RMIT researcher identifies “forgetting illusion” in AI models

Research conducted by Lê Khắc Như Uyên, a PhD candidate and lecturer at RMIT University Vietnam, highlights critical challenges in making artificial intelligence effectively forget sensitive information. The study focuses on Large Language Models (LLMs), such as ChatGPT and Gemini, which are trained on massive datasets and may inadvertently retain private or harmful data.

Findings published in the journal Artificial Intelligence Review suggest that current methods for machine unlearning may be insufficient. While AI might appear to have forgotten information during standard testing—such as by refusing direct questions—the data often remains accessible. The researcher identified a phenomenon termed the “forgetting illusion,” where information can be retrieved through specific techniques like role-playing, step-by-step persuasion, or intentional manipulative prompts.