The hyper-systemizing hypothesis: how the tendency to systemize influences conspiracy beliefs and belief inflexibility in clinical and general populations

The current study seeks to address the limitations of prior research by investigating whether the Systemizing Quotient (SQ) provides a more precise measure of traits associated with conspiracy beliefs and a bias toward counterevidence, and whether this relationship is particularly relevant within autistic populations. Study 1 employed a Latent Profile Analysis (LPA) to compare profiles of individuals from a general sample, evaluating whether the SQ more effectively captures traits linked to conspiracy beliefs across different groups. Following, Study 2 assessed the SQ's predictive value within a clinical sample of individuals diagnosed with autism, examining whether it outperforms broader measures, such as the Autism Quotient (AQ), in its association with conspiracy beliefs. Additionally, Study 2 will explore whether systemizing tendencies (ST) explain the relationship between AQ and conspiracy theory (CT) beliefs. Together, this approach will test the validity of the 'hyper-systemizing hypothesis,' which posits that a preference for systemizing acts as a distinct pathway to conspiracy beliefs and susceptibility to biases against contrary evidence that would debunk conspiracy theories (e.g., the Bias Against Disconfirmatory Evidence task; Georgiou et al. 2023, 2025). Specifically, we hypothesise that individuals with systemising tendencies in both autistic and general populations, will show a higher level of conspiracy theory endorsement as conspiracy theories provide the sense of predictability and control that align with a more systematic belief system.

Study 1 involved 412 participants (Male = 220, Female = 192). Participants were recruited from a range of countries, with the majority residing within the United States of America (23%), the United Kingdom (18%), or Continental Europe (16%). Other demographic results including, level of education and prior mental health history can be found in Table 1.

Demographic information and other measures, (including screening for prior clinical diagnosis) and measures of both scientific reasoning performance and BADE task performance were completed online. The study was approved by the Research Ethics Committee of **anonymised**.

For the purposes of exploratory LPA, in accordance with the recommendations of previous research (i.e. Deleuze et al. 2015; Hussain et al. 2015; Denovan et al. 2018), psychometric measures using Likert scales were recoded. On measures of autistic traits (AQ scores), systemising tendencies (SQ scores) and conspiracy endorsement (GCBS scores), scores indicating disagreement or uncertainty (i.e., scores of 0–3) were recorded as “0”, and scores indicating agreement 4–5 were coded as “1”. Evidence integration scores were recoded to a range of − 10 to 10 as per previous research applying class analysis to the performance task. Scientific reasoning performance was not recoded.

The optimal number of latent classes was determined by considering a range of indices; the Akaike Information Criterion (AIC; Akaike 1987), the Bayesian Information Criterion (BIC; Schwarz 1978) the sample-size adjusted BIC (ssaBIC; Sclove 1987), the Lo-Mendell-Rubin-adjusted likelihood radio test (LMR-A-LRT; Lo et al. 2001), and a standardized measure of entropy (Ramaswamy et al. 1993). For AIC, BIC, and ssaBIC scores, smaller value indicates a better fit of the model. The LMR-A-LRT score does not rely on chi-square distribution for the difference in model likelihood values and normally occur alongside an associated p-value. Progressive class solutions are computed until an LMR-A-LRT value is found that is non-significant, which indicates the model cannot be improved for fit. Lastly, entropy ranges from 0 to 1, with higher values suggesting a better classification of participants. According to Ramaswamy et al. (1993), an entropy value of above 0.80 reflects a sound separation of identified classes in relation to the data. Once LPA exploratory analysis has been conducted, Latent Class membership acted as a group variable for assessing whether differences existed on scientific reasoning and BADE scores.

Prior to conducting the latent profile analysis (LPA), potential covariates of conspiracy theory (CT) beliefs, including level of education, age, gender, and prior clinical diagnosis, were examined to assess their influence on the parameters of interest (e.g., AQ, SQ, SRS, GCBS scores). There was no gender difference in GCBS, t(410) = 0.77, p = 0.44, d = 0.08. GCBS also did not differ across education levels, F(3, 408) = 1.03, p = 0.38, η² = 0.008, or across age groups, F(4, 407) = 0.94, p = 0.44, η² = 0.009. Consequently, gender, age, and education were not retained as covariates in the LPA.

