Fig 1.
Schematic timeline of the gambling task.
Participants performed a gambling task in the presence of a human observer, choosing between a safe (certain) option and a gamble (probabilistic) option. Both the participant and the observer were informed that the observer would evaluate multiple dimensions of the participant’s gambling behavior via a camera and fill in a predefined form and that the evaluation results would not be revealed to the participants. Participants were instructed to look at the observer’s face, which was displayed either as an avatar or a real human and were told that the avatar’s facial expressions and movements mirrored those of the real human observer. Participants were not given any specific guidance regarding the role of the observer’s facial expression in the task. At the start of each trial, a 4-s video of the observer’s face (unbeknownst to participants, this was a pre-recorded video) was presented before the choice phase. If the safe option was chosen, a coin image appeared, and the participant received the reward presented in the option. If the gamble option was chosen, the participant received the outcome and then viewed the observer’s facial expression feedback (a 2-s pre-recorded video displaying either a positive or negative expression). Four different avatars were used (two female avatars: AF1, AF2; and two male avatars: AM1, AM2) along with two real human observers (HF1, HM1). For privacy considerations, we replaced pictures of real human observers (HF1, HM1 in the top-right human panel) with AI-generated images which were created by using the AI Human Images (Beta) (https://www.photo-ac.com/main/genface). Each participant was paired with a same-sex human observer, whom they met in person before the experiment began.
Fig 2.
The avatar condition increased gambling choices compared with the human condition.
Higher gambling rates were observed in the avatar condition in both the behavioral (a) and fMRI (b) experiments. The top panels plot the average gambling rates against the expected value differences between the gamble and safe options for the behavioral (a) and fMRI (b) experiments (red: avatar condition, blue: human condition). The bottom panels show the inter-individual averages of the intra-individual gambling rate differences (avatar–human). Error bars represent the standard error of the means. A t test comparing gambling rates between the avatar and human conditions showed significant differences, with p-values of 0.012 and 7.9 × 10−3, respectively. The most significant p-values for each panel are indicated by asterisks: p = 0.015 (a) and p = 8.7 × 10−³ (b). Yellow shaded boxes highlight additional significant p-values: p = 2.3 × 10−³ (a) and 1.1 × 10−³ (b). Numerical data are provided in S1 Data (a) and S2 Data (a).
Fig 3.
Difference in gambling rates correlated with differential feedback uncertainty coefficient.
. A significant correlation was observed between the differential average gambling rate (%) and the differential feedback uncertainty coefficient (i.e.,
). Both behavioral and fMRI samples are included in the dot plot (p = 1.4 × 10−28, R² = 0.80). Each point represents a single participant. Numerical data are provided in S1 Data (a, b) and S2 Data (a, b).
Fig 4.
Inter-personal reactivity scores correlated with differential .
We examined correlations between self-reported personality scores (IRI, LSAS, and STAI) and differential (a–c). The IRI score was significantly correlated with
(p = 0.027 and R2 = 0.17), whereas LSAS and STAI scores did not show significant correlations. (d) The correlation between IRI sub-scores (IRI_FA, IRI_PT, IRI_EC, and IRI_PD) and
. Among these sub-scores, only IRI_EC showed a significant correlation (p = 0.019 and R2 = 0.093), as indicated by the black regression line. The other sub-scores were not significantly correlated (IRI_FA: p = 0.27 and R2 = 0.078, IRI_PT: p = 0.13 and R2 = 0.067, and IRI_PD: p = 0.24 and R2 = 0.088). Robust regression was used to minimize the effect of outliers, and p-values were corrected for multiple comparisons using the Benjamini–Hochberg method. Numerical data are provided in S2 Data (b, c).
Fig 5.
Amygdala response to feedback uncertainty explains differential risk-taking in avatar and human conditions.
(a) Whole-brain subtraction analysis comparing avatar and human observer conditions revealed lower amygdala activity in the avatar condition. Peak activity was identified in the CM (MNI coordinates: [20 −4 −12] and PFWE = 8.5 × 10−6, T-value = 8.21, Z-value = 6.46). For display purposes, a threshold of punc. < 10−4 and k > 100 was applied (y = −4). (b) To find brain regions associated with differential valuation of feedback uncertainty (), we weighted differential activity correlated with feedback uncertainty (avatar–human) by the differential behavioral coefficient for feedback uncertainty (i.e.
). A significant correlation was observed in the CM, with an FWE-corrected p-value of 0.039 at the peak (small-volume corrected using the Anatomy toolbox CM ROI). For display purposes, a threshold of punc. < 0.001 and k > 10 was applied (y = −6). (c) Activity correlated with differential feedback uncertainty at the center of an independent anatomical ROI of the CM is plotted against the differential feedback uncertainty coefficient (
). The negative slope indicates that participants who exhibited greater risk-taking behavior in the avatar condition had a lower feedback uncertainty-correlated response in the CM during interactions with avatars. (d) Participants were divided into two groups: risk-takers against avatars (RTA; left panel) with higher values of
in the avatar condition
and risk-takers against humans (RTH; right panel) with lower values of
in the avatar condition
. CM responses to feedback uncertainty were compared with zero separately for the avatar and human conditions within each group. In RTA (left), the CM response was significantly negative in the avatar condition (p = 4.7 ×10−2, t = −1.73), while in RTH (right), the CM response was negative in the human condition (p = 5.6 × 10−4, t = −3.79). Asterisks indicate statistical significance. Numerical data are provided in S2 Data (b, d).
Fig 6.
Differential response of the CM to feedback uncertainty correlates with the IRI_EC score.
(a) The differential response of the CM to feedback uncertainty is plotted against the IRI_EC score and shows a significant correlation (p = 0.042, R2 = 0.071). (b) CM responses to feedback uncertainty are plotted against the IRI_EC score separately for the avatar and human conditions. We found a significant collation in the avatar condition (left; p = 5.1 × 10−3 and R2 = 0.098) but not in the human condition (right; p = 0.65, R2 = −0.0021). Regression lines indicate a significant correlation. Robust regression was used in the estimation to mitigate the effect of outliers. Numerical data are provided in S2 Data (c, d).
Fig 7.
Feedback uncertainty-correlated activity in the vSTR and vACC.
(a, d) Differential activity correlated with feedback uncertainty (avatar–human) was weighted by the differential behavioral coefficient (valuation) for feedback uncertainty (i.e. ). Significant activity was identified in the vSTR (a) (peak MNI coordinates: [6,16, −4], FWE-corrected p = 1.5 × 10-5) and the vACC (d) (peak MNI coordinates: [−6, 42, −12], FWE-corrected p = 0.030). (b, e) Activity correlated with differential feedback uncertainty at the center of independent anatomical ROI of the vSTR (b) and vACC (e) is plotted against the differential feedback uncertainty coefficient (i.e.
). (c, f) Responses (beta values) of the vSTR (c) and vACC (f) to feedback uncertainty are displayed separately for the avatar and human conditions within RTA and RTH, following the same format as in Fig 5d. In the vSTR (c), feedback uncertainty-correlated activity was negative in the human condition of RTH (p = 0.024, t = −2.10) but not in the avatar condition of RTA (p = 0.15, t = −1.02). No significant difference was found between the avatar and human conditions in RTA (p = 0.075, t = −1.85). In the vACC (f), feedback uncertainty-correlated activity showed no significant difference between the avatar and human conditions in either RTA or RTH. The asterisk indicates statistical significance. Numerical data are provided in S2 Data (b, d).