2.1 Literature Search and Selection

In January 2025, a comprehensive online database search was conducted utilizing PubMed and Web of Science. The search employed a series of keywords, including (“risk taking” OR “risk preferences” OR “risk seeking” OR “risk avoidance” OR “loss aversion” OR “risk aversion” OR “ambiguity aversion”) AND (“fMRI” OR “functional magnetic resonance imaging”). Furthermore, we delved into additional resources, specifically (i) the bibliographies of published meta-analysis reviews and (ii) targeted searches based on the names of prominent authors within the field.

The initial assessment of the retrieved literature involved screening titles and abstracts, followed by a more rigorous full-text evaluation using the criteria outlined in Figure 1. These criteria encompassed: (i) the utilization of fMRI as the imaging technique; (ii) a focus on whole-brain analysis, rather than region of interest (ROI) analysis; (iii) the inclusion of participants without psychiatric or neurological diagnoses, excluding minors and the elderly (i.e., healthy adult participants); (iv) the presentation of activations in a standardized stereotaxic space, such as Talairach or Montreal Neurological Institute (MNI) coordinates, with a conversion to MNI coordinates conducted for studies reporting Talairach coordinates using the icbm2tal algorithm (accessible at http://brainmap.org/icbm2tal/); and (v) the reporting of a binary contrast (e.g., high-risk choice compared to low-risk choice or baseline, or vice versa) or parametric modulation analysis for high- and low-risk choices. Adhering to these rigorous criteria, our final meta-analysis encompassed 43 experiments (with 422 foci) related to risk seeking (Table 1) and 22 experiments (with 305 foci) related to risk aversion (Table 2).

During activation coordinate extraction, we classified reported activations as follows: Risk seeking: Activations associated with (a) higher-risk choices (e.g., selecting riskier options/balloon pumping; 41 studies), (b) individuals with higher risk preference (1 study: Kruschwitz et al. 2012), or (c) higher-risk trials (two studies: van Leijenhorst et al. 2006; Arioli et al. 2023). Risk aversion: Activations associated with (a) lower-risk choices (e.g., selecting safer/certain options/cashing out; 21 studies), or (b) lower-risk trials (1 study: van Leijenhorst et al. 2006).

2.2 ALE Meta-Analysis

ALE is a quantitative meta-analytic approach that allows for the integration of fMRI results across studies (Eickhoff et al. 2009, 2012; Zhang et al. 2022). This study employed the ALE algorithm to explore the neural correlates of risk seeking and risk aversion in risk decision-making. BrainMap GingerALE version 3.0.2 software (Eickhoff et al. 2012; Turkeltaub et al. 2012) was used for the data analysis. The algorithm is grounded in probability distributions, which evaluate both the spatial probability distributions and the activation statistics of each voxel within the brain (Laird et al. 2005; Turkeltaub et al. 2002). First, all reported foci for each contrast in the different studies were modeled as Gaussian 3D distributions, centered on the reported coordinates. These modeled activation maps (MA maps) were then constructed by combining the maximum probability distributions of voxel-associated foci for each experiment. Next, the MA maps were combined to calculate the overall ALE maps for all experiments (Turkeltaub et al. 2012). Finally, ALE scores from the convergent MA maps were calculated voxel-by-voxel to test for convergent (random-effects) foci, rather than study-specific foci (fixed-effects) (Eickhoff et al. 2009, 2012).

This study conducted two separate ALE analyses for the risk seeking and risk aversion datasets. A cluster-level family-wise error (cFWE) correction at p < 0.05 was applied, with a cluster-defining threshold of p < 0.001 and 10 000 permutations to assess the significant p map (Eickhoff et al. 2017).

2.3 Conjunction and Contrast Analysis

Conjunction and contrast analyses were performed to test the common and distinct neural activations between risk seeking and risk aversion. In the contrast analysis, initial individual ALE analyses were conducted on the datasets pertaining to risk seeking and risk aversion, with subsequent calculation of the differences in their respective ALE values. Subsequently, all experiments were pooled and randomly partitioned into two groups, each matching the size of the original experiment set. The ALE values for these newly formed datasets were then recalculated, and the discrepancy between their ALE values was determined. This entire process was iterated 25 000 times to establish the null distribution of the ALE value differences between risk seeking and risk aversion (Eickhoff et al. 2011; Langner et al. 2018). The ALE maps were then thresholded at a posterior probability threshold of p > 0.95 (Wu et al. 2021; Wang et al. 2022). In the conjunction analyses, the overlapping regions between the two corrected ALE results were identified (Eickhoff et al. 2011).

2.4 Meta-Analytic Connectivity Modeling (MACM) Analyses

MACM analyses were conducted to examine the co-activation patterns associated with risk seeking and risk aversion, providing a more comprehensive analysis of their underlying neural mechanisms. MACM analyses estimate the functional connectivity of regions of interest (ROIs) (Robinson et al. 2010). We selected the clusters that emerged significant from the contrast analyses and utilized these cluster-based ROIs as separate seeds within the BrainMap database (http://www.brainmap.org/) to search for studies reporting activations within each respective cluster (Wu et al. 2021), facilitating the construction of data-driven functional connectivity maps of these ROIs (Laird et al. 2005). ROI was defined as a sphere centered on the peak activation coordinates within a 10 mm radius (Wang et al. 2022; Bellucci et al. 2018). For inclusion in our analysis, we only considered whole-brain neuroimaging studies that reported activations in standard stereotaxic space for healthy participants. Using BrainMap, we screened for experiments featuring at least one activation focus within a specified seed region. Subsequently, we downloaded the activation coordinates and performed ALE analysis, applying FWE correction at p < 0.05 and a cluster-defining threshold of p < 0.001, to evaluate the convergence of co-activations across seeds.

3.1 Results of Single ALE Analyses

ALE meta-analysis of risk seeking revealed significant activations in the bilateral caudate, bilateral paracingulate gyrus, and bilateral insula (Figure 2, Table 3). The insula integrates risk-related information, encodes risk predictions and prediction errors, while the caudate, part of the basal ganglia, plays a key role in reward processing. Additionally, the paracingulate gyrus has important cognitive and emotional functions, mediating bottom-up information processing from limbic inputs to executive control areas. Regarding risk aversion, significant activations were located in the left STG, right middle frontal gyrus (MFG), and left ACC (Figure 2, Table 3). These regions are crucial parts of the cognitive control network. The results suggest that risk seeking mainly engages risk and reward information processing systems, reflecting the predominance of the emotional system in risk seeking. Risk aversion is primarily associated with cognitive control processes, reflecting the predominance of the cognitive system in risk aversion.

3.2 Results of Conjunction and Contrast Analyses

The contrast analysis revealed that compared to risk aversion, significant activations converged in the right insula and left caudate during risk seeking (Figure 3, Table 3). Conversely, the left MTG and left ACC were more activated in risk aversion than in risk seeking (Figure 3, Table 3). However, no significant activations were found in the conjunction analysis. The results of the contrast analysis were consistent with those of the individual analyses. Moreover, our analysis did not identify common activations between risk seeking and risk aversion, indicating that these two behaviors exhibit distinct neural activity patterns.

3.3 Results of MACM Analyses

We selected the distinct activations identified in the contrast analysis for risk seeking and risk aversion as seeds to investigate the brain connectivity networks. Specifically, for risk seeking, the right insula (427 experiments, 6351 subjects, 6838 foci) and left caudate (165 experiments, 2476 subjects, 2293 foci) were used as seeds. For risk aversion, the left MTG (251 experiments, 3613 subjects, 2858 foci) and left ACC (242 experiments, 3591 subjects, 3057 foci) were used as seeds.

For risk seeking, the analysis using the right insula as the seed showed significant activations in the bilateral right insula, bilateral caudate, bilateral thalamus, left superior frontal gyrus (SFG), and bilateral IPL. When using the left caudate as the seed, significant activations were observed in the right insula, vmPFC, dorsal anterior cingulate cortex (dACC), left caudate, bilateral thalamus, and bilateral parahippocampal gyrus (see Table S1 and Figure 4). Among these regions, the caudate and thalamus are important parts of the reward network. The insula is a core region of the salience network. The dACC, SFG, and IPL are involved in the cognitive control network. These findings suggest that ROIs associated with risk seeking are predominantly linked to the reward network, salience network, and cognitive control network.

For risk aversion, the analysis using the left ACC as the seed revealed that significant activations were located in the LPFC, vmPFC, OFC, ACC, posterior cingulate cortex (PCC), parahippocampal gyrus, and left MTG. When using the left MTG as the seed, significant activations were observed in the vmPFC, LPFC, ACC, PCC, left MTG, and left parahippocampal gyrus (see Table S1 and Figure 4). Among these regions, the ACC, PCC, LPFC, and left MTG engage the cognitive control network. The vmPFC and OFC are components of the valuation network. These findings indicate that risk aversion ROIs are primarily linked to the cognitive control network and valuation network.

The results of MACM analyses show that both risk seeking and risk aversion involve a combination of brain networks. The cognitive control network is prominently involved in both risk seeking and risk aversion, while risk seeking is more strongly related to the reward network compared to risk aversion.

4.1 Distinct Neural Activity of Risk Seeking

The current study revealed that risk seeking involves neural activations in the right insula and left caudate. The insula plays a pivotal role in processing risk information, integrating risk-related signals and encoding predictions and prediction errors (Baek et al. 2017; Sun et al. 2017; Wu et al. 2021; Picard et al. 2021; Cui et al. 2022). Prior studies have confirmed the insula as a core neural substrate for uncertain information processing, spanning networks involved in risky, ambiguous, and anticipatory threat decision-making (Feng et al. 2022; Wang et al. 2024). During risk seeking, individuals remain uncertain until the outcome is realized, prompting the insula to evaluate risk and signal uncertainty to other brain regions, especially the basal ganglia (Liu et al. 2017; Wu et al. 2021). Additionally, the left caudate, crucial for reward evaluation, computes subjective values across tasks, reward types, and decision stages, influencing preference formation and choice execution (Clithero and Rangel 2014; Verdejo-Garcia et al. 2018). Notably, its activation induces risk seeking, especially in typical risky decision-making paradigms where higher risks are associated with greater rewards, suggesting that reward incentives drive risk preference (Venkatraman et al. 2007; Gujar et al. 2011; Pletzer and Ortner 2016; Huo et al. 2023). In summary, our findings suggest that processing risk- and reward-related information is a fundamental neural activity in risk seeking, with rewards potentially driving risk seeking despite perceived risks.

4.2 Distinct Neural Activity of Risk Aversion

The current study revealed that risk aversion elicits distinct neural activations in the left MTG and left ACC, contrasting with risk seeking. These regions are integral to the cognitive control network (Dong and Potenza 2016; Khaleghi et al. 2020; Lin and Feng 2024). The ACC plays a role in modulating and inhibiting risk seeking (Fukunaga et al. 2013; Knutson and Huettel 2015; Dong and Potenza 2016), encompassing functions such as error, conflict, and performance monitoring (Holroyd and Coles 2002; Platt and Huettel 2008; Purcell et al. 2021; Román et al. 2019). Elevated ACC activation may signify heightened risk awareness, prompting individuals to prefer risk aversion (Rudorf et al. 2012). For example, reduced ACC, extending to the MTG, activity in nicotine-dependent individuals is linked to impaired impulse control (Luijten et al. 2014; Nestor et al. 2011; Palombo et al. 2015). Consistent with prior research, significant left MTG activation was observed during risk aversion (Zhu et al. 2023). The MTG is not only involved in planning and judgment, but is also associated with social cognitive processing (van der Meer et al. 2010; Van Hoeck et al. 2015; Dong and Potenza 2016). During risky decision-making, self-related decision-making beliefs influence behavioral patterns, a process associated with MTG activation (Rodrigo et al. 2014, 2018). Humans, akin to other animals, strive for benefits, avoid harm, and pursue risk-free or certain solutions (Cuevas Rivera et al. 2018). Thus, the pursuit of certainty may more align with individual decision-making beliefs. In summary, our findings suggest that impulse control constitutes a crucial foundation for risk aversion, with cognitive control processing being a unique and fundamental neural activity underlying it.

