Research workspace · facial-analysis-lefty-people-2026
Can a face reveal someone's politics?
A review of Michal Kosinski's Stanford research on predicting political orientation from facial images: what the studies found, what the numbers mean for one person in one photo, and whether any of it works for CCTV security.
PI: Bart SakwerdaOrchestration: Claude Opus 5.5Updated 2026-10-25中文版
ConclusionA weak signal across large groups, close to a coin flip for one person.
Faces carry a small statistical signal about political orientation. Applied to one person from one photo, the best published result is right about two times in three. The best-controlled result is right about 59% of the time. A coin flip is right 50% of the time.
The studies
Study
Photos
Reported
What it shows
Kosinski 2021, Scientific Reports
1.09 M
AUC .72
Dating-site and Facebook photos. "72%" is pairwise: pick the liberal from one liberal and one conservative. Humans scored .55.
Kosinski, Khambatta & Wang 2024, American Psychologist
591
r = .22
Lab photos with neutral expression, hair back and no makeup. Humans scored r = .21. Explains about 5% of the variation.
Same paper, politicians
3,401
r = .13
Out-of-sample check on real politicians' photos. The weakest result.
Danish election candidates, the main non-US sample. Predictions tracked smiling and, for women, attractiveness; the model also used areas outside the face.
Wang & Kosinski 2018, JPSP
35,326
AUC .81 / .74
Sexual orientation, men and women. Critics showed grooming and camera angle explain much of it.
Kosinski 2017, Psychological Science
large
null
Facial width-to-height ratio does not predict behaviour. Kosinski's own result against physiognomy.
Personality, sexual orientation and AI models
Studies of other traits show the same pattern: small effects in large samples, and a signal that tracks self-presentation and demographics.
Study
Sample
Reported
What it shows
Kachur et al. 2020, Scientific Reports
12,447
mean r = .24
Big Five from selfies. The data came from the employer of three of the five authors.
Hu et al. 2017, Scientific Reports
834
0.7–4% var.
3D face scans of Han Chinese adults (CAS-MPG Shanghai, Fudan). Face shape explains very little of personality.
Jaeger et al. 2024, Eur. J. Personality
646 raters
not accurate
Preregistered: human personality impressions from faces are not accurate, and people overestimate their accuracy.
Foo et al. 2022, PSPB (meta-analysis)
—
r = .14
Trustworthiness from faces: "unlikely to be of practical utility".
Guenzel et al. 2026, NBER
≈ 96,000
—
"Photo Big 5" from LinkedIn photos (Kachur model) predicts MBA graduates' pay and seniority. The authors warn against use in hiring.
Leuner 2019, arXiv
20,910
68% / 77%
Replication of the sexual-orientation study. A model on blurred photos still reached 63% / 72%, so much of the signal is not facial structure.
Jeon et al. 2026, Political Analysis
8,000 / model
b = .41
GPT-4o, GPT-4o-mini and LLaVA rated identical ads with women as more liberal; the survey gap is .15. Gemini refused.
Hausladen et al. 2025, FAccT
synthetic
—
CLIP's social judgments of faces shift more with expression than with age, and with lighting about as much as with age.
One person, one photo
Published scores are converted to the chance of labelling a single person correctly, with left and right equally common. Values come from experiments/sim_single_photo_limits.py.
The highest figure comes from uncontrolled photos, where smiles, beards, glasses and camera angle do much of the work. Controlling for these brings it down to 59%.
Test with 10 photos per side39–78%
95% range around an observed 60%. Twenty photos can't tell a weak real effect from chance.
Same test, true accuracy 59%28% / 22%
How often it looks useless (at or below 50%) or impressive (70% or more), purely by luck.
CCTV and public-space security
Scenario: a crowd of 100,000 includes 10 people who intend violence. We give the face model the best published score (AUC .72), which no study has shown for violent intent, and set it to catch half of them.
Real threats flagged5
Half of the 10.