Initial model comparisons began with 1-class and 2-class solutions. Model fit indices, including AIC, BIC, and ssaBIC, indicated that the 2-class model provided a superior fit. Additionally, the Lo-Mendell-Rubin adjusted likelihood ratio test (LMR-A-LRT) for the 2-class model demonstrated a significant improvement over the 1-class model (see Table 3). Subsequent comparison between 2-class and 3-class solutions favoured the 3-class model, as evidenced by lower AIC, BIC, and ssaBIC values, higher entropy scores (0.77 compared to 0.71), and a significant LMR-A-LRT p-value. Further examination of a 4-class solution suggested it provided an even better fit, with reduced AIC, BIC, and ssaBIC values, increased entropy scores (0.82 compared to 0.77), and a significant LMR-A-LRT p-value. However, exploration of a 5-class solution did not yield significant improvement over the 4-class model. When these tests fail to reach significance, the more parsimonious model is typically preferred to avoid overextraction and ensure replicability (Berlin et al. 2014; Masyn 2013; Nylund et al. 2007). Consequently, the 4-class solution was selected as the final model, as it represented the optimal balance of fit indices and interpretability.

As shown in Table 4, a 4-class solution represented the model most likely to be of best fit. In this model, Class 1 (37.6% of the sample) demonstrated low scores on autistic traits, systemizing tendencies, and conspiracy beliefs, along with the highest levels of scientific reasoning skills and best performance on the BADE task (i.e., lowest evidence integration scores). Class 1 can be considered the group with the strongest reasoning abilities and the least likely to endorse conspiracy theories. Class 2 (29.6% of the sample) also had low scores on autistic traits and systemizing tendencies but showed slightly lower scientific reasoning and BADE performance. Although Class 2 exhibited less adept reasoning skills, they did not show ‘clinical traits’ (i.e., did not score above the cut-off range on the AQ or SQ for likely clinical presentation) or a heightened tendency to endorse conspiracy beliefs. Class 3, representing 14.3% of the sample, was most distinct in exhibiting the highest systemizing tendencies, intermediate levels of autistic traits and conspiracy beliefs, and lower evidence integration ability on the BADE task. However, their scientific reasoning performance was similar to that of Class 2. Class 4 (18.4% of the sample) exhibited the highest levels of autistic traits and conspiracy beliefs, coupled with the lowest scientific reasoning scores and poorest BADE task performance (i.e., the lowest ability to integrate evidence). Both Class 3 and Class 4 may represent ‘clinical’ levels of autistic traits, but key differences in systemizing tendencies, reasoning skills, and BADE performance may distinguish the two groups.

As shown in Table 5, post-hoc pairwise comparisons with Bonferroni correction were used to assess AQ, SQ, SRS, conspiracy theory beliefs, and BADE performance as a function of class membership. In terms of conspiracy theory endorsement (via GCBS scores), Bonferroni correction revealed that Class 3 and Class 4, which likely represent individuals with clinical levels of autistic traits, had significantly higher conspiracy belief scores compared to Class 1 and Class 2. The difference in scores between Class 4 and Class 1 (i.e., those with the highest level of scientific reasoning) highlights the expected pattern from the existing literature: the more adept individuals are at evaluating conspiracy content, the less likely they are to endorse CTs. However, the results here suggest that Class 3 has demonstrated an exception to this argument with both higher SRS and higher CT endorsement.

Prior to commencing analysis, one AQ score was found to be an outlier believed to potentially influence on exaggerating effect sizes (i.e., 2.5 below the Standard deviation). Despite some slight deviations from normality on SRS scores, and the obvious slight negative skew of AQ and SQ scores due to being solely a clinical sample, the data distributions were found to be suitable for parametric testing. The majority of participants scored above the midrange for both the AQ and SQ, which is expected due to the sample consisting solely of individuals diagnosed with Autism Spectrum Disorder.

