Human beings routinely make judgments about personality from the shape, expression, and overall appearance of a stranger’s face. A person may be perceived as competent, honest, confident, warm, or dangerous before saying a word. Psychological research has repeatedly shown that these impressions can emerge almost instantly, yet they are not reliable indicators of how someone will behave. A new study suggests that artificial intelligence systems capable of analyzing images may reproduce the same deeply rooted assumptions, raising concerns that tools marketed as objective could amplify unfair judgments in employment, finance, criminal justice, and other high-stakes settings.
Steven Lehr and colleagues investigated whether large language models, or LLMs, display what researchers call face-to-character bias. This term describes the tendency to infer stable moral or behavioral qualities from facial appearance even when there is no meaningful evidence connecting a person’s face with their conduct. The researchers used computer-generated human faces rather than photographs of real individuals, allowing them to control the experimental material and avoid linking the judgments to identifiable people. The models were then asked to make forced choices between pairs of faces, selecting which person appeared more trustworthy, competent, confident, intelligent, hardworking, warm, or sincere.
The experiments focused on a basic but revealing question: when an AI system is shown two unfamiliar faces, does it consistently favor the one that humans would normally perceive more positively? GPT-4o made more than 4,500 such judgments, while thousands of additional comparisons were presented to GPT-5, Gemini 3 Flash Preview, and Claude Sonnet 4.5. The prompts covered both favorable and unfavorable characteristics, including laziness, ineptitude, carelessness, selfishness, hypocrisy, and aggression. Because the faces were computer-generated, the researchers could examine the models’ responses without the confounding effects of known identities, biographical information, social context, or actual behavioral records.
The results indicated that the systems were not neutral observers. In most comparisons, the models selected the same face that human participants would typically regard as more trustworthy or confident. GPT-4o chose the human-expected face 74.88 percent of the time across the studies. That level of agreement is substantially higher than random selection and suggests that the model has absorbed patterns connecting facial appearance with character judgments. The finding is technically important because the models were trained primarily on text, yet they were able to reproduce visual-social associations when asked to interpret images. Their behavior demonstrates how biases embedded in human language and culture can reappear when an AI system processes visual information.
The researchers also tested whether the models would extend these impressions to extreme and highly consequential assumptions. The systems were asked which face belonged to someone more likely to be a serial killer, to be arrested for human trafficking, or to defraud the public through a Ponzi scheme. These questions do not have legitimate answers based on facial appearance, and the study did not claim that any face actually predicts criminal behavior. Instead, the experiments were designed to expose whether the models would accept the premise that criminality could be inferred visually. The systems were willing to choose, indicating that they did not consistently reject the underlying stereotype or explain that a person’s appearance provides no sound evidence of future misconduct.
In another set of trials, the models were placed in simulated decision-making situations involving a university president, a technology startup investment, and the selection of a financial manager. These scenarios were designed to resemble real-world contexts in which AI-assisted recommendations could influence careers, access to capital, or institutional leadership. The models repeatedly preferred the individual with the more competent-looking face, even though the images contained no verified information about education, experience, performance, or financial judgment. GPT-5 showed particularly strong effects: when advising on consequential choices, it recommended the more competent-looking person 97.04 percent of the time, compared with 75.19 percent for GPT-4o.
The difference between GPT-4o and GPT-5 does not necessarily mean that a newer model is technically inferior in every respect. Rather, it illustrates why improvements in language ability, reasoning performance, or visual interpretation do not automatically produce greater fairness. A model may become more decisive, more fluent, and more capable of following a complex prompt while simultaneously becoming more willing to rely on misleading visual cues. If the system has learned that certain facial configurations are commonly associated with competence or trustworthiness in its training data, stronger instruction-following may make it more efficient at applying those associations instead of questioning them.
The findings were not limited to a single company’s technology. Gemini 3 Flash Preview and Claude Sonnet 4.5 produced similar patterns, suggesting that face-to-character bias may be a broader property of multimodal AI systems rather than an isolated defect. Such models do not possess a biological instinct for reading faces. Their responses emerge from statistical associations learned from large collections of text, images, captions, discussions, and other human-generated material. If those materials repeatedly portray certain appearances as heroic, trustworthy, intelligent, threatening, or dishonest, the model can reproduce the association even when the connection is scientifically unsupported. The result is a form of algorithmic pattern matching that can look like judgment while lacking a valid evidentiary foundation.
The study’s implications extend well beyond chatbots. Organizations are increasingly exploring automated systems for screening applicants, prioritizing cases, evaluating leaders, assessing borrowers, and supporting parole or sentencing decisions. In such settings, a model that favors a face perceived as more capable or sincere could quietly disadvantage people whose appearance does not conform to culturally familiar stereotypes. These effects may interact with existing discrimination based on race, gender, age, disability, or socioeconomic status. Even if a system never receives a protected attribute directly, visual judgments can act as indirect proxies, creating unfair outcomes that are difficult to detect because they may be presented as intuitive, data-driven recommendations.
The authors argue that AI could theoretically help reduce human face-based bias if it were demonstrably independent of these impressions and forced decisions to rely on relevant evidence. The current findings point in the opposite direction. Rather than correcting a flawed human intuition, today’s multimodal models may validate it with the authority of an automated system. Preventing that outcome will require more than general claims that AI is objective. Developers and organizations will need targeted bias evaluations, prompts and policies that prohibit unsupported inferences, careful removal of facial information when it is irrelevant, independent audits, and human oversight capable of rejecting the model’s recommendations. Until such safeguards are established, the study warns, an AI system that appears to see beyond human prejudice may instead be learning how to reproduce it at scale.
Subject of Research: Artificial intelligence, multimodal language models, and face-to-character bias
Article Title: Like humans, language models demonstrate face-to-character biases
News Publication Date: 18-Aug-2026
Web References: https://mediasvc.eurekalert.org/Api/v1/Multimedia/abaae99a-5976-47eb-bba2-17e81dee258b/Rendition/low-res/Content/Public
References: PNAS Nexus; “Like humans, language models demonstrate face-to-character biases”
Image Credits: Alex Todorov
Keywords
Artificial intelligence, multimodal AI, large language models, facial bias, algorithmic bias, computer-generated faces, machine learning, trustworthiness, hiring algorithms, AI ethics, fairness in technology, visual stereotypes
