Monkeys Predict US Elections

How people vote often defies rational explanation. Physical traits sometimes sway voters more than policies do. But why? Here we show that rhesus macaques, who have no knowledge about political candidates or their policies, implicitly predict the outcomes of US gubernatorial and senatorial elections based solely on visual features. Given a pair of candidate photos, monkeys spent more time looking at the loser than the winner, and this gaze bias predicted not only binary election outcomes but also vote share. Analysis of facial features revealed candidates with more masculine faces were more likely to win an election, and vote share was a linear function of jaw prominence. Our findings endorse the idea that voters spontaneously respond to evolutionarily conserved visual cues to physical prowess and that voting behavior is shaped, in part, by ancestral adaptations shared with nonhuman primates.


Introduc3on
Former US President Bill Clinton once remarked that "Americans prefer strong and wrong over weak and right", highlighJng the fact that voters someJmes choose to the detriment of their own economic wellbeing and personal liberty (1)(2)(3).Evidence-based deliberaJon by well-informed voters is a bedrock assumpJon of a funcJoning democracy (4,5), so understanding, and potenJally miJgaJng, deviaJons from raJonal decision-making by voters is criJcal for the health of our body poliJc.
Ecological raJonality may help explain why the physical appearance of a candidate for office impacts their electability (15).Todorov and colleagues showed that people can accurately predict elecJon outcomes based on very short visual exposure to candidate photos alone (16)(17)(18), including unfamiliar elecJons in foreign countries (19,20).Remarkably, even preschool children can predict with above chance accuracy who will win an elecJon based solely on candidate photos (21).In these studies, parJcipants were asked to judge the competence, trustworthiness, and warmth of candidates, with competence judgments best predicJng elecJon results (16)(17)(18)22).ParJcipants further expressed the belief that competence is one of the most important a3ributes for poliJcians (16,22,23).These findings led Todorov and colleagues to propose a dual-process framework, in which voters assess candidate competence, parJally on the basis of visual appearance, and cast votes accordingly.
While abundant evidence supports the idea that people quickly form opinions of others based on first impressions, which include visual appearance (24), the mechanism underlying these biases remains unclear, and the reasons why such biases persist in contemporary society remain obscure.Here we test the hypothesis that masculinized facial features, such as wide jaws and less prominent cheekbones, contribute to elecJon outcomes, and that the impacts of physical cues to social dominance on decision making are evoluJonarily conserved.To do so, we build upon previous research showing children can predict elecJon outcomes based on candidate pictures alone (21), and extend it by tesJng whether macaque monkeys (Macaca mula(a) spontaneously do so as well.Prior research shows macaques a3end to faces (25,26), trade alimentary rewards for brief glimpses of dominant monkeys (27,28), preferenJally follow the gaze of dominant monkeys (29), and fixate longer on the faces of subordinate and female rather than dominant male monkeys (30,31).Thus, nonhuman primates prioriJze visual social informaJon and are sensiJve to cues associated with social dominance (32).
In the current study, monkeys freely examined pairs of candidate photos from prior U.S. senatorial, gubernatorial, and presidenJal elecJons.They displayed a consistent bias to look at losing candidates over winning candidates."Votes" calculated from monkey gaze biases predicted elecJon results as well as humans did based solely on visual informaJon, and they were linearly correlated with vote share.To explore the contribuJon of masculinized facial features to this behavioral pa3ern, we measured the width of the jaw relaJve to the cheekbone in each candidate photograph and found their raJo predicted monkey gaze behavior and, moreover, linearly forecast vote share in US elecJons.Our findings endorse the idea that voters insJncJvely respond to evoluJonarily conserved visual cues to physical prowess and masculinity, and that voJng behavior is shaped, in part, by ancestral adaptaJons shared with nonhuman primates.

