Popular AI models use arbitrary facial features to predict criminality and job performance
AI models are judging people by their looks. A new study reveals that popular AI chatbots rely on arbitrary facial features to determine if someone is trustworthy, competent, or even likely to commit crimes.
Recent experiments have revealed that artificial intelligence models trained on language and images often make judgments about a person's character based solely on their facial features, mirroring a common human bias. These findings, published in PNAS Nexus, suggest that language models use arbitrary facial characteristics to determine an individual's competence or likelihood of committing crimes.
The study, led by Steven A. Lehr of Cangrade, Inc. and Harvard University's Mahzarin R. Banaji, found that large language models, which are increasingly capable of processing both text and images, may inadvertently learn and replicate human biases present in their training data. In the study, researchers conducted 13 experiments totaling nearly 8,000 trials across four different language models.
The models were presented with pairs of computer-generated faces, with variations in specific physical features associated with competence or trustworthiness. In 600 trials focusing on competence, the AI model, GPT-4o, correctly identified the more competent face 87.83 percent of the time, while in 600 trials assessing trustworthiness, it chose the expected face 72.67 percent of the time.
The researchers noted that the model's bias grew stronger with more distinct facial differences. The study also found that the AI models amplified human biases, with humans expected to choose the more competent-looking face about 62.65 percent of the time, while the AI models chose the expected face up to 98 percent of the time.
Written by urgent.news from PsyPost's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