The majority of participants also scored above conventional screening thresholds for both the AQ-10 (cut-off ≥ 6) and SQ-10 (cut-off ≥ 5; Baron-Cohen et al. 2008, 2017), consistent with their verified ASD diagnoses. Fourteen percent of participants (n = 20) scored below the AQ-10 threshold, reflecting known discrepancies between questionnaire cut-offs and formal clinical assessments of autism (Ashwood et al. 2016). Because inclusion was based on a confirmed clinical diagnosis rather than screening criteria, all participants were retained. Sensitivity analysis excluding below-threshold participants produced an equivalent pattern of results (R² = 0.24, ΔR² = 0.01), and no change in the direction or significance of any covariate effects in later analysis. Influence diagnostics indicated that the single AQ outlier did not materially affect the regression results, and the case was therefore kept in all analyses.

A one-way ANOVA revealed no significant differences in conspiracy theory beliefs or systemizing tendencies based on age, gender, or level of education. Table 6 summarizes the descriptive statistics for the different psychometric variables of Study 2. The majority of participants scored above the midrange for both the AQ and SQ, which is expected due to the sample consisting solely of individuals diagnosed with Autism Spectrum Disorder. The sample also scored above the midpoint on the measure of conspiracy theory beliefs (i.e., GCBS scores) (Table 7).

Pearson correlations are presented in Table 8. Consistent with previous research (e.g., Georgiou et al. 2025a, b), there were weak positive correlations between general autistic traits (AQ scores), conspiracy theory endorsement (GCBS scores), and a bias against disconfirmatory evidence (EI scores). Consistent with our hypotheses, systemizing tendencies (SQ scores) showed a positive correlation with general autistic traits (AQ scores) and conspiracy theory endorsement. However, SQ scores also showed a weak positive correlation with both bias against disconfirmatory evidence (BADE) and scientific reasoning performance (SRS scores), while BADE performance showed a negative correlation with scientific reasoning performance. These results suggest a counter-intuitive relationship, wherein higher systemising tendencies were associated with contrasting outcomes: a greater bias against counterevidence, as indicated by BADE performance, but a greater ability to perform scientific reasoning skills, considered to be a robust protective factor against conspiracy beliefs. All correlations remained significant following False Discovery Rate (Benjamini–Hochberg) correction, indicating that the observed relationships were robust to Type I error control. These results also align with the cognitive profiles found in Class 3 of Study 1 (see Table 4).

Based on the results of Study 1 and preliminary correlations from Study 2, a moderated regression analysis was conducted to examine whether systemizing tendencies (SQ) influence the relationship between general autistic traits (AQ) and conspiracy beliefs (GCBS). BADE and scientific reasoning performance (SRS) were included in the model as covariates. All predictors were mean-centred prior to computing the interaction term to reduce multicollinearity between main effects and interaction terms (Aiken and West 1991).

In Step 1, AQ, SQ, BADE, and SRS were entered as predictors of GCBS. In Step 2, the interaction term (AQ × SQ) was added to test for moderation. The Step 1 model accounted for 19% of the variance in conspiracy theory endorsement (R² = 0.19, Adjusted R² = 0.17, F(4, 140) = 8.21, p < 0.001). The addition of the interaction term in Step 2 significantly improved model fit (ΔR² = 0.04, ΔF(1, 139) = 15.52, p < 0.001, f² = 0.052), bringing the total explained variance to 23% (R² = 0.23, Adjusted R² = 0.21). A significant interaction term between systemizing tendencies (SQ) and autistic traits (AQ) on conspiracy belief endorsement was identified (β = 0.21, p < 0.001), indicating that SQ moderates the relationship between AQ and GCBS. Significant main effects were observed for BADE (β = 0.14, p < 0.001) and SRS (β = − 0.14, p < 0.001). Specifically, higher BADE scores were associated with greater endorsement of conspiracy theories, whereas higher SRS scores were associated with lower endorsement. Full regression coefficients for this model are presented in Table 9.