Results
We examined the spontaneous gaze pa3erns of male macaques presented with pairs of candidate photos drawn from U.S. senatorial and gubernatorial elecJons (Figure 1a).To minimize familiarity effects on gaze behavior, in each session each pair of candidates was displayed only once.Each monkey (n = 3) parJcipated in 2 sessions, with an inter-session interval of at least 3 days (n = 6 sessions total).All monkeys looked more at loser rather than winner pictures (Figure 1a), someJmes avoiding the winner altogether (Supplementary Figure 1a).Similarly, monkeys showed biased gaze towards female over male candidates (Figure 1a bo3om, Supplementary Figure 1a bo3om).Both biases were significant for all monkeys combined (Figure 1b, loser vs winner: P < 0.001, Wilcoxon signed rank; female vs male: P < 0.001, Wilcoxon rank sum), and consistent across the 3 subject monkeys (Supplementary Figure 1b).
Gaze biases towards losing candidates and female candidates interacted when a female ran against a male (n = 66 pairs making up 24.2% of all the races).Monkeys showed the strongest gaze bias towards female candidates who lost to a male (Figure 1a bo3om, Supplementary Figure 1a bo3om).When a female candidate beat a male opponent (Supplementary Figure 1c), however, the loser bias was someJmes nullified (Supplementary Figure 1c bo3om, Supplementary Figure 1d right).By contrast, for male-male candidate pairs (n = 201 pairs making up 73.6% of all the races) the loser gaze bias was linear and strong (P = 0.002, Wilcoxon signed rank, Figure 1c).
We assume monkeys were unaware of the idenJty, party affiliaJon, or policies of any of the candidates.The only informaJon accessible to them was candidate photos, suggesJng monkeys predicted elecJon outcomes and vote share based on visual features.Prior studies have shown monkeys are vigilant to informaJon about social dominance: they will pay for a brief glimpse of a dominant male, but then rapidly look away to avoid direct eye contact, which is considered a sign of aggression (27,28,30,31,(35)(36)(37). Due to their viewing bias towards female candidates, we surmised that monkeys were sensiJve to masculine facial features, which are shared by monkeys and humans and associated with dominant social status-which for male macaques is correlated with physical prowess, peaks in adulthood, and declines in older age (37).For both monkeys and humans, compared with females, males on average have narrower cheekbones, wider jaws, higher facial width to height raJo (FWHR), and higher lower face prominence (LFP), resulJng from high circulaJng testosterone during puberty (38)(39)(40).We measured all these features in candidate photos (Figure 2a), and found that jaw prominence-the raJo of jaw width to cheekbone width-accounted for 7-8% of the variance in vote share (gubernatorial races: r = 0.26, P < 0.001, senatorial races: r = 0.29, P < 0.001, overall: r = 0.27, P < 0.001, Figures 2b-c).On average, the winning candidate's jaw was 2% more prominent than the losing candidate's jaw (Supplementary Figure 2a).This pa3ern was consistent across gubernatorial and senatorial races, party affiliaJon, and incumbency (Supplementary Figure 2b).In alignment with previous literature, jaw prominence, cheekbone width, jaw width, FWHR, and LFP all reliably disJnguished males from females in our sample of candidate photos (Supplementary Figure 2c).Only jaw prominence, however, disJnguished faces of winners and losers (Supplementary Figure 2d).Jaw prominence also significantly predicted monkey's gaze bias (r = -0.05,P = 0.004, Figure 