, the surrounding sounds, and the current temperature, while contemplating how you would spend the day. Subsequently, the participants viewed the pertinent weather forecast and associated images.

Subsequently, participants were required to compose a daily schedule. The instructions were as follows: “Please close your eyes and imagine the current air quality, the sounds that can be heard in your immediate vicinity, and the temperature outside.” Please consider how you would spend the day. Please provide a comprehensive plan for the day’s activities, delineating the sequence of events from morning to night. Please allocate approximately five minutes to this task and produce a written plan comprising at least 100 words. Following the completion of the plan, participants were invited to complete the Air Pollution Perception and Travel Willingness scales. Therefore, the identical procedure across experiments allowed us to examine whether day-of-week effects might influence the impact of fake positive air quality reports on travel-related decisions.

Data were analyzed using SPSS 22. To assess the effects of fake positive reports on air pollution perception and travel behavior, multiple statistical analyses were conducted for both weekend and weekday’s experiment.

Firstly, a linear regression analysis was conducted, with report type (fake positive vs. real pollution) as the independent variable and air pollution perception as the dependent variable. The analysis was controlled for initial travel willingness, gender, age, city and SES. Subsequently, a t-test was employed to ascertain any differences in travel willingness between participants who had been exposed to fake positive reports and those who had been exposed to real pollution reports. Furthermore, a binary logistic regression was conducted to ascertain the probability of travel based on report type (fake positive vs. real pollution reports), with travel behavior (coded as 0 = no travel, 1 = travel) as the dependent variable. The analysis controlled for the following variables: initial travel willingness, gender, age, city and SES. Subsequently, a multinomial logistic regression was performed to investigate the impact of fake positive reports on the mode of travel (non-motorised, motorised, or no travel, with no travel as the reference category). Finally, mediation models were constructed to investigate the potential mediating role of perceived air pollution between fake positive news and air quality.

Study 4 methods and data analysis

The pilot study and Study 1 employed publicly available data to analyze the correlations between news reports and actual air quality. Studies 2 and 3 proceeded to examine the effects of fake positive air quality reports on individual behavior and perceptions, utilizing both priming and field experiments. Although there is substantial evidence indicating that individual travel behavior has a significant impact on air quality, there is still a gap in the literature with regard to directly linking fake news reports to actual changes in air quality itself. In particular, no study has hitherto investigated whether fake positive air quality reports could exert a measurable influence on air quality. To address this gap, Study 4 employs ABM to directly simulate the relationship between fake positive air quality reports and air quality outcomes, thereby offering a novel approach to understanding the broader environmental implications of media manipulation.

ABM is a computational simulation technique that allows researchers to model systems composed of interacting agents within a defined environment over time71. Each agent operates according to specific behavioral rules and interacts with other agents and the environment, leading to emergent, system-wide behaviors that may not be apparent at the individual level. ABM is particularly well-suited to the exploration of complex, dynamic systems, rendering it an ideal method for the investigation of how individual behaviors influenced by media reports can collectively affect air quality. In this study, the ABM simulation will be conducted using the NetLogo platform, thereby enabling the manipulation of media reporting parameters and the observation of their potential effects on air quality in real time. The simulation was implemented using the NetLogo platform, with a time step of one hour. The study investigates the effects of fake air quality reports in a polluted environment; therefore, the initial AQI was set to 180, corresponding to a moderate pollution level. This setting is justified by China’s official six-tier AQI classification where an AQI above 100 indicates pollution: 0–50 (Excellent), 51–100 (Good), 101–150 (Mild Pollution), 151–200 (Moderate Pollution), 201–300 (Heavy Pollution), and above 300 (Severe Pollution). In the present simulation, the parameter for the environment was selected based on these standard indicators. In real-world scenarios, air quality reports provide an estimation of the AQI for the subsequent 24 hours; therefore, in our model, the air quality is reported at 00:00 hours as the anticipated average AQI for the following day. Two distinct global environments were developed: one representing real reports and another representing artificially positive (fake) air quality reports.

