Towards Trustworthy Generative AI

Speaker: Karthik Nandakumar, Associate Professor in the Department of Computer Science and Engineering at Michigan State University (MSU); Affiliated Associate Professor in the Computer Vision department at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)

Date: 06 July 2026

YouTube link: https://youtu.be/GTi_srgkblA

While generative AI models have revolutionised many domains including computer vision and natural language processing, they have introduced new dangers to society.

In this talk, Karthik Nandakumar first reviewed the trust issues in generative AI and identified key unsolved challenges. Next, he focussed on three specific problems in generative computer vision and discuss potential solutions. Firstly, powerful capabilities of these models can be misused to cause privacy harm to individuals. To address this problem, he showed how the inherent vulnerabilities of these models can be leveraged to protect against misuse of facial data. Secondly, it is necessary to ensure that the content generated by these models is safe and aligned with human values. In this context, he demonstrated how models can robustly unlearn toxic information that has already been imbibed during pre-training. Finally, people live in a world where they can no longer believe what they see (hear or read), especially in digital form. Distinguishing between natural and synthetically generated or modified image/video content (e.g., deepfakes) is one of the grand cybersecurity challenges of the near future. To solve this problem, he considered both proactive approaches that watermark all content generated by honest actors and reactive techniques like generalisable multimedia forensic detectors that identify content generated/modified by malicious adversaries. He concluded with an immediate call for action to make generative AI models trustworthy, before they begin to cause irreparable real-world harm.