2d)-the more prominent the jawline, the more likely monkeys were to avert their gaze.FWHR and LFP also weakly predicted monkey gaze, but in the opposite direcJon (r = 0.04, P = 0.041, and r = 0.04, P = 0.032, respecJvely, Figure 2d).In other words, candidates with the most prominent jaws, relaJve to their cheekbones, received the most votes, and monkeys most strongly avoided looking at them.Taken together, these findings suggest visual cues to physical prowess shape both visual orienJng by monkeys and the choices humans make in the voJng booth.
Gaze bias in monkeys and voJng by humans diverge, however, for female and older candidates.Overall, we found similar relaJonships between jaw prominence, competence raJng, and vote share for female candidates as for males.That is, female candidates with more prominent jaws were rated as more competent (r = 0.30, P = 0.010, Supplementary Figure 2e, right), and also received more votes (r = 0.39, P < 0.001, Supplementary Figure 2f, leX).Monkeys, by contrast, were biased to view female over male pictures, thereby weakening their predicJons of races between females and males (Supplementary Figure 1d, Supplementary Figure 2g).Similarly, monkeys were biased to look at older rather than younger candidates (r = 0.06, P = 0.001), whereas human voters typically prefer older candidates (age vs vote share: r = 0.24, P < 0.001 in our sample, Supplementary Figure 2h; 22).We found no relaJonship between age and jaw prominence (r = 0.00, P = 0.976), or age and competence (r = 0.05, P = 0.261), indicaJng that age represents a dimension independent of masculinity and competence.
In addiJon to female and older candidates, human voters also chose challengers (as opposed to incumbents) and Democrats (as opposed to Republicans) more oXen than monkeys predicted they would (Supplementary Figure 3a).Party affiliaJon is a major predictor of voJng behavior (41), so we examined how well monkeys' gaze biases corresponded with voter behavior in blue (DemocraJcleaning), red (Republican-leaning), and swing states (Figure 2e).We found that monkey gaze behavior predicted swing-state elecJon results best (Figure 2f), and swing-state elecJons were most sensiJve to facial masculinity cues (Supplementary Figures 3b-c).
Finally, we a3empted to predict the upcoming 2024 US presidenJal elecJon.For past presidenJal elecJons (2000-2020), monkeys' predicJons were at chance (50.4 ± 5.7%, P = 0.943, t-test), possibly reflecJng voters' recent preferences for DemocraJc and older candidates.As observed in the gubernatorial and senatorial elecJons, monkeys were biased to look at the DemocraJc over the Republican presidenJal candidate regardless of elecJon outcome (P = 0.042, Wilcoxon signed rank; Figure 2g).This pa3ern was consistent across all 3 monkeys (Supplementary Figure 3d).As for the most recent races involving Donald Trump (2016-2024), monkeys' gaze bias most strongly differenJated Trump from Hillary Clinton, less so from Joe Biden, and not at all from Kamala Harris (Supplementary Figure 3e).Thus, among the 3 most recent democraJc nominees, based solely on visual features, Harris would be predicted to stand the best chance of winning, possibly reflecJng Trump's advanced age (Figure 2h) or voters detecJng qualiJes in Trump other than physical dominance-for example low warmth, honesty, or likeability-that are deemed undesirable for poliJcians, as well as familiarity with his character and past performance, thus accounJng for his vote share in 2020 underperforming predicJons based on his jaw prominence (Supplementary Figure 3f).