To ensure ecological validity, the model’s AQI parameterization is grounded in real-world air quality fluctuations observed in Beijing, a major urban center frequently experiencing moderate to severe pollution. For the real-report environment, hourly true AQI data from the first ten days of each month in Beijing during 2023 (a total of 120 days) were collected. The difference between the AQI recorded at 00:00 and the average AQI over the subsequent 24 hours (defined as the “AQI fluctuation value”) was calculated for each day. Outliers (values beyond three standard deviations) were removed from the dataset to minimize the impact of extreme fluctuations. The remaining data produced a normally distributed fluctuation with a mean of 1.924 and a standard deviation of 11.538, and a range from −34.792 to 42.125. In the model, the “reported AQI” in the real-report environment is defined as the AQI at 00:00 minus an independently sampled fluctuation value drawn from the N(1.924, 11.538) distribution, but constrained within the interval [–34.792, 42.125].

In contrast, under the fake positive report environment—motivated by our focus on the impact of overly optimistic reporting—the reported AQI is modified when the actual AQI exceeds 100. Specifically, if the real AQI is greater than 100, the reported AQI is randomly drawn from a uniform distribution over the interval (85, 100], and if the real AQI is 100 or lower, it remains unchanged. This formulation is mathematically summarized as follows:

$$\left\{\begin{array}{c}{reported\; AQI}={real\; AQI},{if\; realAQI}\, \le \, 100\\ {reported \; AQI} \sim U\left(85,100\right),{if\; realAQI} \, > \, 100\end{array}\right.$$ (2)

The initial travel willingness of 580 participants in Studies 2 and 3 was collected and analyzed, yielding a mean of 0.665 and a standard deviation of 0.258. Subsequently, 10,000 agents were initially created, with each agent assigned an initial travel probability that followed a normal distribution (mean = 0.665, SD = 0.2582, range = 0–1). The simulation’s agents represent individual decision-makers regarding travel. The baseline travel willingness data were derived from the participants that we collected in Studies 2 and 3, where 580 participants’ initial travel intentions were collected and transformed from a 1–5 scale to a normalized 0–1 scale. The analysis revealed an average travel probability of 0.665 with a standard deviation of 0.258. In the simulation, a total of 10,000 agents were generated. Therefore, based on the experimental results that we collected, each agent was assigned an initial travel probability sampled from a normal distribution with a mean of 0.665 and a standard deviation of 0.258. This probability was truncated to lie within the interval [0, 1]. In addition, agent movement decisions during the day were modeled stochastically, with each agent randomly selecting a time point to determine whether to travel. Once an agent decided to travel, a return journey was always scheduled. The specific distributions for departure and return times were based on empirical data derived from studies of resident travel patterns on non-working days56 (see Supplementary Fig. 7).

Prior research has consistently demonstrated that individuals tend to avoid travel on days with poor air quality. For example, ref. 72 used social media check-in data to show that increases in pollutants (PM₂.₅, NO₂, SO₂, PM₁₀) are associated with significant reductions in various travel activities, while ref. 73 reported that the proportion of outdoor activities decreases by 9.7%–57% on polluted days. In line with these findings, our model posits that an agent’s decision to travel is influenced both by its initial travel probability and by its perception of air quality.

The maximum travel probability is given by the agent’s initial travel probability, and the actual probability decreases as the perceived air quality worsens. The travel probability update rule was formulated as follows:

$${P}_{{\rm{travel}}}=\frac{{P}_{{\rm{initial\; travel}}}* (1+{e}^{\frac{0-200}{60}})}{1+{e}^{\frac{{perceivedAQI}-200}{60}}}$$ (3)

The perceived AQI was a weighted combination of real AQI and reported AQI:

$${perceivedAQI}=\alpha \, * \, {AQI}+\left(1-\alpha \right)\, * \, {reportedAQI}$$ (4)

Here, the parameter α represents an agent’s environmental sensitivity—that is, the extent to which the agent’s perception of air quality is influenced by the actual AQI relative to the reported AQI.

The parameter α represents the agent’s sensitivity to the environment, with higher sensitivity indicating a greater impact of actual air quality on perceived air quality. The value of parameter α is estimated based on the field experiment data that we collected in Study 3. Initially, the scale of air pollution perception from the field experiment, ranging from 1 to 7 points, was converted to a scale of 0–100 points. Table 4 presents the air pollution perception levels for each city group after conversion. For instance, considering Xi’an during the weekend, where the AQI was 241, the air pollution level perceived by the fake positive air quality reports group was 55.42, while that perceived by the real reports group was 78.38. This suggests that fake positive air quality reports reduced the air pollution perception level of the agents by 29.29% (\(\frac{{perceivedAQ}{I}_{{real}}-{perceivedAQ}{I}_{{fake}}}{{perceivedAQ}{I}_{{real}}}=\frac{78.38-55.42}{78.38}\)).

Table 4 The air pollution perception levels of each city group after the conversion Full size table

Under real report conditions, the reported AQI is identical to the actual AQI, leading to a perceived AQI of:

$${perceivedAQ}{I}_{{real}}={AQI}$$ (5)

Under fake reports conditions,

$${perceivedAQ}{I}_{{fake}}=\alpha * {AQI}+\left(1-\alpha \right)* {reportedAQI}$$ (6)

Therefore, we can derive the following equations:\(-\)

$$\frac{{{perceivedAQI}}_{{real}}-{perceivedAQ}{I}_{{fake}}}{{perceivedAQ}{I}_{{real}}}=\frac{{AQI}-[\alpha * {AQI}+(1-\alpha )* {reportedAQI}]}{{AQI}}$$ (7)

In the field experiment, regardless of the actual AQI, all cities consistently received fake positive reports indicating “good” air quality, corresponding to an air quality index range of (50, 100]. Therefore, in Eq. 1, the reported AQI is substituted with 75. Based on the field experiment results, the equation is constructed as follows:

$$\left\{\begin{array}{c}{\frac{241-[{\alpha }_{1}*241+(1-{\alpha }_{1})\,*\,75]}{241} = \frac{78.38-55.42}{78.38} {\text{X}}{\mbox{‘}}{\text{ian}}} \\ {\frac{159-[{\alpha }_{2}\,*\,159+(1-{\alpha }_{2})\,*\,75]}{159} = \frac{71.08-55.88}{71.08} {\rm{S}}{\rm{h}}{\rm{a}}{\rm{n}}{\rm{g}}{\rm{h}}{\rm{a}}{\rm{i}}} \\ {\frac{328\,-[{\alpha }_{3}\,*\,328\,+(1-{\alpha }_{3})\,*\,75]}{328} = \frac{80-56.25}{80} {\rm{H}}{\rm{a}}{\rm{r}}{\rm{b}}{\rm{i}}{\rm{n}}} \\ {\frac{202\,-[{\alpha }_{4}\,*\,202+(1-{\alpha }_{4})\,*\,75]}{202} = \frac{74.24-47.06}{74.24} {\text{X}}{\mbox{‘}}{\text{ian}}} \\ {\frac{211\,-[{\alpha }_{5}\,*\,211\,+(1-{\alpha }_{5})\,*\,75]}{211} = \frac{70-54.3}{70} {\rm{Z}}{\rm{h}}{\rm{e}}{\rm{n}}{\rm{g}}{\rm{z}}{\rm{h}}{\rm{o}}{\rm{u}}}\end{array}\right.$$ (8)

Hence, \({\alpha }_{1}\approx 0.575,{\alpha }_{2}\approx 0.595,{\alpha }_{3}\approx 0.615,{\alpha }_{4}\approx 0.418,{\alpha }_{5}\approx 0.652\). Recognizing individual variations in environmental sensitivity, this study sets the distribution of α among agents to follow a normal distribution with a mean value randomly selected within the range [0.418, 0.652] and a standard deviation of 0.2.

When agents decide to travel, they must select a transportation mode. Prior research indicates that, under polluted conditions, residents tend to alter their travel modes to reduce exposure to contaminants. For instance, ref. 47 found that under moderate pollution, the proportion of non-motorized travel (such as walking or cycling) decreases while the usage of public transport or cars increases. Moreover, within the group choosing motorized travel, further mode shifts toward car usage are observed as pollution levels escalate.

Travel distance is a primary factor in determining the mode choice. Based on data from ref. 74, which characterizes weekend travel behaviors, our model assigns travel distances to agents according to the following distribution:

26.55% of agents are assigned a travel distance randomly drawn from the interval (0, 5] km.

33.32% of agents are assigned a travel distance from (5, 10] km.

21.5% of agents are assigned a travel distance from (10, 15] km.

9% of agents are assigned a travel distance from (15, 20] km.

9.63% of agents are assigned a travel distance from (20, 30] km.

These percentages are derived from empirical observations (ref. 74), ensuring that our simulation reflects the travel behavior distribution observed in real-world scenarios.

The probability of selecting a non-motorized mode is assumed to be a function of travel distance. According to the 2022 Guangzhou Transportation Development Report, average travel distances vary by transport mode: the average one-way distance for cycling is ~1.5 km (with 80% of trips under 2 km), whereas taxi and ride-hailing services average 7.0 km and 9.0 km, respectively, and private car trips average 12.2 km. Based on these findings, we model the initial probability of choosing non-motorized travel as a decreasing function of travel distance. Specifically, we define this probability by:

$${P}_{{initial\; choice\; of\; non}-{motorized\; travel}}=\frac{1+{e}^{-3}}{1+{e}^{({distance}-3)}}$$ (9)

As distance approaches zero, this probability trends towards one, reflecting a high likelihood of choosing non-motorized travel for very short trips. Conversely, as distance increases, the likelihood of choosing a non-motorized mode declines.

Given that motorized travel is partitioned into public transportation and car usage, and based on the 2022 Guangzhou Transportation Development Report, which indicates that 47.2% of motorized trips in urban areas are via public transport, the initial mode selection is set as follows:

Public transport: 0.472 × (1−P initial choice of non-motorized travel );

Car travel: 0.528 × (1−P initial choice of non-motorized travel ).

Each agent selects an initial travel mode based on these probabilities, and the agent’s return mode is assumed to be consistent with the mode chosen for departure.

Empirical studies have shown that air pollution significantly influences travel behavior, as individuals attempt to reduce pollutant exposure. For instance, during severe pollution episodes, residents are more likely to prefer door-to-door travel modes (e.g., private cars) over modes that involve prolonged outdoor exposure, such as walking or waiting for public transport75. Data from that study revealed that the proportion of people using taxis on weekends increased by 10.17% during heavy pollution red alert periods compared to normal days.

Further, ref. 47 conducted a survey comparing mode choice under varying pollution levels. The findings showed that:

Under moderate pollution (AQI 150–200): The proportion of individuals choosing walking or cycling decreased by 6.5%; Car travel increased by 6.5%; Public transport usage remained unchanged.

Under heavy pollution (AQI 200–300): Walking and cycling decreased by 14%; Car travel increased by 14.5%; Public transport usage decreased slightly by 0.5%; These empirical changes in travel behavior provided the basis for the agent-based model’s mode-switching rules:

If perceived AQI ≤ 150:

Agents retain their initially selected mode.

If perceived AQI is between 150 and 200:

For agents traveling more than 1.5 km and initially choosing non-motorized modes, there is a 6.5% probability of switching to car travel.

If perceived AQI is between 200 and 300: Agents using non-motorized modes (with distance >1.5 km) switch to car travel with a 14% probability; Agents using public transport switch to car travel with a 0.5% probability.

If perceived AQI > 300:

Agents on non-motorized modes (with distance >1.5 km) switch to car travel with a 20% probability;

Public transport users switch to car travel with a 1% probability.

Finally, temporal constraints were also implemented: since most public transport systems in China operate between 5:00 AM and 11:00 PM, agents who initially chose public transport but travel during non-operational hours were automatically reassigned to car travel.

