Main

There is near-universal consensus (97–99.9%) in the peer-reviewed scientific literature that the climate is changing as a result of human activity1,2,3. However, the public often considerably underestimates this consensus4, which is evident even in recent data from six European countries (estimates range from 65% in the United Kingdom to 71% in Ireland)5. These misperceptions have several negative consequences that can impede efforts to mitigate climate change6,7. People who underestimate the scientific consensus on climate change are less likely to believe in (human-caused) climate change, to worry about climate change and to support public action on climate change compared to those who perceive the scientific consensus more accurately6,7,8,9,10.

On a more optimistic note, these misperceptions offer an opportunity for low-cost interventions that can be applied at scale. In recent years, communicating the message that 97% of climate scientists agree that human-caused climate change is happening has been one of the most studied strategies to correct misperceptions of the scientific consensus and influence climate change attitudes11,12. A large body of experimental studies supports the notion that communicating this scientific consensus can shift consensus perceptions, beliefs in the reality of climate change and human activity as its primary cause, climate change worry and support for public action13,14,15,16,17,18,19,20,21,22,23,24,25. In addition, two meta-analyses show that informing people about the scientific consensus can substantially reduce consensus misperceptions (Hedge’s g = 0.56)26 and—to a smaller extent—increase several proclimate attitudes (Hedge’s g = 0.09−0.12)26,27, namely personal beliefs in and worry about climate change. One of these meta-analyses shows that messaging interventions—including but not limited to scientific consensus messages—had no effect on support for public action27.

However, the current evidence base relies overwhelmingly on data from Western, democratic, high-income countries. This includes most studies from the United States and other native English-speaking countries (United States, k = 18; Australia, k = 2; New Zealand, k = 1; k is the number of studies in the most recent meta-analysis26) as well as two exceptions, one from Japan15 and one from Germany28. Given that climate change requires global action and cooperation29, the lack of knowledge on whether and to what extent scientific consensus messages can reduce consensus misperceptions and shift climate change attitudes outside the United States and the few studied Western nations is a pivotal gap that needs addressing12,30,31,32,33. The present study aims to extend the evidence on scientific consensus messaging to include 27 countries on six continents (Asia, Africa, Australia, Europe and North and South America). As the public in many countries underestimates the scientific consensus on the reality of human-caused climate change5,34, a message that emphasizes this consensus is expected to increase consensus perceptions and proclimate attitudes in a diverse, multicountry sample.

Beyond the consensus on the reality of human-caused climate change5,24,25,28,33,35, climate experts emphasize very high certainty of the adverse consequences of climate change and the urgency of climate action to curb these impacts in the Sixth IPCC report35. In line with this, 88% of surveyed IPCC authors report that they think climate change constitutes a crisis36. To align communication about the scientific consensus with these more up-to-date climate science assessments and potentially improve its effectiveness, we test a combined message communicating the 97% consensus that human-caused climate change is happening in addition to the 88% agreement that climate change is an urgent matter (a crisis). Such an updated message that emphasizes the negative impacts of climate change and implies the need for public action might prove more effective at increasing belief in climate change as a crisis, climate change worry and support for public action than the classic message28. This might be especially useful in contexts where the public consensus on the reality of climate change is high but a substantial proportion still doubts the urgency of climate action28. Initial support for the effectiveness of such an updated message comes from the finding that communicating the social consensus on the urgency of climate action increases support for public action on climate change more than a social consensus message on the reality of climate change37.

While we expect both scientific consensus messages to be overall effective, their effectiveness may depend on several individual-level (message familiarity, trust in climate scientists and political ideology) and country-level (individualism–collectivism and power distance38) characteristics. Regarding individual-level characteristics, Bayesian approaches to information processing postulate that people are generally motivated to be accurate and update their beliefs according to relevant information39, especially when the new information challenges previous beliefs40,41. Previous research shows that scientific consensus messages on climate change are more effective for individuals with lower initial consensus perceptions24,26. In line with this reasoning, scientific consensus messages might be less effective when individuals are more familiar with the message42. Moreover, according to some motivated cognition accounts43,44, people whose previous worldviews and/or identities are not aligned with a given issue or its implications might not be receptive to scientific (consensus) information on the topic or even revise away from the scientific evidence. In line with this view, trust in climate scientists and political ideology might moderate the effectiveness of any consensus intervention, with those lower in trust and more politically on the right6 being less likely to accept the scientific consensus on climate change. To date, however, the evidence for such effects is mixed. Although there is some suggestive evidence that those with higher trust in science/scientists might be more receptive to scientific consensus messages15,45,46, strong evidence using larger samples is lacking to support these patterns. With respect to political ideology, studies show smaller13, similar19,22, larger23,24,31 and even backfiring47 effects of scientific consensus interventions among conservatives compared to other political groups and the most recent meta-analyses found no moderating effect of political ideology26,48.

If the effectiveness of the intervention varies across countries, this could be due to country-level differences in individualism–collectivism and power distance38. Consensus information is a form of a descriptive norm, and norms have been shown to be more predictive of support for environmental policy in collectivistic cultures compared to individualistic ones49. Further considering that scientific consensus messages are an expert norm, it is possible that they are more effective in cultures with higher power distance, where greater weight is given to source expertise/authorities50,51,52. However, more precise predictions are difficult because of the lack of studies on scientific consensus interventions outside the United States12,30,31,32,33.

In sum, this study expands the evidence base on the effectiveness of scientific consensus messaging on climate change in several ways (Table 1). We first test the effectiveness of the classic (reality of human-caused climate change) and the updated (reality of human-caused climate change and crisis) consensus messages across 27 countries. To do so, we focus on the main effects of the interventions on perceptions of the reality consensus and crisis agreement, personal climate change beliefs (reality, human causation and crisis), climate change worry and support for public action in a between-participants design (H1a–e and H2a–g). Second, we test whether supplementing the scientific consensus on the reality of climate change with the broad scientific agreement that climate change constitutes a crisis can further increase personal belief in climate change as a crisis, climate change worry and support for public action (H3a–c). Third, this study provides an opportunity for a comprehensive, high-powered investigation into individual-level characteristics, such as message familiarity, trust in climate scientists and political ideology, which might moderate the effectiveness of both interventions on reality consensus and crisis agreement perceptions (H4a–c and Q5). Last, we explore whether and to what extent the effectiveness of both messages varies across countries (Q6) and whether country-level characteristics, such as individualism–collectivism and power distance, can predict potential variation (not preregistered). Tests corresponding to each hypothesis are summarized in Supplementary Table 1.

