Changing minds about climate change: a pervasive role for domain-general metacognition

The effects of climate change on planetary life are among the primary existential threats we are facing. Scientific consensus on the existence, causes and impacts of climate change is well established, yet a considerable percentage of people remain only moderately concerned, and some deny anthropogenic climate change. Climate information has become ideologically polarized, with political affiliation, values, and ideologies influencing beliefs as much as factual knowledge. People tend to update beliefs more in response to desirable or self-confirming information and less to undesirable or disconfirming information. Prior work shows that climate change sceptics and believers update asymmetrically to good versus bad news. Recent research highlights cognitive underpinnings of polarized beliefs, including metacognitive sensitivity and cognitive inflexibility. Impaired domain-general metacognitive sensitivity and reduced adjustment of confidence to corrective evidence predict radical beliefs. Domain-specific meta-knowledge predicts climate beliefs more effectively than factual knowledge. However, the role of metacognition in changing climate-related beliefs remains underexplored. The present work investigates whether domain-general metacognitive abilities underlie attitudes towards climate change and influence the updating of climate-related beliefs, also considering the mode of science communication (textual versus visuo-textual).

Design and recruitment: Two online studies were hosted on Gorilla Experiment Builder and administered via Amazon Mechanical Turk. Both were approved by the Institutional Ethics Committee (Royal Holloway, University of London). Participants provided demographics and informed consent. Participants: Study 1 recruited 424 US participants (Dec 2019–Jan 2020); after exclusions for performance/attention checks, n=364 (36.54% women; mean age 36.67, SD 11.17). Study 2 recruited 435 US participants (Mar 2020); after exclusions, n=335 (42.39% women; mean age 35.86, SD 10.11). Only US citizens participated; Study 1 participants could not take Study 2. Procedure overview: Participants completed (1) a perceptual metacognition task to assess domain-general metacognitive ability, (2) a climate belief-updating task with good/bad news scenarios, and (3) questionnaires including climate knowledge (with confidence ratings) to assess domain-specific metacognition, climate scepticism, dogmatism, social dominance orientation, and political orientation. Study 2 was identical except the belief-updating task included visual material (maps) supporting textual information. Perceptual metacognition task: Adapted from Rollwage et al. (2018) using JsPsych. On each trial, two 250×250 px black squares comprised 625-cell grids with flickering white dots; one square had more dots. Difficulty was controlled via the dot difference using a 2-down-1-up staircase to target ~71% accuracy. Calibration included 33 trials with feedback to set individual stimulus strength; difficulty was continuously adapted thereafter. The main phase had 80 trials with a left/right choice followed by a 9-point confidence rating (0–100% in 12.5% steps; midpoint 50%). A quadratic scoring rule incentivized accurate confidence. Domain-general metacognitive sensitivity (meta-d'), confidence bias (mean confidence), and perceptual performance (d' or accuracy) were estimated via hierarchical Bayesian methods (HMeta-d). Climate belief-updating task: Based on Sunstein et al. (2017), two scenarios were presented: (a) temperature rise and (b) sea-level rise by 2100. Initial statements conveyed commonly cited projections (e.g., temperature ≥6°F; sea level ~3 ft). Participants provided initial numeric estimates (temperature 1–12°F; sea level 0.5–6 ft) and confidence (1–9). They were then shown either good news (lower projected change; temperature 1–5°F; sea level 0.5–2.5 ft) or bad news (higher projected change; temperature 7–11°F; sea level 3.5–6 ft), after which they provided revised estimates and confidence. Scenario order and news valence were counterbalanced. Mode of communication (Study 2): Visual maps accompanied text. Temperature maps (Impact Lab) displayed current (2020) versus projected 2100 temperature distributions under different RCPs; sea-level maps (UCS/NOAA-based) showed population at risk of chronic inundation. Three-map displays were used for good/bad news (current, initial projection, and better/worse projection). A binary variable coded mode: 0=text-only (Study 1), 1=visuo-textual (Study 2). Climate change knowledge and domain-specific metacognition: A 16-item True/False knowledge test (8 true, 8 false) covered precipitation, air, temperature, health, species extinction, desertification, food, wildfires, and disease. Participants rated confidence (1–5) for each item. Accuracy scores were computed (number-right method and 1-0-0). Domain-specific metacognitive sensitivity (meta-d'), confidence bias (mean confidence), and accuracy (d') were estimated (HMeta-d for meta-d'). Questionnaires: Climate Change Scepticism Scale (20 items; Epistemic and Response scepticism subscales), Dogmatism Scale (20 items), Social Dominance Orientation SDO7s (8 items; Dominance and Anti-Egalitarianism), and a 3-item continuous political orientation scale (0 liberal–100 conservative). Analyses: Correlations used Spearman's rho. Linear mixed-effects models (lme4) predicted belief updating from news valence and cognitive/attitudinal predictors with participant random intercepts; variables were mean-centered and scaled; multicollinearity was checked; confidence intervals via bootstrapping (ggeffects). Structural equation modeling (lavaan) tested pathways among domain-general and domain-specific metacognition, confidence biases, climate change scepticism, mode of communication, and belief updating (good/bad), with model fit assessed by χ², RMSEA, CFI, and SRMR.

Belief updating and climate scepticism: - Study 1: Greater climate change scepticism predicted less updating after bad news (β = −0.27, CI [−0.35, −0.20], p < 0.0001) and interacted with valence (β = 0.14, CI [0.04, 0.26], p = 0.008): sceptics updated more with good news and less with bad news, replicating Sunstein et al. (2017). Higher dogmatism also predicted reduced updating to bad news (β ≈ −0.13, p = 0.009). - Study 2: Replicated the scepticism main effect (β = −0.28, p < 0.001) and valence interaction (p = 0.003). A main effect of valence indicated participants were less likely to update prior beliefs when positive evidence was communicated visuo-textually (p = 0.03). Domain-general metacognition and updating: - Study 1: Higher domain-general metacognitive sensitivity (meta-d') predicted greater updating to bad news, controlling for perceptual performance (β = 0.144, CI [0.06, 0.22], p < 0.001). Greater confidence bias (higher mean confidence) predicted reduced updating to bad news (β ≈ −0.091, p = 0.03). More conservative political orientation and higher dogmatism were associated with less updating to bad news (ps ≤ 0.03). - Study 2: Replicated the effect of metacognitive sensitivity on updating to bad news (β = 0.14, CI [0.059, 0.235], p = 0.001). A sensitivity × valence interaction showed higher sensitivity increased updating to bad news but decreased updating to good news (β = −0.13, CI [−0.26, −0.02], p = 0.02). Higher perceptual accuracy additionally predicted more updating to bad news under visuo-textual communication. Conservatism/dogmatism again predicted reduced updating (β ≈ −0.12, p = 0.03). Domain-specific metacognition and updating: - Study 1: Domain-specific metacognitive sensitivity did not significantly predict updating. However, higher domain-specific confidence bias predicted reduced updating to bad news (β ≈ −0.10, CI [−0.20, 0.01], p = 0.032). Accuracy on climate knowledge interacted with valence (β = 0.14, CI [0.04, 0.25], p = 0.009): more accurate participants updated more with good news. - Study 2: Overconfidence effect replicated (β ≈ −0.11, CI [−0.21, −0.01], p = 0.02). Domain-specific metacognitive sensitivity predicted greater updating to bad news (β ≈ 0.104, CI [0.01, 0.19], p ≈ 0.037). Path analyses (SEM): - Study 1 model fit was good (CFI = 0.972; RMSEA = 0.039; SRMR = 0.031). Both domain-general and domain-specific metacognitive sensitivities negatively predicted climate change scepticism (β = −0.14, p = 0.001; β = −0.39, p < 0.001). Perceptual and epistemic confidence biases positively predicted scepticism (β = 0.34, p = 0.001; β = 0.15, p = 0.001). Scepticism predicted decreased updating to bad and good news (bad: β = −0.27; good: β = −0.13; ps = 0.000). Indirect effects: higher domain-general meta-d' increased updating via reduced scepticism (good: β = 0.04, p = 0.002; bad: β = 0.01, p = 0.02); domain-specific meta-d' similarly (good: β = 0.10, p < 0.001; bad: β = 0.05, p = 0.003). Confidence biases indirectly reduced updating (perceptual: good β = −0.096, p < 0.001; bad β = −0.045, p = 0.003; epistemic: good β = −0.044, p = 0.003; bad β = −0.021, p = 0.0021). Perceptual performance did not predict scepticism or updating. - Study 2 model fit was good (CFI = 0.989; RMSEA = 0.022; SRMR = 0.029). Key pathways replicated; the indirect effect of domain-general meta-d' via scepticism on updating to bad news was a trend (β ≈ 0.013, p = 0.062). Mode of communication: Combining both studies in SEM showed no significant main or interaction effects of communication mode (text-only vs visuo-textual) on updating; both modes were similarly effective. Effect sizes: Measures explained 39% (Study 1) and 34% (Study 2) of the variance in climate change scepticism, underscoring the importance of (meta)cognitive characteristics.

The findings address the central question of whether metacognitive abilities, beyond domain-specific knowledge, influence changes in climate-related beliefs. Across two independent samples, both lower domain-general metacognitive sensitivity and higher confidence bias were associated with greater climate change scepticism and reduced belief updating, particularly for negatively framed evidence. This pattern aligns with theories of confirmation bias and the role of confidence in filtering new information. Domain-specific metacognitive processes also mattered: overconfidence in climate knowledge diminished openness to bad news, while better calibrated domain-specific metacognition fostered updating. Structural models showed that metacognition affected updating indirectly via its effect on scepticism, indicating that metacognitive capacities shape broader attitudes that, in turn, modulate evidence incorporation. The asymmetry in updating—stronger for self-confirming information—was replicated, and domain-general metacognitive sensitivity specifically increased responsiveness to disconfirming (bad) evidence. Together, these results suggest that failures of metacognition contribute to the persistence of erroneous or polarized beliefs about climate change and that improving metacognitive insight could enhance receptivity to corrective evidence.

This work extends prior research by demonstrating a pervasive role for domain-general metacognitive abilities in shaping both climate change scepticism and belief updating. While climate scepticism biases updating toward self-affirming information, metacognitive sensitivity—both domain-general and domain-specific—predicts lower scepticism and greater willingness to revise beliefs, especially in response to negatively framed evidence. In contrast, confidence biases reduce updating. The mode of communication (text alone versus visuo-textual) did not significantly alter updating efficacy. These results highlight metacognition as a core capacity underpinning both scepticism and belief revision in the climate domain. Given evidence that metacognitive sensitivity can be trained, future research should test whether metacognitive training can reduce dogmatic, inflexible beliefs and improve information-seeking and evidence integration in educational, professional, and policy contexts.

Samples were relatively large but not nationally representative, and findings are limited to US participants; cross-cultural generalizability remains to be tested. Although Study 2 added visual materials, the mode of communication showed no significant effect on updating; more nuanced or emotive visual designs might yield different outcomes and warrant further investigation. Extracted confidence intervals and some reported statistics in text reflect study reports; perceptual performance showed no predictive value in these datasets, but alternative tasks or measures might capture additional variance. The online, self-report nature of measures and the artificiality of laboratory-style updating tasks may limit ecological validity.