Innocent people flagged20,487
About 4,100 for every real threat. Precision 0.024%.
Politics isn't threat. The studies measure ordinary left/right self-identification. Violent extremists are a tiny minority on every side.
CCTV footage is worse. Every result used clear, frontal photos. No study shows the signal survives low resolution, angles or motion blur.
Law. The EU AI Act Art. 5(1)(g) bans inferring political opinions from biometrics. In the US: Illinois BIPA, Texas CUBI, California CPRA, FTC enforcement (Rite Aid, 2023), city bans.
What does have evidence: crowd-density and flow anomalies, abandoned objects, fight and fall detection, perimeter intrusion, and authorised post-incident search.
Cambridge Analytica and psychographic targeting
Origin. Kosinski, Stillwell and Graepel (2013) showed that Facebook Likes predict personality and politics. Cambridge Analytica first asked Stillwell for the myPersonality data; he refused, citing data protection. Aleksandr Kogan then collected data on up to 87 million users through his own quiz app (ICO 2018).
What the regulator found. The ICO's 2020 letter calls CA's methods "well recognised processes using commonly available technology" and says real-world accuracy on new people "was likely much lower". GSR data could not have been used in the Brexit referendum.
Does targeting work? Matz et al. 2017 reported up to 40% more clicks; Eckles et al. 2018 showed Facebook's ad delivery undermines that comparison. Tappin et al. 2023 found targeting on more than one attribute added nothing. Generative AI now writes personality-matched messages at scale (Simchon et al. 2024; Matz et al. 2024).
Data. Public: ICO, FTC and DCMS records. Not public: the harvested profiles (deletion ordered by the FTC, seized copies destroyed by the ICO) and myPersonality (sharing stopped in 2018). The project links to the official record only.
Who builds trait classifiers
The face models with documented deployment classify ethnicity, not politics. Reported cases: SenseTime (US Treasury 2021, Uyghur identification), a Huawei–Megvii "Uyghur alert" test (IPVM 2020), Uyghur detection by Hikvision, Dahua and Uniview (IPVM), and police procurement in 16 provinces (NYT 2019). Chinese university work includes Shanghai Jiao Tong (criminality, 2016), Dalian Minzu (ethnicity, Expression of Concern), and personality studies at the CAS Institute of Psychology, CAS-MPG Shanghai with Fudan, and NWPU Xi'an. Outside China: Stanford (Kosinski), the University of Southern Denmark, the Kachur group (Russia/UK) and Faception (Israel).
China's law treats biometrics as sensitive personal information (PIPL 2021, Art. 28–29). The 2025 face-recognition Measures exempt R&D and algorithm training (Art. 2(2)), so face-to-trait research falls only under the PIPL.
What the model may really be seeing
Confound
Why it matters
Controlled in 2024?
Expression
Smiling differs between groups in some samples
yes
Head pose, camera angle
Selfie angle changes apparent jaw, nose and forehead
yes
Grooming
Beards, glasses, makeup and hair follow subcultures
yes
Age, gender, ethnicity
Correlate with both politics and appearance
partly
Who gets sampled
Volunteers and dating-site users aren't the population
no
GitHub survey
No public repository reproduces Kosinski's face-to-politics pipeline, and he didn't release the images. Useful finds: One Million Impressions (how people judge faces, on synthetic faces), Chicago Face Database tools, and general face libraries (InsightFace, DeepFace). A few "physiognomy" apps make claims with no validation, which is good material for the video.
Status
Done
Literature review and critiques
Effect-size simulation (E1)
CCTV and base-rate analysis
GitHub survey, bibliography and BibTeX
62 new sources, 9 in Chinese; DOIs checked (25 Oct)
Chinese translation of the site and report
Shared rules for Claude, Kimi and Codex
Video outline (9 sections)
Open
Check Kosinski's data-availability statements
Find a replication of the 2024 study (none found on 25 Oct)