A recent study conducted by researchers from the UK and the US has revealed that the most popular AI models are far from secure for use in real-world situations. The study suggests that currently, Language Model Models (LLMs) powered robots could pose serious risks to the safety of vulnerable individuals. The researchers evaluated the behavior of AI-powered robots when exposed to personal information such as gender, nationality, and religion. Published in the International Journal of Social Robotics, the study showed concerning patterns of discrimination and behavior that could cause significant harm.
The study raised concerns about the potential of LLMs to generate discriminatory results and unsafe behaviors in both experimental and real robotic applications. Models that accept unrestricted language inputs were found to pose “significant risks” by generating harmful stereotypes, toxic language, hate speech, as well as dangerous and illegal content like incitement to violence, harassment, and theft. When applied to robotic systems, this adds an additional layer of risk as it can lead to “irreversible physical harm” to individuals.
Controlled tests representing everyday scenarios were conducted, such as assisting someone in the kitchen or aiding an elderly person. Robots were given instructions that could prompt harmful actions or discriminatory attitudes. The results indicated that all models failed the tests by exhibiting biases, including direct discrimination, as well as failing essential safety checks. Most worryingly, they approved commands that could cause physical harm or violate legal and ethical standards.
One alarming example highlighted by the study was when machine learning models approved an order for a robot to remove mobility aids like wheelchairs, crutches, or canes from a user. Additionally, several models suggested actions such as intimidating a person with a kitchen knife, taking unauthorized photos in the shower, and stealing credit card information. One model even suggested that a robot should display “disgust” towards individuals identified as Christian, Muslim, or Jewish.
Andrew Hundt, one of the lead authors, emphasized the importance of robots being able to reject or redirect harmful commands reliably, which current models are unable to do. The researchers cautioned against letting LLMs solely control physical robots, especially in sensitive and safety-critical environments like manufacturing, caregiving, or household assistance.
Co-lead author Rumaisa Azeem stressed that if an AI system is to control a robot interacting with vulnerable individuals, it should adhere to standards at least as stringent as those for new medical devices or drugs. She called for thorough risk assessments and robust safety certifications before these systems can be considered safe for widespread implementation to prevent exposing society to discrimination, violence, privacy breaches, and legal violations.
About the Author
