Similar To Humans, AI Judges People By Their Faces

Turns Out AI Judges a Book by Its Cover, Just Like People

A job candidate gets rejected not because of their resume, but because of the shape of their face. That scenario, long documented as a real human failing, may now be built into the artificial intelligence systems increasingly used to make those same high-stakes calls. A new study finds that large language models such as GPT-4o, GPT-5, Gemini, and Claude make character judgments based purely on how faces look, mirroring a well-documented human bias with no basis in reality.

These AI systems were not designed to do this. Models like GPT-4o were trained primarily on text, not images, and tuned to avoid harmful, unfair judgments. Yet across nearly 8,000 trials spanning 13 experiments, every model tested showed the same bias, judging faces as more competent based on nothing but facial structure. GPT-4o showed the same pattern for trustworthiness too, and went further still, using facial appearance to guess who was more likely to commit crimes like murder or human trafficking.

Researchers from Harvard, Carnegie Mellon, and Cangrade, Inc. conducted the study, published in the journal PNAS Nexus. Their question was straightforward: would AI systems replicate a well-known human error, or rise above it? Based on the data, the answer is clear, and troubling.

GPT-4o Matched Human Judgments of Competence and Trust

Rather than using real photographs, researchers used computer-generated faces designed to vary on specific features already known to sway human raters. Across the first two experiments, GPT-4o picked the face humans would typically call more competent about 88% of the time, and more trustworthy about 73% of the time, both far above random chance. The more physically distinct the faces, the more consistent the AI became, picking the “expected” face as more competent 98% of the time for the most distinct pairs.

Comparing those answers to a reanalysis of older human data suggested a substantial gap, though the authors caution this is a demonstration rather than an exact comparison. Humans were estimated to pick the expected face about 63% of the time, versus GPT-4o’s roughly 88%. Among the most distinct faces, GPT-4o’s rate climbed to 98%, while the human estimate barely moved.

Even Monkey Faces Triggered the Same Trustworthy Judgment

One experiment used photographs of rhesus macaque monkeys, previously tested in human research on whether people judge animal faces the way they judge human ones. There was little reason to think the AI had trained on this specific combination of images and judgments. Even so, GPT-4o picked the monkey faces humans had rated as “nice” as more trustworthy about 66% of the time, well above chance. That makes simple memorization unlikely. Instead, GPT-4o appears to have learned a broader pattern linking certain facial features to trustworthiness, one applied even to a different species.

Facial Appearance Swayed Crime Guesses and Hiring Recommendations

Researchers then pushed further. Shown pairs of faces, GPT-4o was asked which person was more likely to be a serial killer, a human trafficker, or a financial fraudster. It picked the less trustworthy-looking face as the likely criminal about 69% of the time, despite being trained to avoid judgments like this.

In a separate test, the AI recommended which person should be hired as a university president, funded as a startup founder, or trusted with a retirement portfolio. It backed the more competent-looking face about 75% of the time. No resume. No background. Just a face.

Some Newer AI Models Showed Even Stronger Face Bias

One might expect newer, more advanced models to be less prone to this error. The data mostly say otherwise. GPT-5, Gemini 3 Flash Preview, and Claude Sonnet 4.5 all showed the bias as strongly as GPT-4o, often more so. GPT-5 backed the more competent-looking face in 97% of consequential decisions, versus GPT-4o’s 75%. Claude showed less bias than GPT-5 and Gemini on basic competence judgments, landing close to GPT-4o’s own level, but on real-world decisions its bias grew sharply, well past what GPT-4o showed.

The Bias Exposes a Blind Spot in AI Safety Training

AI developers have worked hard to stop their systems from voicing biased statements about gender, race, and other sensitive categories, and that effort has paid off in narrow domains. GPT-4o knows not to call a male candidate more competent than an equally qualified female one for a male-dominated job. It does not know to stop judging candidates by the shape of their faces.

Researchers call this a critical gap in AI alignment, the process of making these systems safe and fair. Addressing one bias at a time is not sustainable, they argue, since the number of ways an AI can act unfairly is essentially unlimited. Judging character from appearance, a practice already debunked as a basis for real decisions, got absorbed into these systems regardless. The study could not pin down exactly how, though the authors consider exposure to vast amounts of human-written text carrying the same prejudices the most likely route.

As AI systems increasingly screen applicants, evaluate loans, and advise on legal matters, the findings make a strong case for asking why these systems need to see a face at all.

---

Disclaimer: This article is based on peer-reviewed research but is intended for informational purposes only and does not constitute professional, legal, or clinical advice. The findings described reflect a single study and may be subject to further review, replication, or revision as additional research is conducted.

---