4.3 Brain Networks in Risk Seeking and Risk Aversion

Surprisingly, we found no overlap between risk seeking and risk aversion. This suggests that these two behavioral patterns may be regulated by distinct neural networks. Dual systems theory proposes that both emotional and cognitive systems influence decision-making (Steinberg 2008; Cui et al. 2022; Fryt et al. 2023), and they include different brain networks (Bjork et al. 2007; Jiang et al. 2024). Our study may support this view by showing distinct effects of these systems on risk seeking and aversion. The emotional system, involving the caudate and right insula, drives risk seeking, while the cognitive system, including the MTG and ACC, promotes risk aversion. MACM analyses revealed that risk seeking regions co-activate with the reward, salience, and cognitive control networks, whereas risk aversion regions involve the cognitive control and valuation networks. These findings align with previous research identifying the reward, default, salience, and cognitive control networks in risky decision-making (Krain et al. 2006; Mohr et al. 2010; Poudel et al. 2020; Wu et al. 2021; Wang et al. 2022) and a recent meta-analysis linking risk-taking to cognitive control, reward, and salience networks (Wang et al. 2024). These activation patterns overlap with resting-state networks, where reward, salience, and cognitive control connectivity predict risk decision-making (Ren et al. 2023; Ji et al. 2025). Previous research has shown that functional networks exhibit dynamic activity during risky decision-making, with increased activation of the reward network and decreased activation of the cognitive control network during risk seeking (Jiang et al. 2024). While cognitive control networks are crucial for impulse control, risk aversion, and preferences for safe choices (Rudorf et al. 2012; Dong and Potenza 2016), significant activation of the reward network by large rewards can override the cognitive system, leading to a preference for risk seeking and altering behavioral patterns (Liu et al. 2017; Nagel et al. 2018). The asynchronous development of these two networks during adolescence also positions the neural substrates of risk seeking and risk aversion within distinct yet interacting neural networks (Bjork et al. 2007; Marquez-Ramos et al. 2023; Jiang et al. 2024). In summary, our findings suggest that risk seeking is primarily governed by the emotional system, while risk aversion is primarily regulated by the cognitive system, both residing within two relatively distinct yet interconnected neural networks, and the imbalance between these two determines an individual's decision-making pattern.

However, the results of the MACM analyses may also reveal an alternative possibility. Both risk seeking and risk aversion seeds showed co-activation in the cognitive control network, suggesting potential unified network governance. Many studies emphasize the interaction between these two networks—specifically, the reward network's potential to promote cognitive control (Somerville and Casey 2010), and the cognitive control network's role in restricting the reward system (Liu et al. 2017; Nagel et al. 2018). Furthermore, during risk decision-making, individuals making rule-based choices, whether impulsive or deliberate, are likely to undergo a process of value computation (Van Duijvenvoorde et al. 2015; Wang et al. 2024). Thus, their distinct brain regions may function as integrated nodes modulating behavioral patterns. However, our primary analyses revealed no significant overlap, likely due to overlapping literature inclusion. Contrasting conditions within studies may suppress shared activation. This methodological approach, while limiting overlap detection, enhances differentiation between risk seeking and aversion mechanisms. Therefore, future research should probe whether these structures form unified networks and elucidate their organizational and modulatory dynamics.

4.4 Limitations

The present study is subject to several limitations. Primarily, there is an insufficiency of neuroimaging studies with a focus on risk aversion, and different experimental paradigms, definitions of risk aversion and risk seeking, as well as experimental manipulation, are utilized in different experiments. The results obtained may be considered inadequate due to these factors. Therefore, it is recommended that future studies conduct further analyses across different paradigms and designs. Secondly, in different decision-making contexts (e.g., loss vs. gain), risk aversion and risk seeking may involve distinct neural activity patterns. However, this study does not aim to differentiate between these factors. Future research should more comprehensively focus on this difference. Thirdly, risk decision-making is generally divided into two stages: valuation and response selection (Liu, Liu, et al. 2022). The present study has not classified different neural networks recruited for these stages. The analysis and comparison of the different stages of decision-making is a research avenue that could be pursued in future studies, which would help to better understand the reasons for the occurrence and shift of different behavioral modes. Finally, while our findings more strongly support the modulation of risk seeking and risk aversion by two interacting neural networks, this interpretation is not exclusive. Whether the implicated brain structures constitute distinct networks or an integrated network requires further analytical validation.