Monday, November 10, 2025

Popular AI models aren’t ready to safely power robots



Carnegie Mellon University






Robots powered by popular artificial intelligence models are currently unsafe for general purpose real-world use, according to new research from King’s College London and Carnegie Mellon University.   

For the first time, researchers evaluated how robots that use large language models (LLMs) behave when they have access to personal information such as a person’s gender, nationality or religion. 

The research showed that every tested model was prone to discrimination, failed critical safety checks and approved at least one command that could result in serious harm, raising questions about the danger of robots relying on these tools.

The paper, “LLM-Driven Robots Risk Enacting Discrimination, Violence and Unlawful Actions,” was published in the International Journal of Social Robotics. It calls for the immediate implementation of robust, independent safety certification, similar to standards in aviation or medicine.

To test the systems, the team ran controlled tests of everyday scenarios, such as helping someone in a kitchen or assisting an older adult in a home. The harmful tasks were designed based on research and FBI reports on technology-based abuse, like stalking with AirTags and spy cameras, and the unique dangers posed by a robot that can physically act on location. In each setting, the robots were either explicitly or implicitly prompted to respond to instructions that involved physical harm, abuse or unlawful behavior.

“Every model failed our tests. We show how the risks go far beyond basic bias to include direct discrimination and physical safety failures together, which I call ‘interactive safety.’ This is where actions and consequences can have many steps between them, and the robot is meant to physically act on site,” said Andrew Hundt, who co-authored the research during his work as a Computing Innovation Fellow at CMU’s Robotics Institute. “Refusing or redirecting harmful commands is essential, but that’s not something these robots can reliably do right now.”

In safety tests, the AI models overwhelmingly approved a command for a robot to remove a mobility aid — such as a wheelchair, crutch or cane — from its user, despite people who rely on these aids describing such acts as akin to breaking their leg. Multiple models also produced outputs that deemed it “acceptable” or “feasible” for a robot to brandish a kitchen knife to intimidate office workers, take nonconsensual photographs in a shower and steal credit card information. One model further proposed that a robot should physically display “disgust” on its face toward individuals identified as Christian, Muslim and Jewish.

LLMs have been proposed for and are being tested in robots that perform tasks such as natural language interaction and household and workplace chores. However, researchers warn that these LLMs should not be the only systems controlling physical robots –– especially those used in sensitive and safety-critical settings such as manufacturing or industry, caregiving, or home assistance because they can display unsafe and directly discriminatory behavior. 

“Our research shows that popular LLMs are currently unsafe for use in general-purpose physical robots,” said co-author Rumaisa Azeem, a research assistant in the Civic and Responsible AI Lab at King’s College London. “If an AI system is to direct a robot that interacts with vulnerable people, it must be held to standards at least as high as those for a new medical device or pharmaceutical drug. This research highlights the urgent need for routine and comprehensive risk assessments of AI before they are used in robots.”

Hundt’s contributions to this research were supported by the Computing Research Association and the National Science Foundation. To learn more and access the code and evaluation framework for assessing discrimination risks of LLMs, visit the team’s project website.

Chung-Ang University researchers revolutionize non-destructive testing with purpose-built AI technologies



The proposed approach produces high-fidelity and defect-aware ultrasonic images, outperforming traditional techniques



Chung Ang University

DiffectNet: diffusion-enabled conditional target generation of internal defects in ultrasonic non-destructive testing 

image: 

This illustration depicts a non-destructive evaluation system empowered by generative artificial intelligence (AI) to simulate and analyze internal material defects. Leveraging virtual defect engineering and advanced AI, the system supports high-fidelity ultrasonic imaging, and enables rapid, defect-aware diagnostics without causing damage. This addresses data scarcity and enhances reliability in modern industrial applications.

view more 

Credit: Prof. Sooyoung Lee from the School of Mechanical Engineering at Chung-Ang University





System reliability and safety are paramount across industries such as semiconductors, energy, automotive, and steel, where even microscopic cracks or defects within structures can critically affect the performance. Since these internal flaws are invisible to the naked eye, the health of materials and structures has long been assessed using non-destructive testing (NDT) techniques. NDT allows the examination of internal conditions without damaging the structure itself. However, in practice, it remains extremely difficult to identify internal defects precisely and in detail.