Table 1 Overview of preregistered research questions and hypotheses
Full size table

Testing climate change consensus messaging across various countries has ramifications at two levels. Theoretically, it addresses generalizability concerns about the effects of expert norm communication. In addition, well-powered moderation analyses enable us to address conflicting theoretical standpoints (Bayesian information processing and motivated cognition) about human cognition in the face of contested scientific evidence. On a translational level, a messaging approach that is effective across diverse contexts and audiences would provide a general guideline for climate change communication and could thus facilitate a more rapid move toward urgently needed climate policies. If the effectiveness varies according to individual and country-level characteristics, this could inform targeting specific audiences within countries and/or calibrating consensus messaging interventions to different country contexts. However, if consensus messaging is ineffective when tested across a diverse set of countries, this would signal the limits of this intervention and the need to focus on different strategies to mobilize support for climate action.

Results

Participants

The analytical sample consisted of 10,527 total participants. Country sample sizes ranged from 9 (Lebanon) to 634 (Germany). In terms of gender, women were slightly more represented than men (female, 57%; non-binary or prefer not to say, 1%). Most of the sample held a university degree (68.1%) and lived in urban areas (81.6%). About one-third of the sample were studying at the time of data collection (33.7%). Demographic overviews per country are presented in Table 2, while population descriptions for each country are available in Supplementary Table 5.

Table 2 Demographic characteristics across countries
Full size table

Our recruitment approach is described in detail in the Methods section ‘Participant recruitment’ and further details about the sample are reported in the section ‘Sample details’.

Overview

The preregistered data analyses focus on three broad questions. First, we test whether the classic and the updated scientific consensus messages can reduce misperceptions and increase climate change beliefs, worry and support for public action (Q1 and Q2 in Table 1). Second, we investigate whether the updated message is more effective than the classic one at shifting personal belief in climate change as a crisis, climate change worry and support for public action (Q3). Third, we test whether the effectiveness of both interventions varies by several individual and country-level characteristics as well as by country (Q4–Q6 and further exploratory analyses).

To do so, we rely on Bayesian model-averaging approaches53,54. Our analyses (Supplementary Information Section 1; Open Science Framework (OSF) https://osf.io/z6quh/) take into account the uncertainty regarding the model structure (for example, constant versus heterogeneous intervention effects across countries). Furthermore, we used Bayesian mixed-effects linear and ordinal regressions, with participants (level 1) nested in countries (level 2), controlling for relevant demographic characteristics, including age, gender, university degree and political ideology (for further details, see Methods section ‘Data analysis’). Instead of treating the one-item outcomes measured on seven-point scales (that is, climate change beliefs, worry and support for public action) as continuous55, we apply cumulative probit regression models that appropriately treat the data as ordinal and can account for skewed response patterns56. We further specify informed hypotheses based on earlier research which allow us to test the presence versus absence of even small intervention effects57. All analyses are based on group differences in post-intervention outcomes. In sum, our analytical approach enables us to draw robust and valid conclusions about the data.

To simplify the interpretation and integration of the results into the existing literature, we supplement the informed Bayes factors with meta-analytic estimates of the overall intervention effect (Cohen’s d) and the between-country heterogeneity (τc) including confidence intervals (CI) derived from frequentist random-effects meta-analyses (Supplementary Information Section 2.1). As these estimates do not correspond to the specified models used for assessing the presence versus absence of intervention effects and can thus diverge from them, especially in the case of ordinal outcomes, the confidence intervals should not be interpreted as statistical significance tests. For future meta-analyses, we provide summary tables of the results for each outcome per country (Supplementary Table 1).

Misperceptions of the reality consensus and crisis agreement

In this section, we provide misperceptions of the reality consensus and crisis agreement per country before message exposure. These descriptives are unlikely to be representative of misperceptions per country due to the convenience sampling approach. Instead, they demonstrate that misperceptions are present in our samples—a prerequisite for consensus messaging to be effective.

Across all 27 countries (n = 10,527), the scientific consensus that human-caused climate change is happening (97%) is underestimated by, on average, −12.11% (95% CI (−12.43, −11.80)). This underestimation ranges from −20.91% (95% CI (−22.80, −19.03)) in the Chinese sample to −7.54% (95% CI (−8.53, −6.56)) in the German sample (Fig. 1). In total, 72.2% (95% CI (71.3, 73.0)) of participants underestimate this consensus, ranging from 57.5% (95% CI (52.2, 62.6)) in the US sample to 83.7% (95% CI (79.9, 87.0)) in the Chinese sample (Supplementary Information Section 2.2).

Fig. 1: Mean pre-intervention (mis)perceptions of the scientific consensus on the reality of climate change and agreement on climate change as a crisis per country sample.
figure 1

The error bars represent the 95% CI for each country. The dashed blue line represents the actual scientific reality consensus (97%). The dashed red line represents the actual scientific crisis agreement (88%).

Full size image

The scientific agreement that climate change constitutes a crisis (88%) is slightly underestimated by, on average, −4.14% (95% CI (−4.47, −3.81)). This ranges from an underestimation of −13.18% (95% CI (−15.07, −11.28)) in the Chinese sample to an overestimation of 1.18% (95% CI (0.14, 2.22)) in the German sample (Fig. 1). However, this scientific agreement is not consistently underestimated. A total of 44.5% participants across all countries (95% CI (43.6, 45.5)) underestimate the crisis agreement, ranging from 29.2% (95% CI (25.7, 32.9)) in the German sample to 69.5% (65.0, 73.7) in the Chinese sample (Supplementary Information Section 2.2).