To further explore this moderation effect, SQ scores were standardized and categorized into low (−1 SD), average (mean), and high (+1 SD) groups. Conditional-effects analysis revealed that at low levels of SQ, the relationship between AQ and GCBS was weak and non-significant (β = 0.09, p = 0.23), whereas at high levels of SQ, the relationship was moderate and significant (β = 0.24, p < 0.01). These conditional effects are summarized in Table 10. This finding further emphasises the importance of an individual's cognitive profile, rather than their diagnosis alone (i.e., Roels et al. 2024), for how people may believe in conspiracy theories, as discussed by Georgiou et al. (2024).

The finding that systemizing tendencies are associated with both scientific reasoning (a known barrier to conspiracy beliefs) and conspiracy belief endorsement underscores the need for a more integrated research framework that extends beyond existing cognitive models. While some integrated models have been proposed to examine the antecedents of conspiracy beliefs—either through comparisons of predictors (e.g., Georgiou et al. 2019) or cross-examinations via reviews (e.g., Bowes et al. 2021; Biddestone et al. 2022; Goreis and Voracek 2019)—the statistical methods used to analyse conspiracy beliefs still largely assume that reasoning ability is the primary determinant (Pilch et al. 2023).

For instance, whether through Bayesian modelling, signal detection analyses, or frequentist methods such as those in this study (e.g., latent profiles), most approaches implicitly frame belief formation as a matter of distinguishing truth from falsehood or assessing how reasoning and thinking styles affect belief accuracy. However, they often overlook whether conspiracy believers prioritize accuracy in the first place. Douglas and Sutton (2023, 2024) have also raised concerns about conspiracy belief outcome measures, noting that ‘endorsement’ is typically framed in terms of accuracy rather than accounting for broader motivational influences. This study highlights this issue by demonstrating that systemizing tendencies predict conspiracy beliefs independently of reasoning ability, reinforcing the need to consider alternative frameworks. While current measurement approaches fail to account for how motivational and identity-based factors interact with cognitive tendencies, motivated reasoning provides a potential lens for future research. This conceptual area remains largely unexplored, but the hyper-systemizing hypothesis could help examine how these factors collectively shape belief formation. However, a clearer framework is needed to effectively integrate these perspectives.

The findings of Georgiou et al. (2024), alongside the present study on systemizing tendencies, suggest that engagement with conspiracy content may persist regardless of reasoning ability when an individual has a strong preference for systemizing within an autistic population. This has implications for research that has primarily examined conspiracy beliefs in autistic populations through diagnostic group comparisons without incorporating broader cognitive profiles. Roels et al. (2024) examined group-level differences in conspiracy beliefs but did not account for systemizing tendencies or other cognitive factors that may interact with belief formation. The present findings indicate that within-group variability, particularly in cognitive style, may be key to understanding why some autistic individuals are more prone to conspiracy beliefs. Rather than focusing solely on autism as a diagnostic category, future research should consider the associated traits—such as systemizing tendencies and specific cognitive profiles—that may drive engagement with conspiracy content.

More broadly, these results highlight the risks autistic individuals may face in digital media spaces and raise questions about how to address conspiracy theory engagement in potentially vulnerable groups. If engagement is driven by a preference for systemizing rather than an ability to discern truth based on reasoning, the common strategy of providing an "inoculation" against conspiracy content may be ineffective. Many existing interventions focus on countering the validity of conspiracy claims but fail to address their motivational purpose or alternative functions. Indeed, Williams et al. (2024a) has emphasized that people rarely change their conspiracy beliefs over time, suggesting that current intervention methods insufficiently account for factors beyond logic and reasoning. This may explain the weak effect sizes often found in studies testing interventions for conspiracy beliefs in autistic samples (e.g., Georgiou et al. 2023, 2025).

Therefore, future research should explore alternative intervention approaches tailored to individuals with high systemizing tendencies. In particular, qualitative research is needed to better understand the motivations underlying engagement with conspiracy content in autistic individuals. This may provide insights into how interventions can address not only accuracy but also the broader appeal and function of such beliefs.