These switching probabilities and behavioral thresholds were directly based on observed empirical data and transport reports, ensuring the model’s parameters reflect realistic travel patterns and provide a robust foundation for simulation validity. In our simulation, the ambient air quality—as measured by the AQI—is primarily affected by the travel behaviors of the agents. In particular, motorized travel by agents leads to an increase in the AQI. This relationship was estimated using real-world data from Guangzhou. Specifically, we collected data on public transport usage (i.e., daily per capita public transport trips) in Guangzhou from January 2015 to December 2021, covering 84 months. After controlling for factors such as average temperature, relative humidity, rainfall, and sunshine duration, the per capita public transport trip count was found to significantly predict AQI (B = 20.786, t = 2.793, p = 0.007). Hence, the coefficient representing the impact of public transport travel on AQI is set to 20.786.

Considering that public transport vehicles (e.g., buses) typically have a capacity approximately four times that of a private car, the pollution load contributed by private car travel is assumed to be four times that of public transport. As a result, the coefficient for private car travel is set to 83.144.

Based on these empirical findings, the AQI update rule in the model is formulated as follows:

$${AQ}{I}_{t+1}= \, {AQ}{I}_{t}+20.786* \frac{{Pquantit}{y}_{t+1}}{{N}_{{agent}}}+83.144* \frac{{Cquantit}{y}_{t+1}}{{N}_{{agent}}}\\ -{{\rm{Diffusion}}}_{t+1}* {AQ}{I}_{t}$$ (10)

In the Eq. 10, ‘t’represents the time step, \({Pquantity}\) represents the number of agents choosing public transport for t ravel (both outbound and return), \({Cquantity}\) represents the number of agents opting for car travel, and ‘N agent ’ represents the total number of agents. ‘Diffusion’ signifies the AQI diffusion at each time step.

To account for natural pollutant dispersion and minimize the direct effects of travel on AQI, we determined the AQI diffusion rate using 720 hours (30 days) of AQI data collected during the final month of the Wuhan lockdown—from March 9, 2020, to April 7, 2020. This period was selected because:

In the initial phase of the lockdown, large-scale movements such as the construction of emergency hospitals (e.g., Huoshenshan and Leishenshan) and increased engineering vehicle traffic could skew travel-related pollution effects.

The Chinese Lunar New Year period (January and February) often shows anomalous patterns due to fireworks and related activities, which would also affect AQI readings.

For each hourly time step in this 30-day period, the AQI diffusion rate was calculated as:

$${\rm{Diffusion\; Rate}}=\left(\frac{{{\rm{A}}{\rm{QI}}}_{t}-{{\rm{AQI}}}_{t+1}}{{AQ}{I}_{t}}\right)$$ (11)

This diffusion rate reflects the degree to which the ambient AQI naturally declines from one hour to the next. In the simulation, to account for natural variability, at each time step, the diffusion coefficient is randomly selected from a uniform range between 0.9 and 1.1 times the measured diffusion rate. For instance, in the diffusion rate data for Wuhan, the fourth-hour data is 0.067. Consequently, in the model, the diffusion for the fourth time step takes a random value within the range of [0.9 × 0.067, 1.1 × 0.067]

Summary of data sources and parameter settings is as following:

Public transport impact : The coefficient of 20.786 for public transport per capita trip impact is based on Guangzhou’s public transport data spanning 2015–2021, controlled for meteorological factors.

Car travel impact : The coefficient of 83.144 is derived from the public transport coefficient scaled by the relative vehicle capacity (with private cars generating approximately one-fourth of the passenger capacity of public vehicles).

AQI diffusion: The diffusion coefficient, reflecting the natural decay of AQI over time, is estimated from 720 hours of AQI data during the lockdown in Wuhan, with a random variation of ±10% to simulate environmental variability.

Table 5 summarizes all key model parameters, their numerical values, and the corresponding empirical data sources or literature references supporting their selection.

Table 5 The model parameters Full size table

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.