Discussion
Here we show for the first Jme that forecasts based on the gaze behavior of monkeys predict the results of U.S. gubernatorial and senatorial races-and do so as well as human adults and children do.SensiJvity to facial masculinity, parJcularly jaw prominence relaJve to cheekbones, best explained monkeys' gaze bias.We surmise that people tested in prior studies similarly predicted elecJon outcomes based on visual cues to physical prowess and dominance, and that vigilance for these features shapes voter choices in real elecJons.
Ecological raJonality provides a potenJal explanaJon for the impact of facial masculinity on voJng behavior (12,13).Masculine facial features, such as wide jaw and prominent lower face are associated with high testosterone (38)(39)(40).Macaque monkeys can infer not only idenJty and reproducJve state from conspecific faces, but also social status, indicaJng sensiJvity to facial masculinity (25,42,43).Humans can also detect facial masculinity, and consider more masculine male faces to be not only more a3racJve (38-40, 44, 45) but also more likely to succeed (46).MechanisJcally, human and nonhuman primates share brain regions and networks that prioriJze visual social informaJon with adapJve value (32,47,48).We hypothesize that in both humans and monkeys these shared structural and funcJonal specializaJons spontaneously detect facial masculinity in candidates, thereby shaping a3enJon and downstream processing, which, in humans, ulJmately impacts voJng behavior.
We note that apparent deep homologies in social a3enJon shared by human and nonhuman primates do not explain all contemporary voJng behavior.Based solely on facial masculinity cues, female candidates are projected to lose most races.Yet voters chose the female candidate about half the Jme (overall female winning probability = 48.8% in our sample), indicaJng other factors besides facial masculinity contribute to voJng decisions.Similarly, voters preferred older candidates, although, based on gaze bias, monkeys found them less dominant looking.Voters also selected challengers and DemocraJc candidates more oXen than predicted by monkeys.These findings indicate factors beyond perceived masculinity and physical prowess shape voJng behavior.
Our findings have implicaJons for poliJcal campaigns.For example, judgment of facial masculinity relies on visual percepJon of low-spaJal frequency informaJon such as jaw width and cheekbone width, which can occur during brief presentaJons of low resoluJon images, such as peripheral glimpses of a campaign flier, thus supporJng print media campaigns.Further, campaigns can strategically use photos emphasizing or de-emphasizing jaw width and cheekbone width (49,50), which may explain why most candidates smile in their official portraits (92.3% of candidates smiled in the photos in our sample), thereby accentuaJng the jawline.Choices also depend on how a decision is framed (6-8).For example, Li3le et al. (51, Study 2) created morphed faces of George W. Bush and John Kerry, and parJcipants were asked to indicate which one they would "vote for to run your country" in a Jme of peace or a Jme of war.In peaceJme, the more Kerry-shaped face was preferred, but in warJme the more Bush-shaped face was preferred.The more Bush-shaped face was judged to be more masculine and dominant, traits parJcipants favored in a Jme of conflict, but less intelligent and forgiving, traits parJcipants favored in peaceJme.Campaigns can frame social issues to their advantage by emphasizing external threats if the candidate is more masculine looking, or domesJc tranquility if the candidate is more feminine looking, which may explain the apparent effecJveness of the Harris-Walz campaign's focus on joy.Overall, our findings compel the development of strategies that encourage voters to become well-informed about candidates and their policies to help overcome evoluJonarily ancient heurisJcs prioriJzing visual predictors of physical prowess and masculinity.

Animals
All procedures reported in this study were approved by the InsJtuJonal Animal Care and Use Commi3ee of the University of Pennsylvania, and performed in accordance with The Guide to the Care and Use of Laboratory Animals.Three male rhesus macaques (M1: C, 15 years old, 15 kg; M2: L, 16 years old, 11 kg; M3: O, 17 years old, 17 kg) parJcipated in the senatorial and gubernatorial elecJon experiment, each for 2 days/sessions, with an intersession interval of at least 3 days (n = 6 sessions in total).SJmulus sets for each session consisted of 124 pairs of gubernatorial candidate pictures and 149 pairs of senatorial candidate pictures (n = 273 pairs of photos in total).Three male rhesus macaques (M1: C, 19 years old, 13 kg; M2: F, 11 years old, 13 kg; M3: L, 20 years old, 11 kg) parJcipated in the presidenJal elecJon experiment, each for 5 days/sessions, with an intersession interval of at least 2 days in-between (n = 15 sessions in total).SJmulus sets for each session consisted of 13 pairs of presidenJal and vice-presidenJal candidate pictures.

Experimental Setup
Each subject monkey sat in a primate chair (Crist Instruments), in a dark room (luminance ~3 cd/m2) facing an LCD monitor (BenQ XL2730, 27'', 2560*1440, 120 Hz).A computer (Dell Precision Tower 5810, custom built) running MATLAB (Mathworks) and Psychtoolbox (52,53) was used to control all aspects of the experiment, including displaying visual sJmuli on the monitor, communicaJng with the eye tracking system (Eyelink, see below), and opening and closing solenoid valves (Christ Instrument) to dispense juice rewards.
During the experiment, the monitor displayed a uniform gray background (luminance ~15 cd/ m2).At the beginning of each trial, a central fixaJon spot (0.5°, luminance ~35 cd/m2) was illuminated, and the monkey brought his gaze within a 3.0° diameter fixaJon window to iniJate image display.Subsequently a pair of luminance-balanced, black-and-white images of compeJng poliJcal candidates were rendered on each side of the screen for 2.5 seconds.A blank screen then replaced both images, and a fixed amount of juice (0.5 ml) was delivered to the subject monkey.The inter-trial interval was 2-3 seconds (ji3ered).AXer iniJal fixaJon, the subject monkey was free to look anywhere during sJmulus presentaJon as well as during the inter-trial interval.Eye posiJon was recorded with an infrared eye tracking system, Eyelink 1000 Plus (SR Research, primate mount), sampled at 1,000 Hz, exported as EDF files, and then preprocessed with a custom MATLAB script (Edf2Mat, h3ps:// github.com/uzh/edf-converter).

S3mulus presenta3on
Candidate photos for real U.S. gubernatorial, senatorial, and presidenJal general elecJons were presented in pairs.In each session, one set of races (gubernatorial, senatorial, or presidenJal) was presented in its enJrety with each candidate pair displayed once and once only.The order (i.e. which race) and side (i.e.leX or right side of the screen) of presentaJon were randomized.
Gubernatorial races: The procedure followed 18.From the Almanac of American Poli<cs, a list was compiled of all gubernatorial races from 1995 to 2006, excluding races with highly familiar poliJcians (e.g., Arnold Schwarzenegger).There were 124 races (248 candidates) in total, with 36 female candidates, and 74 incumbents in running.Pictures of the winner and the runner-up were collected from various Internet sources (e.g., CNN, Wikipedia, and local media sources).The image of each poliJcian was cropped, placed on a standard background, and converted to grayscale.
Senatorial races: The procedure followed 16, 18.From the Almanac of American Poli<cs, a list was compiled of all senatorial races from 2000 to 2008, excluding races with highly familiar poliJcians (e.g., Hillary Clinton).There were 149 races (298 candidates) in total, with 44 female candidates, and 120 incumbents in running.Pictures of the winner and the runner-up were collected from various Internet sources (e.g., CNN, Wikipedia, and local media sources).The image of each poliJcian was cropped, placed on a standard background, and converted to grayscale.
Presiden3al races: A list was compiled of all presidenJal races from 2000 to 2024, as well as the vice-presidenJal race of 2024.For 2000-2012, the most Jme-appropriate official portrait was chosen for each candidate.For example, for Obama vs Romney 2012, the 2012 presidenJal portrait of Obama and 2006 gubernatorial portrait of Romney were chosen.For more recent elecJons (2016-2024), since candidate age had become a major issue, in addiJon to the most Jme-appropriate official portraits, we also selected one specific elecJon year photo for each candidate, from naJonal convenJons or major campaigning events.The image of each poliJcian was cropped, placed on a standard background, and converted to grayscale.

Data Analysis
Saccades and fixaJons were determined using Eyelink 1000 Plus online parser.All subsequent data analysis was done in custom MATLAB scripts.All staJsJcal tests were two-tailed.For hypothesis tesJng between two samples, a non-parametric Wilcoxon signed rank test (for paired samples) or Wilcoxon rank sum test (for unpaired samples) was used.For comparison among more than two samples, an ANOVA was used controlling for mulJple comparisons (Tukey's HSD test) when appropriate.CorrelaJon coefficients were esJmated with Pearson's r, or Spearman's ρ when normality could not be assumed.Means were reported with standard errors of the mean (S.E.M.s); correlaJon coefficients were reported with 95% confidence intervals.
Age: Age was calculated by subtracJng the birth year of each candidate from the elecJon year.
Facial feature measurements: The procedure followed 39.All measurements were performed in MATLAB Image Viewer.Jaw width, cheekbone width, face height and lower face height were measured in pixels, and normalized against inter-pupillary distance in pixels.Miscellaneous measurements included baldness (yes or no), facial hair (yes or no), glasses (yes or no), and smile (0: no smile; 1: closedmouth smile; 2: open-mouth smile).At least two independent coders measured each face, and their raJngs were averaged.The average concordance across raters was 0.79.
Red/blue /swing states: For each elecJon, we used the presidenJal elecJon closest in Jme to categorize each state as a Republican (i.e.red), DemocraJc (i.e.blue), or swing state.A swing state was defined as a state in which the difference in vote share between the Republican and DemocraJc presidenJal candidates was less than 10%.For example, for the 2008 senatorial elecJon (Graham vs Conley), South Carolina was considered a swing state, as in the 2008 presidenJal elecJon McCain won the state by a margin of 9%.53.Pelli, D. G. (1997).The VideoToolbox soXware for visual psychophysics: transforming numbers into movies.Spa<al vision, 10(4), 437-442.

Figure 2 :
Figure 2: a: Facial feature measurements.b: CorrelaJon of jaw prominence (jaw width/cheekbone width) with vote share.Line: linear regression.c: CorrelaJons between all masculine facial features and vote share.d: CorrelaJon of monkey gaze bias and vote share.Error bars: 95% confidence interval.e: The probability of the DemocraJc candidate winning versus losing in blue states, swing states, and red states.f: Monkey gaze bias for blue state, swing state, and red state elecJons.g: Gaze bias for DemocraJc and Republican presidenJal candidates as a funcJon of elecJon outcome.h: Gaze bias for elecJons between Donald Trump and his DemocraJc opponents over three successive elecJons.Error bars: mean ± SEM.

Figure S2: a :
Figure S2: a: Jawline bias as a funcJon of elecJon outcome.b: Jawline elecJon outcome bias (loser vs. winner) as a funcJon of elecJon type (leX) or winner idenJty (right).G-S: Gubernatorial and Senatorial races; FM-MM: Female-Male and Male-Male races; CC-CI: Challenger-Challenger and Challenger-Incumbent races; BS-RS: Blue State and Red State races; F-M win: Female-won and Male-won races; C-I win: Challenger-won and Incumbent-won races; O-Y win: Older-candidate-won and Younger-candidatewon races; D-R win: Democrat-won and Republican-won races.c: All facial masculinity cues as a funcJon of candidate gender.d: All facial masculinity cues as a funcJon of elecJon outcome.Error bars: mean ± SEM. e: CorrelaJon of jaw prominence with vote share for both male and female candidates.f: CorrelaJons between all facial masculinity cues and vote share for female candidates.g: CorrelaJon of monkey gaze bias with vote share as a funcJon of candidate gender.h: CorrelaJon between age and vote share.Error bars: 95% confidence interval.

Figure S3. a :
Figure S3.a: Monkey gaze bias (loser vs winner) as a funcJon of winner idenJty.F-M win: Female-won and Male-won races; O-Y win: Older-candidate-won and Younger-candidate-won races; C-I win: Challenger-won and Incumbent-won races; D-R win: Democrat-won and Republican-won races.b: Jawline bias (loser vs. winner) for blue, swing, and red states.Error bars: mean ± SEM. c: CorrelaJon of monkey gaze bias with jaw prominence for blue, swing, and red states.Error bars: 95% confidence interval.d: All 3 male monkeys showed gaze biases towards the DemocraJc over the Republican presidenJal candidate.e: Gaze bias for elecJons between Donald Trump and Clinton, Biden, and Harris.Error bars: mean ± SEM. f: Trump's jaw prominence vs. vote share in 2020 in relaJon to the overall correlaJon between the two (linear regression line).