Top AI models judge people by their faces, just like humans do. A new study finds the systems use facial features to guess who is more trustworthy, who is more likely to be a criminal, and who deserves to be hired. Researchers from Harvard, Carnegie Mellon and Cangrade tested GPT-4o, GPT-5, Gemini and Claude across nearly 8,000 trials. The results were published in the journal PNAS Nexus. Every model showed the same bias. In 540 trials, GPT-4o picked the less trustworthy-looking face as more likely to be a serial killer 68.7 percent of the time. It picked the more competent-looking face for hiring and funding calls 75.2 percent of the time. Newer models were worse. GPT-5 chose the expected face in 94.3 percent of competence judgments and 97 percent of real-world decisions. Gemini and Claude Sonnet 4.5 also showed heavy bias. The researchers say AI safety training blocks obvious biases like race and gender but missed this one. The finding raises hard questions about using AI in hiring and legal settings.