Consistent with previous studies17,19,20,24,25,58, we find substantial misperceptions of the reality consensus, now in a diverse 27-country sample, indicating a gap between the actual and the perceived scientific consensus that could be reduced with consensus messaging interventions. Although, on average, people also underestimate the crisis agreement, these misperceptions are relatively small.

Effectiveness of the classic scientific consensus message

We first examine whether perceptions of the reality consensus, climate change beliefs (reality and human causation), worry and support for public action are higher in the classic scientific consensus (n = 3,488) compared to the control condition (n = 3,512).

Controlling for pre-intervention perceptions of the reality consensus, we find extremely strong support for H1a that post-intervention perceptions of the reality consensus are higher and thus more accurate in the classic scientific consensus compared to the control condition (BF+0 = 2.01 × 1012; Fig. 2). The Bayes factor implies that the data are 2.01 × 1012 more likely under the hypothesis that participants in the classic consensus condition perceive the reality consensus as higher than those in the control condition (H+) compared to the hypothesis that there is no difference between conditions (H0). This corresponds to substantial effects between conditions across all countries, with Cohen’s d = 0.47 (95% CI (0.41, 0.52)). However, these analyses also show extremely strong support for between-country heterogeneity (BF10 = 1.49 × 106; τc = 0.06, 95% CI (0.00, 0.15)), meaning that the effect on reality consensus perceptions is positive in all countries but varies across countries in terms of magnitude. Despite this strong evidence, the confidence interval includes zero and spans a wide range of values, indicating that the magnitude of heterogeneity is uncertain.

Fig. 2: Effects of the classic and updated scientific consensus intervention on all post-intervention outcomes.
figure 2

a, shows that both the classic and the updated consensus messages increase perceived scientific consensus and agreement. The upper vertical line represents the actual scientific reality consensus (97%); the lower vertical line represents the actual scientific crisis agreement (88%). b, shows that both messages increase belief in climate change, its human causation and climate change worry but not support for public action. The updated condition does not further increase belief in crisis, worry or support for public action. a,b, the means of each outcome per condition and the 95% CI (which are too small to be visible in panel a) are presented on the left-hand side. On the right-hand side, Bayes factors for between-group comparisons are shown. We only indicate Bayes factors for the tested hypotheses, not all comparisons. ‘Classic’ refers to the message communicating the scientific consensus on the reality of climate change. ‘Updated’ refers to the message communicating the scientific consensus on the reality of climate change and the scientific agreement on climate change as a crisis.Across both panels, sample sizes for all outcomes are: nclassic = 3,488; nupdated = 3,527; and ncontrol = 3,512, except for belief in human causation of climate change, where: nclassic = 3,443; nupdated = 3,490; and ncontrol = 3,464. These results are reported in the sections ‘Misperceptions of the reality consensus and crisis agreement’, ‘Effectiveness of the updated scientific consensus message’ and ‘Effectiveness of the updated scientific consensus message’. Complete results are described in the Results and the Supplementary Results.

Full size image

In line with H1b and H1c, we find strong and extremely strong support that people believe more in climate change (BF+0 = 25.51) and human activity as its primary cause (BF+0 = 467.86; Fig. 2) after being exposed to the classic (H1c, n = 3,443) compared to the control message (H1c, n = 3,464). Both intervention effects are small (reality: Cohen’s d = 0.06, 95% CI (0.01, 0.12); human causation: Cohen’s d = 0.10, 95% CI (0.04, 0.15)), with evidence against any between-country heterogeneity (reality: BF10 = 4.78 × 10−6, τc = 0.07, 95% CI (0, 0.20); human causation: BF10 = 1.51 × 10−5, τc = 0.06, 95% CI (0, 0.17)).

Similarly, there is moderate support for a small but consistent effect of the classic scientific consensus intervention on climate change worry (H1d; BF+0 = 5.03; Cohen’s d = 0.05, 95% CI (−0.01, 0.10); Fig. 2), with evidence against between-country heterogeneity (BF10 = 9.93 × 10−10, τc = 0.07, 95% CI (0, 0.21)). In contrast to H1e, we find weak evidence against an effect of the classic scientific consensus message on support for public action (BF+0 = 0.62; Cohen’s d = 0.02, 95% CI (−0.03, 0.08); Fig. 2), with evidence against between-country heterogeneity (BF10 = 2.93 × 10−10, τc = 0.06, 95% CI (0, 0.20)).

Additionally, we explore whether the classic message influences belief in climate change as a crisis (not preregistered). We find strong evidence that the classic scientific consensus message increases belief in climate change as a crisis (BF+0 = 35.80, Cohen’s d = 0.06, 95% CI (0.01, 0.11)), with evidence against between-country heterogeneity (BF10 = 5.18 × 10−10, τc = 0, 95% CI (0, 0.12)).

Effectiveness of the updated scientific consensus message

We next compare the updated consensus (n = 3,527) to the control (n = 3,512) and the classic consensus condition (n = 3,488). Controlling for pre-intervention perceptions of the reality consensus and crisis agreement, respectively, we find extremely strong support for H2a and H2b that perceptions of both the reality consensus (BF+0 = 2.12 × 1012; Cohen’s d = 0.47; 95% CI (0.41, 0.52); Fig. 2) and the crisis agreement (BF+0 = 1.54 × 105; Cohen’s d = 0.23; 95% CI (0.16, 0.31); Fig. 2) are higher in the updated compared to the control condition, with substantial evidence for relatively small between-country heterogeneity (reality consensus: BF10 = 1.19 × 103, τc = 0.05, 95% CI (0, 0.14); crisis agreement: BF10 = 1.53 × 108, τc = 0.15, 95% CI (0.09, 0.23)). For climate change beliefs (H2c and H2d), worry (H2f) and support for public action (H2g), the effects of the updated condition are consistent with those of the classic condition in terms of evidence strength and effect size (Fig. 2 and Supplementary Information Section 2.3). Contrary to H2e, there is only weak support for an effect of the updated message on belief in climate change as a crisis (BF+0 = 1.80; Cohen’s d = 0.04, 95% CI (−0.01, 0.09); Fig. 2), with evidence against between-country heterogeneity (BF10 = 9.15 × 10−09, τc = 0.06, 95% CI (0, 0.14)).

Comparing the updated and classic consensus condition (H3a–c) reveals consistent moderate-to-strong evidence for no added benefit of the updated message in terms of crisis belief, worry and support for public action (crisis belief: BF+0 = 0.09, Cohen’s d = −0.02, 95% CI (−0.07, 0.03); worry: BF+0 = 0.11, Cohen’s d = −0.01, 95% CI (−0.06, 0.04); public action: BF+0 = 0.25, Cohen’s d = 0, 95% CI (−0.04, 0.05)), all with extremely strong evidence against any heterogeneity across countries (crisis belief: BF10 = 1.72 × 10−8, τc = 0.07, 95% CI (0, 0.14); worry: BF10 = 9.84 × 10−11, τc = 0.06, 95% CI (0, 0.17); public action: BF10 = 3.82 × 10−11, τc = 0, 95% CI (0, 0.08)). We also find extremely strong exploratory evidence that participants are more confident in their agreement perceptions after seeing the updated compared to the classic consensus message (not preregistered; BF10 = 4.90 × 108; Cohen’s d = 0.44; 95% CI (0.38, 0.51); between-country heterogeneity: BF10 = 73.60, τc = 0.10, 95% CI (0.02, 0.19)).

Moderators of consensus messaging effects on misperceptions

We next investigate whether message familiarity, trust in climate scientists and political ideology moderate (1) the effectiveness of the classic consensus message on reality consensus perceptions and (2) the effectiveness of the updated consensus message on perceptions of the reality consensus and crisis agreement, compared to the control condition. In all analyses, we control for pre-intervention perceptions of the reality consensus and/or crisis agreement, depending on the respective outcome. We do not estimate cross-country heterogeneity for any of the moderation effects due to limited statistical power for these tests.

In line with H4a, there is extremely strong evidence (BF10 = 1.43 × 1016) that the classic consensus message is more effective at correcting misperceptions for those who reported being less familiar with the message before the study. Contrary to H4b, we find extremely strong evidence (BF10 = 0) against the assumption that the classic consensus intervention is more effective for people who trust climate scientists more. Further unplanned exploratory analyses suggest a three-way interaction. Those who trust climate scientists more have higher and more accurate perceptions of the scientific consensus before the intervention (r = 0.20), leaving relatively less room for updating beliefs in the experimental conditions (Supplementary Information Section 2.4; Fig. 1a). Finally, we find weak evidence (BF10 = 1.89) for a moderating effect of political ideology. In contrast to H4c, the intervention seems slightly more effective for people on the right of the political spectrum. As for trust in climate scientists, this moderation probably also depends on the higher degree of possible improvement among right- compared to left-leaning participants (Supplementary Information Section 2.4; Fig. 1b).

Comparing the updated and control messages (Q5), there is extremely strong support for a moderation by message familiarity, trust in climate scientists and political ideology. These moderations are consistent with the ones observed for the effectiveness of the classic intervention, such that people who are less familiar with the message (reality consensus: BF10 = 2.53 × 1022; crisis agreement: BF10 = 2.76 × 1019), trust climate scientists less (reality consensus: BF10 = 3.34 × 1050; crisis agreement: BF10 = 2.60 × 1090) and are more right-leaning (reality consensus: BF10 = 2.52 × 105; crisis agreement: BF10 = 3.10 × 1016) update their perceptions of the reality consensus and crisis agreement more. Similar to the interactions with trust in climate scientists and political ideology for the classic consensus message, additional exploratory analyses suggest that these moderations are qualified by pre-intervention perceptions of the reality consensus and crisis agreement. People with more right-leaning ideology and lower trust in climate scientists had lower initial consensus/agreement perceptions and thus a wider margin to update (Supplementary Information Sections 2.4; Fig. 1c,d).

Additional exploratory analyses

As specified in the preregistration, we run several exploratory analyses. Rerunning all main confirmatory analyses without demographic control variables yields highly similar results (Supplementary Information Section 2.5). Further moderation analyses show extremely strong evidence that the classic compared to the control message is more effective at correcting misperceptions of the reality consensus for people with lower (that is, more incorrect) pre-intervention perceptions of this consensus (BF10 = 6.87 × 10211). We also find strong evidence for a similar moderation on belief in the reality of climate change—the effect of the classic consensus (versus control) message was larger for those with lower pre-intervention perceptions of the reality consensus, such that they increased their belief in the reality of climate change to a greater extent (BF10 = 12.47). We find weak evidence against the same moderation for belief in the human causation of climate change (BF10 = 0.69), climate change worry (BF10 = 0.89) and support for public action (BF10 = 0.78).

Following a reviewer’s suggestion, we also explore whether country-level characteristics, such as individualism–collectivism and power distance, moderate the effects of both interventions on perceptions of the reality consensus and crisis agreement. We find no convincing evidence for any moderation by these two cultural dimensions (Supplementary Information Section 2.6). However, we detect only weak evidence against or for any country-level moderation effects, which suggests that this study is underpowered to robustly probe such moderations. These results should, therefore, be seen as tentative and followed up by analyses on datasets including more countries.

Discussion

Across 27 countries on six continents, we test the effectiveness of two climate science consensus messages, a classic message on the reality of climate change and an updated message that additionally emphasizes the agreement among scientists on climate change as a crisis. We find substantial misperceptions of the scientific consensus that human-caused climate change is happening across all country samples—indicating a gap between the actual and the perceived scientific consensus that could be reduced with consensus messaging interventions. However, perceptions of the crisis agreement are relatively accurate in all country samples.