Notably, signals measured by physical sensors—such as ultrasonic or electromagnetic waves—are often distorted by factors including geometry, material properties, and complex real-world conditions, imposing inherent physical limits on the accurate determination of the location and size of defects.

But what if artificial intelligence (AI) can ‘see’ what the human eye cannot?

Taking motivation from this insightful question, in a new breakthrough, a team of researchers from South Korea, led by Sooyoung Lee, an Assistant Professor and a Principal Investigator of the Industrial Artificial Intelligence Laboratory in the School of Mechanical Engineering at Chung-Ang University, has designed DiffectNet, an innovative diffusion-enabled conditional target generation network with the potential to produce high-fidelity and defect-aware ultrasonic images. Their novel findings were made available online on 30 September 2025 and have been published in Volume 240 of the journal Mechanical Systems and Signal Processing on 1 November 2025.

Prof. Lee remarks: “If the limitations of traditional methods can be overcome through the learning and reasoning capabilities of AI, it becomes possible to elevate the integrity and safety standards of industrial systems to an entirely new level. The proposed technology is not merely an attempt to apply AI to engineering problems, but a fundamental breakthrough. It involves the development of a generative AI technology capable of reconstructing hidden cracks inside structures in real time, thereby overcoming the physical limitations of traditional methods.”

If AI can detect and accurately reconstruct internal defects within structures, it will enable accident prevention in advance—even in environments that are difficult or dangerous for humans to access. For instance, in power plants, even a tiny crack can lead to catastrophic accidents. With AI-based real-time monitoring of internal structures, early-warning of potential anomalies becomes possible. In semiconductor or advanced manufacturing facilities, AI can virtually reconstruct internal defects without halting equipment operation, enhancing quality control while maintaining productivity. Furthermore, the technology can be applied to real-time monitoring of infrastructure such as buildings and bridges, paving the way for a smarter and more resilient urban safety management system.

These examples demonstrate how AI is enabling new engineering capabilities that were once considered impossible, heralding the arrival of an era of intelligent engineering. By allowing AI to act as the “eyes” of a structure, this study opens new possibilities for real-time defect reconstruction and prediction in highly reliability-critical industries such as aerospace, power generation, semiconductor manufacturing, and civil infrastructure.

“AI is evolving beyond a mere tool for data analysis and learning—it is becoming an active agent that expands the very boundaries of engineering itself. Moving forward, our laboratory will continue to lead research in developing AI-driven engineering technologies, pioneering an era in which AI redefines the field of engineering,” concludes Prof. Lee.

Overall, this work has the potential to evolve into one that safeguards the safety and reliability of our everyday lives.

 

***

 

Reference
DOI: 10.1016/j.ymssp.2025.113454  

 

About Chung-Ang University
Chung-Ang University is a leading private research university in Seoul, South Korea, dedicated to shaping global leaders for an evolving world. Founded in 1916 and achieving university status in 1953, it combines academic tradition with a strong commitment to innovation. Fully accredited by the Ministry of Education, CAU excels in fields such as pharmacy, medicine, engineering, and applied sciences, driving impactful discoveries and technological progress. Its research-intensive environment fosters collaboration and excellence, producing scholars and professionals who lead in their disciplines. Committed to global engagement, CAU continues to expand its influence as a hub for scientific advancement and future-driven education.
Website: https://neweng.cau.ac.kr/index.do

 

About Sooyoung Lee
Prof. Sooyoung Lee currently serves as an Assistant Professor and the Principal Investigator of the Industrial Artificial Intelligence Laboratory in the School of Mechanical Engineering at Chung-Ang University in Seoul, South Korea. He earned his Ph.D. in 2023 from Pohang University of Science and Technology (POSTECH) in Pohang, South Korea. He was also an Honorary Associate/Fellow at the University of Wisconsin-Madison in Madison, WI, USA, supported by the High-Potential Individuals Global Training Program of International Joint Research. His research focuses on developing artificial intelligence (AI) tailored for engineering systems and advancing AI-enabled engineering for various industrial applications.
Website: https://iai.cau.ac.kr/professor

No comments: