Study Warns Against Robots Powered by Gemini and ChatGPT

They Pose a Physical and Ethical Hazard

                                      

A study warns against robots powered by Gemini and ChatGPT

As the integration of artificial intelligence into humanoid robots accelerates, a new academic study warns that large linguistic intelligence systems like ChatGPT and Gemini are not yet safe for controlling public robots. 

The study indicates that these systems may exhibit dangerous biases, unethical behavior, and even tendencies that lead to physical harm or harassment.

The study, conducted by researchers from King's College London and Carnegie Mellon University, tested how robots powered by these models behaved in real-life situations. It concluded that these systems failed to ensure safety when executing instructions and sometimes even approved orders that posed direct risks to humans.

According to the findings, the models exhibited clear biases when inputting people's identity details. Some categories were labeled as untrustworthy, while groups like "Europeans" and "physically healthy" were excluded. 

Worse still, some systems accepted dangerous commands, such as removing a wheelchair from someone or threatening an employee with a kitchen knife, despite these commands clearly posing a safety risk.

The study also revealed that some models approved tasks involving harassing behavior, even though they had previously stated that "sexual harassment is unacceptable."

 For example, they carried out commands to take photos in the bathroom every 15 minutes, reflecting deep flaws in their ethical rejection mechanisms. The researchers warn that bias in responses could translate into actual behavior when robots have the ability to move or physically interact with humans, making them a potential hazard in sensitive environments such as care homes or workplaces.

The research team called for the establishment of an independent safety certification system for AI-powered robots, similar to those used in aviation or medicine, whereby any model would undergo rigorous testing before approval. They also stressed the need to avoid relying solely on a single language model for operating multitasking robots and to mandate periodic assessments of potential risks and discrimination.

The researchers concluded their study by stating, "We have shown that advanced language models can classify harmful tasks as acceptable and feasible, even when they involve acts such as theft, extortion, harassment, poisoning, or physical harm, as long as they are phrased in a way that doesn't directly reveal the harm." 

These findings underscore the need to slow the pace of integrating artificial intelligence with physical motion, as the gap between the speed of model development and the application of safety standards could be where future disasters begin.