Complementing previous studies that relied primarily on US samples26,32,33, informing people about the 97% scientific consensus on the reality of climate change is largely effective in a more diverse 27-country sample. Specifically, the classic scientific consensus intervention substantially increases perceptions of the scientific consensus (Cohen’s d = 0.47, 95% CI (0.41, 0.52)) as well as—to a smaller extent—beliefs in the reality (d = 0.06, 95% CI (0.01, 0.12)) and human causation (d = 0.10, 95% CI (0.04, 0.15)) of climate change, as well as worry (d = 0.05, 95% CI (−0.01, 0.10)). However, we find weak evidence (BF+0 = 0.62) against a direct effect on support for public action (d = 0.02, 95% CI (−0.03, 0.08)). The magnitude of the effects is generally in line with recent meta-analyses which found moderate effects on revising consensus perceptions and small effects on outcomes that are more remote from the intervention, such as climate change beliefs and worry26. As the conceptually most remote outcome from the intervention, it is perhaps unsurprising that this study finds no direct effect on support for public action, although some studies, conducted predominantly in the United States, report significant indirect13,25 and direct effects24. In sum, while consensus messaging on climate change can produce small shifts in personal beliefs and worry on the topic of climate change, a one-time messaging intervention alone seems insufficient to alter preferences toward major policy topics, which has also been noted in previous meta-analytic work26,27.

The updated consensus message shows similar effects as the classic message on beliefs in the reality (d = 0.07, 95% CI (0.02, 0.12) and human causation (d = 0.09, 95% CI (0.04, 0.15)) of climate change, as well as worry (d = 0.04, 95% CI (−0.01, 0.08)), with weak evidence (BF+0 = 0.65) against a direct effect on support for public action (d = 0.02, 95% CI (−0.02, 0.07)). We find only weak evidence for a positive effect on crisis belief (BF+0 = 1.80; d = 0.04, 95% CI (−0.01, 0.09)). Comparing the updated to the classic message showed no added value in additionally communicating the 88% scientific agreement on climate change as a crisis beyond strengthening confidence in perceptions of the crisis agreement. These findings might be due to participants’ perceptions of the scientific agreement being already relatively accurate before message exposure. Additionally, given that even small perceived dissent among environmental scientists can undermine message effectiveness15,59,60, the tested scientific agreement (88%) might not be high enough, and consequently convincing enough, to further shift belief in climate change as a crisis, worry and support for public action. Therefore, more attention needs to be devoted to effective ways of communicating the very high scientific confidence about the adverse consequences of climate change and the urgency of climate action to curb these impacts expressed in the Sixth IPCC report35.

The results of the current study also provide several useful indicators for selecting target audiences for consensus messaging interventions. Consistent with Bayesian models of information processing39,40,41, the message seems to be more effective for individuals who report being less familiar with it before exposure. Our exploratory analyses suggest that people with lower initial perceptions of the consensus increase their estimates to a greater extent, probably because they have more ‘room’ to update their perceptions. They also increase their belief in the reality of climate change more than people with higher initial perceptions. Consequently, repeated exposure to scientific consensus messages might have diminishing returns. As the effect on consensus perceptions is detectable days and weeks after exposure31,61,62, people might become more accurate and familiar with the message as they are repeatedly exposed to it, which, in turn, would yield increasingly smaller effects. While this decay in effectiveness is probable when people are not exposed to contrarian views, the information ecosystem contains climate misinformation and disinformation63,64, particularly in contexts where climate change is a politicized topic. In such contexts, consensus messaging can neutralize counterarguments23 and repeated consensus message exposure is effective for those who report being exposed to a mix of contradicting information between two exposures61.

Contrary to some motivated cognition accounts and findings43,44, consensus messaging does not seem to backfire for people whose worldviews might not align with the scientific consensus on climate change, such as right-leaning individuals or those with lower trust in climate scientists. Rather, the present study supports previous research that found consensus messages to result in larger belief updating for those with right-leaning political ideologies24,31 and extends this to those with lower trust in climate scientists, as these groups tend to have higher initial misperceptions. This means that a left-leaning person with a pre-intervention consensus estimate of 75% is likely to update their consensus perceptions more than a right-leaning person with the same consensus estimate. However, at the group level, targeting low-trust and right-leaning individuals corrects misperceptions to a greater extent. As people across the ideological and trust spectrums still update their estimates, consensus messaging represents a non-polarizing tool useful for reaching a social consensus on climate change across different audiences.

While the present study tested the practical use and general effectiveness of scientific consensus messages across countries, we recognize several limitations. First, as we focus on the direct effects of the interventions on several outcomes, this work does not speak to theoretical predictions of the Gateway Belief Model24,33—the main theoretical framework for scientific consensus messaging—that focuses on cascading indirect effects of consensus messaging through changes in perceived consensus and further through climate change beliefs to support for public action. We did not measure pre-intervention estimates for all beliefs, which precludes formal modelling of the ‘gateway’ process.

Second, this study finds nominally small effects65 of scientific consensus messaging on personal climate change beliefs and worry. However, these effects are in line with previous research26,27 and can be practically relevant66, as the intervention is easily scalable to reach many people because of its brevity. Targeting specific subgroups, such as those on the political right who are most likely to underestimate the consensus, might also increase its overall effectiveness.

Third, social-media users are generally younger, more educated, more liberal, more likely to be female and pay more attention to politics67,68, which is also reflected in our current samples. On the one hand, our social-media-based sampling approach may have led us to underestimate the intervention effects because, for example, younger and more educated individuals are more likely to believe in climate change69, which is, in turn, associated with higher perceptions of the scientific consensus70. On the other hand, several previous studies have shown that average treatment effects can be accurately estimated in experiments using convenience samples71,72. This is also supported by the fact that the effect size estimates for the effectiveness of the classic message in our study (misperception correction, d = 0.47; climate change attitudes, d = 0.05–0.10) relatively closely align with effect size estimates from previous meta-analyses on scientific consensus messaging (misperception correction, g = 0.56; climate change attitudes, g = 0.09–0.12) that rely on mostly US-based studies with nationally representative samples. From a practical perspective, hard-to-reach populations (for example, people who do not have access to the internet or do not use social media) will probably not be exposed to and thus influenced by a scientific consensus message when used by policy-makers in, for example, online campaigns. We do not discount the importance of those populations; we simply highlight this consideration in the context of the effectiveness of this specific intervention.

