Beauty and the Bias: Exploring the Impact of Attractiveness on Multimodal Large Language Models

MLLMs exhibit widespread attractiveness bias: Seven different AI models showed attractiveness-influenced decisions in 86.2% of tested scenarios, meaning these AI systems make different judgments about people based on how attractive they appear—a problem that mirrors human bias but occurs in automated systems used for consequential decisions.

The "halo effect" extends to AI: In 94.8% of relevant scenarios, attractive individuals were attributed more positive traits like intelligence and confidence by the AI models, demonstrating that these systems don't just make different decisions about attractive people, but actively assign them better personal qualities based solely on appearance.

Attractiveness bias intersects with other forms of discrimination: The study found that gender, age, and race biases were significantly amplified by attractiveness, meaning that the impact of appearance-based bias isn't uniform across all groups—some populations may face compounded discrimination when their attractiveness is judged through cultural or racial stereotypes.

Beauty filters amplify the problem: The researchers tested the same 462 individuals both with and without beauty filters, revealing that AI systems' biased behavior can be manipulated by simple image editing tools, raising concerns about gaming these systems in high-stakes applications like hiring or lending.

This bias matters for real-world AI applications: Since these models are increasingly used in domains like hiring, credit decisions, and criminal justice—areas where human attractiveness bias already causes documented harm—the presence of this bias in AI systems could automate and scale existing societal inequities.