Neural Representation in mPFC Reveals Hidden Selfish Motivation in White Lies

Identifying true motivation for Pareto lies, which are mutually beneficial for both the liar and others, can be challenging because different covert motivations can lead to identical overt behavior. In this study, we adopted a brain-fingerprinting approach, combining both univariate and multivariate analyses to estimate individual measures of selfish motivation in Pareto lies by the degree of multivoxel neural representation in the mPFC for Pareto lies conforming with those for selfish versus altruistic lies in human participants of either sex. An increase in selfish motivation for Pareto lies was associated with higher mean-level activity in both ventral and rostral mPFC. The former showed an increased pattern similarity to selfish lies, and the latter showed a decreased pattern similarity to altruistic lies. Higher ventral mPFC pattern similarity predicted faster response time in Pareto lies. Our findings demonstrated that hidden selfish motivation in white lies can be revealed by neural representation in the mPFC. SIGNIFICANCE STATEMENT True motivation for dishonesty serving both self and others cannot be accurately discerned from observed behaviors. Here we showed that fMRI combining both univariate and multivariate analyses can be effectively used to reveal hidden selfish motivation of Pareto lies serving both self and others. The present study suggests that selfish motivation for prosocial dishonesty is encoded primarily by increased activity of the ventromedial and the rostromedial prefrontal cortex, representing intuitive self-serving valuation and strategic switching of motivation depending on beneficiary of dishonesty, respectively.

Pareto lies (Erat and Gneezy, 2012), where the results of dishonesty are mutually beneficial 49 for both the liar and others. Two different psychological mechanisms have been proposed to 50 contribute to increasing Pareto lies. The presence of another beneficiary 1) may help justify 51 dishonesty that will benefit oneself, or 2) may trigger genuine care and concern about the 52 benefits others receive (Gino et al., 2013). As Pareto lies are both self-serving and altruistic, 53 recognizing the exact mechanisms engaged from the dishonest behavior alone poses a 54 challenge. 55 We applied the concept of brain fingerprinting technique (Ahuja and Singh, 2012) to 56 neuroimaging data to gain further evidence for inferring an individual's covert motivation of 57 Pareto lies. In this approach, a target without an explicit label can be classified based on the 58 degree to which the brain response to the target resembles the two known categories. More 59 specifically, contexts in which dishonesty may benefit both self and others may appear as a 60 selfish opportunity to some as they may benefit from dishonesty, whereas the same context 61 may be viewed by others as an altruistic opportunity to benefit others. 62 Several sub-regions of the medial prefrontal cortex (MPFC) serve a crucial role in 63 moral judgment and generation of dishonest behavior. For example, judgments of the 64 display (0.5 sec) followed by the number of points assigned to each side of the screen (1~3s), 171 dot-screen display for the individually calibrated length of time, question display (until 172 decision), and the result of choice display (0.7 sec). 173 174

Behavioral data analyses 175
The overall effect of point and beneficiary on dishonest decisions was assessed by entering 176 the percentage of wrong choices to a repeated-measures analysis of variance (rmANOVA) 177 with the beneficiary (Self, Other, Both) and point (0, 1, 2) as within-subject factors. As the 178 points could only be obtained by reporting the wrong answer, we expected a higher 179 percentage of dishonest decisions in point 1 and 2 conditions compared to point 0 condition. 180 We first normalized the RT data within each subject over all trials, and then averaged 181 them separately for dishonest decisions in each condition. For participants who were always 182 honest in certain conditions and whose average RT could not be calculated were excluded 183 from correlation analyses that includes RT data. The correlation between RT data and other 184 indices were obtained using Spearman's rank correlation as the sample size after exclusion 185 resulted in 28, which may be insufficient to use Pearson's correlation. The average 186 normalized RT of each condition were calculated and entered in the rmANOVA for all the 36 187 participants. 188 189

Neuroimaging procedures and analyses 190
FMRI data acquisition and preprocessing 191 FMRI data were acquired using a 3.0 T Siemens Magnetom Trio MRI scanner with a 12-192 channel head matrix coil located at the Korea University Brain Imaging Center. We obtained 193 the T2*-weighted functional images using gradient-echo echo-planar pulse sequences We preprocessed the data using the SPM12 (Wellcome Department of Imaging 206 Neuroscience, University College of London, London, UK). Images were temporally 207 corrected for interleaved slice acquisition, and then realigned to the first volume to correct for 208 head motion and a mean image was created for each participant. The realigned images were 209 normalized to the standard Montreal Neurological Institute (MNI) EPI template, resampled to 210 2 × 2 × 2 mm voxels, and spatially smoothed using a Gaussian kernel with an 8 mm full 211 width at half maximum (FWHM). 212

st level univariate analyses 213
A first-level generalized linear model (GLM) was estimated to create contrasts for each 214 beneficiary condition. Onset times for the three beneficiaries (Self, Other, and Both), with the 215 three points (0, 1, and 2 points) information presentation and decisions for each nine 216 condition as well as six head-motion parameters were included as regressors after being 217 convolved with a standard hemodynamic response function. The brain regions reflecting the 218 point by beneficiary interaction effect were identified by first generating three contrast 219 images (i.e., one for each beneficiary condition) by combining Point 1 and 2 conditions and subtracting point 0 condition at decision onset (e.g., [Point 1+ Point 2] − Point 0 for Self 221 condition), and then entering the contrasts into an rmANOVA. These three contrasts were 222 used in the pattern classification analyses as well. We used these contrasts rather than the 223 contrast of dishonest vs. honest decisions because 1) some participants do not have enough 224 trials of dishonest decision in some conditions, 2) the focus of this research was to distinguish 225 individual motivation and neural mechanisms that underlie the processing of immoral 226 opportunities to gain from Pareto lies. 227

nd level univariate analyses 228
To explore brain regions representing the main effects of beneficiary and point, and the 229 interaction effect between beneficiary and point, three rmANOVAs were conducted. The 230 beneficiary main effect was assessed by constructing first-level contrast images for each 231 beneficiary at decision onset by combining trials overall points for each beneficiary (i.e., 232 Point 0 + Point 1 + Point 2 separately for each of Self, Other, and Both), which were entered 233 into a rmANOAV. In addition, contrasts for each point overall beneficiaries were built and 234 entered into a rmANOVA to examine the main effect of the points. All the statistical maps 235 reported were thresholded at the whole-brain FWE corrected p < 0.05 at voxel-level. 236 conditions in which participants were motivated to lie (i.e., points would be given when 246 lying) with the condition in which participants had no reason to lie (i.e., no point would be 247 given when lying) for each beneficiary. Supporting our rationale for this analysis, the 248 behavioral data showed that participants were induced to lie by the existence, rather than the 249 amount, of available point to be earned. The classifier was first trained on the MPFC activity 250 pattern for Self and Other beneficiary conditions to distinguish between neural patterns 251 associated with the opportunities to lie for Self and Other. Conducting classification using 252 moderately smoothed data is thought to be effective ( The individual measure of selfish motivation in Pareto lies was defined as how 266 certain each individual's MPFC activity pattern during Both condition was classified as Self. 267

Neural signatures of selfish or altruistic motivation for dishonesty: Multivariate analysis
As such, the signed distance of individual Both contrast to the hyperplane separating Self and 268 Other was calculated and used as the self-class confidence scores (SCCS), where a higher 269 score translates into higher certainty of being classified into Self. Computationally, this score 270 Selfish Motivation in White Lies 14 was calculated by taking the dot product of individual Both contrast and the classifier weight 271 map and adding the intercept term. 272

Second-level regression and correlation analyses with the SCCS 273
Multiple regression analyses were performed to explore the neural mechanisms behind selfish 274 motivation in each beneficiary's opportunities. In these analyses, the SCCS were regressed on 275 the contrast maps of Self, Other, and Both conditions separately. 276

Representational similarity analyses 277
The VMPFC, RMPFC, and precuneus masks were generated from the result of the whole-278 brain FWE corrected multiple regression analysis of the SCCS with Both contrasts (VMPFC 279 cluster peak: x = -2, y = 46, z = -8; RMPFC cluster peak: x = 8, y = 34, z = 14; precuneus 280 cluster peak: x = 10, y = -60, z = 34). We extracted the neural activity of Self and Both 281 conditions in the ROIs for each participant from the Self, Other, and Both contrasts used in 282 the univariate analyses. We also calculated the pattern similarity as the Kendall's Tau (Popal 283 et al., 2019) between the neural activity patterns in each ROI of Self and Both conditions, and 284 those of Other and Both conditions for each participant. Then, the calculated pattern 285 similarities were correlated with the SCCS. 286

289
Behavioral results 290 We first tested whether participants were more likely to report incorrectly when points were 291 available, as this suggests dishonesty, and whether such dishonesty is modulated by the 292 beneficiary of the points. We carried out a two-way repeated-measures ANOVA (rmANOVA) 293 to assess the effect of points and beneficiary on the participants' decisions to be dishonest.

Univariate analysis result 322
We first investigated how opportunities to gain from dishonesty for different beneficiaries 323 and different amounts of points are represented in the brain. A first-level generalized linear 324 model (GLM1) was built, including onset times for three beneficiaries (Self, Other, and Both), 325 three points (0, 1, and 2 points) information presentation, and decisions for every nine 326 combinations of beneficiaries and points, which were all convolved with a standard 327 hemodynamic response function. The model also included six motion parameters as nuisance 328 regressors. We created the first-level contrasts for each beneficiary (e.g., Point 0 + Point 1 + 329 Point 2 for Self trials), and each point (e.g., Self + Other + Both for Point 0 trials) to examine 330 brain regions showing the difference in the activation at the time of decision based on the 331 beneficiaries and points, and then entered them into two separate second-level rmANOVAs 332 to assess the main effects of beneficiary and points. The analyses revealed a unique RMPFC 333 response for each beneficiary ( Fig. 2A, x = 0, y = 40, z = 20; whole-brain FWE corrected at 334 voxel-level p < .05 unless stated otherwise), showing the highest activity during Self, and 335 lowest activity during Both conditions. Furthermore, a larger RMPFC cluster extending into the DMPFC was revealed to show differences in the activity to the different amounts of 337 points (Fig. 2B, x = 0

Neural signatures of selfish or altruistic motivation for dishonesty: Univariate analysis 345
We first conducted a second-level t-test on Self vs Other contrast and Other vs Self contrast 346 to identify the distinctive neural features related to selfish or altruistic motivation for 347 dishonesty. No voxels survived the correction in both contrasts, which confirms our 348 prediction that a univariate analysis may not be sensitive enough for detecting subtle 349 differences in neural representation between selfish and altruistic motivation for lying. 350

Neural signatures of selfish or altruistic motivation for dishonesty: Multivariate analysis 351
For a further differentiation of the neural signatures of selfish or altruistic motivation for 352 dishonesty in Both condition, we trained a pattern classifier (For more detailed information, 353 see Methods section 'Multivariate classification of the neural representation of motivations 354 for Pareto lies') to differentiate neural patterns in the MPFC associated with the opportunities 355 to lie for Self and Other. We used the trained classifier to classify individuals' neural patterns 356 for Both conditions to estimate one's covert motivation underlying moral decisions in 357 situations where dishonesty would benefit both Self and Other (Fig. 3A). The classifier was 358 trained across, rather than within, participants to ensure its generalizability. The final 359 classifier showed 98.61% accuracy in distinguishing Self vs Other contrast images. The 360 classification results showed that Both was classified as Self in 17 out of 36 participants and as Other in the remaining 19 participants. The percentage of Pareto lies would not differ 362 between the two groups (t(34) = .664, p = .511), consistent with our hypothesis. 363

Neural evidence for selfish motivation in Pareto lies 364
Next, we identified neural regions related to the degree of selfish motivation in Pareto lies, 365 which was defined as the self-class confidence scores (SCCS). We calculated the SCCS by 366 taking the signed distance of individuals' Both contrast to the hyperplane separating Self and 367 Other. The SCCS ranged from -2.06 to 2.11 with the mean value of 0.03 and SD of 0. two-sided; Fig. 5D). In addition, the SCCS correlates negatively with the degree of similarity of the RMPFC activity pattern between Other and Both conditions (Pearson's r(36) = -.401, p 387 = .013, two-sided; Fig. 5C), but not between Self and Both conditions (Pearson's r(36) = -388 .071, p = .681, two-sided; Fig. 5A). Tests for differences in dependent correlations showed 389 that the correlation coefficients of Self-Both similarity and Other-Both similarity in RMPFC 390 cluster are significantly different (z = 1.651, p = .049, one-tailed; Fig. 5E), and the correlation 391 coefficients of Self-Both similarity and Other-Both similarity in VMPFC cluster are 392 marginally different (z = 1.567, p = .057, one-tailed; Fig. 5F). In the precuneus cluster, the 393 SCCS showed no correlation with the degree of pattern similarity between Self and Both 394 We also examined whether and how selfish motivation in Pareto lies is differentially 419 associated with the neural representations in self-and other-benefiting dishonest 420 opportunities. To achieve this, we regressed the SCCS on the contrast map of the Self and 421 Other conditions separating them into two multiple regression analyses. During Self 422 condition, the activities in the VMPFC (x = 8, y = 44, z = -10) and the ventral striatum (VS: x 423 = -16, y = 8, z = -8) showed significant positive correlations with the SCCS (Fig. 6B). This 424 suggests that as individuals consider opportunities for Both to be closer to opportunities for 425 Self, self-benefiting dishonest opportunities engaged VMPFC and ventral striatum to a larger 426 extent. During the Other condition, a significant positive correlation was observed between 427 individual SCCS and the activities in VMPFC (x = -6, y = 48, z = -6), and VS (x = -18, y = 8, 428 z = -2), similar to the observation made for the Self condition. However, the RMPFC (x = 6, y 429 = 52, z = 16) and left anterior insula (AI: x = -26, y = 20, z = -12) additionally showed 430 significant positive correlations with the SCCS during Other condition (Fig. 6C). These 431 findings indicate that other-benefiting dishonesty additionally engages RMPFC and AI 432 among those Pareto lies that appeared to be primarily driven by selfish motivation. beneficiary was Self, and the lowest when the beneficiary was Both. The same region was 509 also more active when more points were available. This activity may not be related to the 510 increased motivation for dishonesty because participants lied more for Both than for Other 511 even to the level of Self, which is opposite to the pattern of neural activity in this region 512 across conditions. This observation led to a more plausible speculation that the activity in this 513 region reflects a conflict between the urge to gain points and the guilt resulting from 514 dishonesty, which is in line with the previous research showing increased ACC activity 515 associated with moral conflict or guilt (Fourie et al., 2014;Abe et al., 2018). The fact that this 516 region showed the lowest activity in Both condition suggests that people experience the least 517 moral conflict when dishonesty can benefit both the liar and another person. In addition, the 518 activity in this region was also stronger among those with higher SCCS, possibly reflecting 519 an increased moral conflict or guilt due to higher selfish motivation for Pareto lies. 520 We found no evidence for neural signatures of altruistic motivation for Pareto lies 521 because there was no cluster in the brain showing a negative correlation with the SCCS even 522 at a lenient threshold (p < 0.005 uncorrected). It has been established that the magnitude of 523 the BOLD response is sensitive to change in excitation-inhibition balance in the cortical 524 microcircuits involving the pyramidal projection neurons interacting with local GABAergic 525 interneurons, which may reflect mismatch or prediction error-related feedback signals 526 (Logothetis, 2008). Given this, larger negative SCCS, or higher Other-classification 527 confidence score may not necessarily involve significant increase in excitation-inhibition 528 balance, because the multivoxel representation analysis can be immune to such a change in 529 excitation-inhibition balance (Logothetis, 2008). 530 This study provides a novel methodological approach combining the potential 531 benefits of univariate and multivariate analyses. Despite its superior sensitivity to detecting 532 subtle differences in neural representation among different psychological states, MVPA has 533 Selfish Motivation in White Lies 25 not been considered appropriate for identifying the exact neural mechanisms leading to the 534 psychological state at question (Kohoutová et al., 2020), and insensitive to inter-subject 535 variability in mean activation across voxels within a region of interest, which can be better 536 captured by a conventional univariate analysis (Davis et al., 2014). Consistent with the 537 dissociation between univariate and multivariate analyses, multivariate patterns showed a 538 higher similarity of Both to Self vs. Other, whereas univariate patterns showed the opposite, 539 that is, the higher similarity of Both to Other vs. Self, with the RMPFC clusters additionally 540 recruited in Both and Other. This study demonstrated that a univariate analysis can be 541 combined with MVPA to effectively locate the neural regions where the neural 542 representations contributed maximally to the global pattern classification. 543 In conclusion, this study demonstrates that fMRI can be used to infer hidden selfish 544 motivation in Pareto white lies by adopting the brain fingerprinting approach combining both 545 univariate and multivariate analyses. This technique allowed us to estimate individual 546 differences in motivation for Pareto lies, based on distinctive patterns of activity across 547 functionally dissociable MPFC subregions, including VMPFC and RMPFC. We believe that 548 this study will provide a novel and powerful research method and theoretical contributions to 549 the current efforts of understanding complex motivations underlying moral behaviors. 550 551 Figure 5. Correlation between the SCCS and the representational similarity between pairs of 724 conditions. In the RMPFC, the SCCS correlated negatively with the degree of pattern 725 similarity between Other and Both conditions (C), but not between Self and Both conditions 726 (A). In the VMPFC, the SCCS correlated positively with the degree of pattern similarity 727 between Self and Both conditions cluster (B), but not between Other and Both conditions (D). 728 Fisher's r-to-z transformed correlation coefficients in the RMPFC (E) and VMPFC (F). 729 730 Figure 6. Correlation with Self-class confidence score (SCCS) for Self and Other conditions. In addition, their neural representations in the VMPFC were similar between selfish and