Fourth, we are unable to draw definitive conclusions about the extent of between-country heterogeneity and make concrete recommendations as to where scientific consensus messaging might be most effective, due to the convenience sampling approach and insufficient statistical power to detect moderations by country-level predictors (that is, cultural dimensions). We encourage future research to continue testing message effectiveness within countries using representative samples and, possibly, our materials and translations, to ultimately make practical recommendations for climate change communication tailored to specific countries (for example, as previously done in Germany73). In addition, datasets including many countries are essential for robustly testing country-level factors that might determine consensus messaging effectiveness.

Last, we recognize the limitations of single items, especially for broader constructs, such as support for public action24. Future research might investigate the effects of scientific consensus messages on specific climate change mitigation policies. As beliefs in climate change and human causation may not only be associated with support for mitigation but also adaptation policies74, we encourage further studies to investigate the effectiveness of scientific consensus messaging on climate change adaptation policies32. Effects on climate change mitigation are especially impactful in countries with high carbon emissions32, whereas climate change adaptation might prove more useful in nations with comparably lower carbon emissions that are, at the same time, disproportionately affected by climate change75.

Conclusion

Across more than 10,000 participants and 27 countries, this study shows that scientific consensus messages on climate change can reduce consensus misperceptions and produce small shifts in climate change beliefs and worry. This effect does not directly extend to support for public action. Communicating the scientific agreement that climate change is a crisis, along with the consensus that human-caused climate change is happening, seems to have no added value beyond strengthening confidence in perceptions of the crisis agreement. This underscores the importance of continuing to investigate effective ways to communicate climate science projections, beyond the consensus that human-caused climate change is happening. Crucially, scientific consensus messages are most effective among people who were less familiar with the message and had less accurate initial consensus perceptions, including those with lower trust in climate scientists and right-leaning political ideologies. In sum, scientific consensus messaging is an effective, non-polarizing tool for substantially reducing scientific consensus misperceptions and slightly shifting personal climate change beliefs and worry across samples and audiences.

Methods

Ethics

We obtained ethical approval from the Institutional Review Board from the University of Amsterdam (the Netherlands; protocol FMG-1123) and the University of Porto (Portugal; protocol 2023/06-12). All participants provided informed consent at the beginning of the survey experiment. Participation was voluntary and not compensated. However, in Canada and Mexico, we supplemented the unpaid samples with paid participants using Prolific. We collected paid samples in countries in which (1) we could not achieve the target sample size through convenience sampling and (2) enough participants were available on Prolific.

Participant recruitment

We recruited participants using an existing network of researchers that used scalable methods to collect large, diverse samples in 27 countries (Argentina, Australia, Austria, Brazil, Canada, China, Egypt, Georgia, Germany, India, Indonesia, Israel, Italy, Lebanon, the Maltese Islands, Mexico, the Netherlands, Poland, Portugal, Serbia, Singapore, Slovenia, Sweden, Tunisia, Türkiye, the United Kingdom and the United States) from 27 July to 4 August 2023. The final set of countries was selected according to our collaborators’ familiarity and connections with the countries, aiming for geographic spread with at least one country per continent, except Antarctica76.

We collected convenience sampling using snowballing, mailing lists, social media and Prolific (only in Canada and Mexico). On social media, we posted in special interest groups that relate to current events, popular culture or media discussions. We also posted comments on discussion threads of major news stories unrelated to climate change or sustainability. These approaches have been effective at recruiting a diverse body of participants in similar research where comparable sample sizes were required76,77. Crucially, because climate change is a prominent topic in public discourse, we advertised the study as a survey on popular media topics, to prevent selection bias of participants with strong opinions on climate change.

Sample details

To be able to participate in this study, participants needed to be at least 18 years old, live in one of the 27 target countries and speak the language in which the survey was conducted (that is, the most prominent locally spoken language/s) fluently. Of 21,462 individuals who clicked on the link, 11,702 participants completed the study, while 676 were filtered out at the beginning of the survey because they did not reside in any of the 27 countries. Out of all people who dropped out, most did so after seeing the informed consent (2,687; 29.6%), after the introduction that they are randomly assigned to one topic but before seeing that this topic is climate change (804; 8.8%) and right after the control/intervention message (931; 10.2%). Consensus and agreement perceptions of individuals in the two intervention conditions who dropped out directly after seeing the message (consensus, 78.4% and 79.1% in the classic and updated condition; agreement, 75.3% and 77.1%) are slightly lower compared to perceptions of those who completed the study (consensus, 84.2% and 84.5%; agreement, 83.4% and 83.3%), indicating selective dropout. However, the dropout rate (that is, number of dropouts directly after seeing the control/intervention message versus number of overall dropouts) is comparable between both intervention (classic: 267, 2.9%; updated: 296, 3.3%) and control conditions (368, 4.1%). This indicates that the dropout is unlikely to result from a specific backfire of the consensus messages and is suggestive of a more general tendency of less motivated participants to trickle out of the survey in its initial stages. After data exclusions (see section ‘Data analysis’), 10,527 participants (including Canada, npaid = 179 and Mexico, npaid = 143) remained for the analysis.

Design and procedure

We conducted an online study using Qualtrics. The study implemented a between-participants design with three conditions—consensus, updated consensus and control. As the only exception to this, we measured reality consensus and crisis agreement perceptions both before and after the intervention (including masking the study aims and distractors to minimize demand effects). This is because this ‘estimate and reveal’ technique makes the intervention more effective20, as it highlights the gap between the participants’ perceptions and the scientific norm33,61. Therefore, we measured pre-intervention reality consensus and crisis agreement perceptions and control for them in estimating between-participants effects of the two interventions on post-intervention consensus perceptions. The outline of the procedure is depicted in Supplementary Fig. 2. The median completion time was 6.33 min.

Pre-intervention

As part of the informed consent, the topic of the research was described as ‘opinions about and reactions to popular news topics’ to reduce potential self-selection biases as well as biasing participants’ responses. After providing informed consent, we asked for participants’ current country of residence. If they did not live in any of the targeted countries, they were redirected to the end of the survey. The remaining participants were told that they were now asked to provide their opinion on one out of 20 randomly selected news topics. In fact, all participants answered questions about climate change.

Next, to ensure equal understanding of the topic across testing contexts, all participants saw a brief description of climate change as a news topic (‘You may have noticed that climate change has been getting some attention in the news. Climate change refers to the idea that the world’s average temperature has been increasing over the past 150 years, may be increasing more in the future and that other aspects of the world’s climate may change as a result’) used in previous research78.

Participants then responded to two items assessing climate change consensus perceptions in a randomized order (see section ‘Materials’ for item wording).

Intervention

Following previous studies on climate change consensus messaging44, participants were informed that they would see a random statement from a large database of media statements the researchers maintain and were randomly, double-blind assigned to one of the three experimental conditions (consensus, updated consensus or control). Participants in the consensus condition saw the classic message on the scientific consensus emphasizing the reality of climate change: ‘97% of climate scientists agree that human-caused climate change is happening’. Participants in the updated consensus condition were shown the classic consensus message, supplemented with the IPCC scientists’ agreement that climate change is a crisis: ‘97% of climate scientists agree that human-caused climate change is happening. In addition, 88% of climate scientists agree that climate change constitutes a crisis’. We distinguish between reality consensus and crisis agreement to emphasize that scientific consensus on the reality of climate change was obtained by analysing abstracts of scientific publications, while the scientific agreement with regards to climate change as a crisis was obtained by surveying IPCC authors (that is, percentage of IPCC authors who agree with the statement that climate change is a crisis). See Supplementary Information Sections 3.3 and 3.4 for further information on the wording choice and the pilot study. To reduce anchoring effects, the control group was shown an unrelated consensus message: ‘97% of dentists recommend brushing your teeth twice per day’61.

Post-intervention

Similar to previous studies on consensus messaging24,61, as a distractor task, all participants read a paragraph about an upcoming science fiction film, Dune 2, and were asked one filler question about the movie. They also responded to an attention check.

After, they again reported their perceptions of the scientific consensus on the reality of climate change and agreement that climate change is a crisis (randomized and consistent with pre-intervention order for each participant), their confidence in these estimates directly after each one, as well as their personal beliefs in climate change beliefs, worry and support for public action (see following section ‘Materials’). After reporting demographic information, they answered several questions that tap into potential moderators (message familiarity, trust in scientists and political ideology). Finally, they completed a comprehension check and were debriefed.

Materials

Outcomes

Perceived scientific consensus on the reality of human-caused climate change

‘To the best of your knowledge, what percentage of climate scientists agree that human-caused climate change is happening?’. Participants responded on a slider scale from 0% to 100%. See ref. 24.

Perceived scientific agreement about climate change as a crisis

‘To the best of your knowledge, what percentage of climate scientists agree that climate change constitutes a crisis?’ Response options again ranged from 0% to 100%.

Confidence in both scientific consensus and agreement estimates (exploratory)

‘How certain are you about your answer above?’. Participants indicated the confidence in their estimates on a scale from 0 (very uncertain) to 100 (very certain). See ref. 20.

Belief in climate change

‘How strongly do you believe that climate change is or is not happening?’. Responses were provided on a scale from 1 (I strongly believe climate change is not happening) to 7 (I strongly believe climate change IS happening). See ref. 24.

Belief in the human causation of climate change

‘How much of climate change do you believe is caused by human activities, natural changes in the environment or a combination of both?’. Response options ranged from 1 (I believe climate change is caused mostly by natural changes in the environment) to 7 (I believe climate change is caused mostly by human activities). Participants also had the option to report that they believe climate change is not happening. Adapted from ref. 24.

Belief in climate change as a crisis

‘How strongly do you believe that climate change constitutes a crisis?’. Response options ranged from 1 (I strongly believe climate change is not at all a crisis) to 7 (I strongly believe climate change is a crisis).

Climate change worry

‘How worried are you about climate change?’ Response options ranged from 1 (I am not at all worried about climate change) to 7 (I am very worried about climate change). See ref. 24.

Support for public action on climate change

‘Do you think society should be doing more or less to reduce climate change?’. Response options ranged from 1 (much less), 4 (same amount) to 7 (much more). Adapted from ref. 24.

Checks

Filler item

‘How likely are you to watch Dune 2?’, including the response options 1 (very unlikely) to 5 (very likely).

Attention check

‘This is a test item. Please select ‘somewhat agree’’. The response scale included 1 (strongly disagree), 2 (somewhat disagree), 3 (neither agree nor disagree), 4 (somewhat agree) and 5 (strongly agree).

Comprehension check

‘As part of this survey, you may have viewed one of the statements below. Please select the statement, if any, you have seen’. Participants could select one of four options: the control, classic consensus or updated consensus message as well as none of the above.

Demographic information

We assessed current country of residence (‘Do you currently live in (country)?’), age (‘What is your age in years?’), gender (‘What is your gender?’, including male, female, other and prefer not to say as response options), current region of residence (‘In which region do you currently live?’; response categories were adapted to each country), urbanicity (‘Would you describe the area where you live as urban or rural?’, including urban, rural and don’t know), highest education level (‘What is the highest level of education you have received?’, including seven levels, from 1 less than high school to 7 doctoral degree), student status (‘Are you currently a college/university student?’, yes/no) and ethnicity (‘Please choose which best describes you’; response categories were adapted to the country-specific context). While age, gender, education and political ideology were also used as covariates in confirmatory analyses, the remaining demographic information was used to describe the samples.

Moderators

Familiarity with the consensus/agreement statements

We assessed participants’ familiarity with the consensus messages using two items: (1) ‘Before taking this survey, to what extent were you familiar with the following statement: ‘97% of climate scientists agree that human-caused climate change is happening’?’ and (2) ‘Before taking this survey, to what extent were you familiar with the following statement: ‘88% of climate scientists agree climate change constitutes a crisis’?’. Response options for both questions ranged from 1 (not at all familiar) to 7 (very familiar). Adapted from ref. 42.

Trust in climate scientists

‘In general, how much do you trust or distrust climate scientists as a source of information about climate change?’ Response options ranged from 1 (strongly distrust) to 7 (strongly trust). Adapted from ref. 61.

Political Ideology

We assessed ideology using one item: ‘In politics, people sometimes talk of ‘left’ and ‘right’. Where would you place yourself on this scale, where 0 means the left and 10 means the right?’ from 0 (left) to 10 (right).

Translation

We used a standard forward-back translation procedure for the instrument, following the Psychological Science Accelerator’s79 recommendations. Specifically, the instrument was first translated into the local language by at least one researcher who is fluent in English and the local language. Another independent researcher, also fluent in both languages, then back-translated this version to English, which was compared to the original English instrument. The translators resolved any disagreements through discussion. After the wording had been agreed upon, independent team members performed fidelity checks for each Qualtrics survey to ensure equal formatting and verify the survey flow. The final version was pretested in the target population (n ≈ 5) for clarity checks. All necessary cultural adjustments are reported in Supplementary Information Section 3.5.

Data analysis

Preprocessing

Exclusions

All countries were included in the analyses, regardless of the achieved sample size. At the participant level, we excluded respondents who failed the attention check (n = 1,032) or finished the survey in <2 min (n = 0; based on pilot data where none of the respondents completed a similar survey in <2 min). In addition, for paid samples, responses identified as potential bots (Qualtrics variable: Q_RecaptchaScore < 0.50; n = 3), duplicates (Qualtrics variable: Q_RelevantIDDuplicateScore ≥ 75; n = 9) or fraudulent (Qualtrics variable: Q_RelevantIDFraudScore ≥ 30, as per Qualtrics recommendations; n = 6) were excluded. Given that these are not equally functional and thus informative across countries, we kept respondents for whom these metrics were not recorded. All exclusion criteria were preregistered.

Missing data

The following responses were coded as missing values: ‘prefer not to answer/say’ (items: gender and ethnicity) and ‘I believe that climate change is not happening’ (item: human causation of climate change). Participants with missing data were excluded on an analysis-by-analysis basis. Before data collection, we expected few missing values due to the forced-response format of all items. This was confirmed, with 130 missing values on belief in the human causation of climate change and 128 missing values on gender.

Outliers

We did not define or remove any outliers, as all measures are bounded, which effectively prevents any outliers.

Main analysis

We analysed the data using the Bayesian model-averaging framework53,54,80 with mixed-effects models. This allowed us to evaluate the evidence in favour and against the preregistered hypotheses81,82 while accounting for uncertainty in the model structure (for example, constant versus heterogeneous intervention effects across countries). For hypotheses about the continuous outcomes (perceptions of the reality consensus and crisis agreement), we ran mixed-effects linear regression models estimated using the BayesFactor83 R package and 100,000 iterations. For hypotheses about the ordinal outcomes (personal belief in climate change, human causation and climate change as a crisis, climate change worry and support for public action), we used mixed-effects cumulative probit regression models in Stan84 and through the Rstan R package85 and computed the marginal likelihood by means of bridge sampling using the bridgesampling R package86. These models were run with two chains, of which each included 2,000 warm-up and 3,000 sampling iterations. All models converged with \(\hat{R}\) < 1.02.

The cumulative probit regression models allowed us to deal with the skewed responses and different ordinal scale response patterns (no, random, constant and dominant)56.

In all models, we accounted for nesting of participants (level 1) within countries (level 2). In all analyses, we model-averaged across models assuming the presence versus absence of random-slopes of intervention (that is, differences in the intervention effect across countries) and adjusted for demographic covariates (age, gender, university degree and political ideology), unless stated otherwise. We supplemented the interpretation of Bayes factors with the evidence labels of ref. 87 based on ref. 88.

Previous settings

In the linear mixed-effects models, we set a prior scale of r = 0.50 for the fixed-effects regression coefficients, the ‘medium’ prior scale closely corresponding to the previously reported meta-analytic effect of g = 0.55 (ref. 26) and r = 0.25 for the random-effects regression coefficients, assuming that the between-country variability is approximately half the effect size. The common intercept and residual variance use the default Jeffreys prior.

In cumulative probit mixed-effect models, we set the prior standard deviation on the latent scale shift to θ = 0.14 (which converts to the Cohen’s d of approximately 0.10 were the data analysed as continuous which corresponds to recent meta-analyses26,27) and the standard deviation of the normal distribution for the random effects to τ = 0.07, again assuming that the between-country heterogeneity is approximately half the effect size. The common thresholds use the default standard normal prior distribution56.

The Supplementary Information Section 3.6 contain a detailed overview of our analytical approach and the Bayes factor design analyses. Detailed model specifications are available on OSF (https://osf.io/z6quh/).

Deviations from the preregistration

We preregistered this study on OSF before data collection (https://osf.io/b6cmp; 19 July 2023). We deviated from the preregistered plan in the following ways. First, we placed the comprehension check after, instead of before (as preregistered), the demographics and moderator sections to avoid compromising moderator validity. More specifically, if we had followed our preregistered item order, the message familiarity measure would have been less valid because the participants would have already seen all messages in the comprehension check question. Participants were exposed to one of the three messages, indicated their familiarity with both consensus messages and then reported which of the messages they had seen. Since all participants have seen both messages as part of the familiarity measure, we did not use the comprehension check in any of our analyses.

Second, we did not measure subjective income as specified in the preregistration, to keep the survey brief and because we did not plan to use it in the main analyses (unlike, for example, education). Both decisions were made before preregistering and we missed revising the protocol before preregistering. In addition, there was a typo in the preregistration about exclusion criteria: for paid samples, we excluded participants with the Q_RelevantIDFraudScore ≥ 30, instead of <30 as stated erroneously in the preregistration. This is in line with official Qualtrics guidelines on using this fraud indicator (https://www.qualtrics.com/support/survey-platform/survey-module/surveychecker/fraud-detection).

Reporting summary

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