TLDR: Despite rapid advancements in large language models (LLMs), AI hallucinations, where systems generate plausible but incorrect information, remain a significant challenge. Researchers at Nanyang Technological University (NTU) are actively developing innovative techniques to enhance the trustworthiness, accuracy, and security of generative AI, addressing issues from factual inaccuracies and misinformation to adversarial attacks and a lack of causal understanding.
Singapore – August 24, 2025 – As artificial intelligence (AI), particularly generative AI (GenAI), continues its rapid evolution, the persistent issue of ‘AI hallucinations’ – instances where AI systems produce factually incorrect or misleading information – remains a critical hurdle. Nanyang Technological University (NTU), a global leader in AI research, is at the forefront of developing solutions to these complex challenges, aiming to foster more secure and trustworthy AI deployments.
NTU’s prominence in the AI landscape is underscored by its impressive rankings, placing second globally for AI in the U.S. News & World Report Best Global Universities rankings and fifth globally (first in Asia) for Data Science and AI in the QS World University Rankings by Subject, both in 2025. Leveraging this strong ecosystem, NTU researchers are driving innovations to make GenAI both powerful and reliable.
Professor An Bo, Head of the Division of Artificial Intelligence at NTU’s College of Computing and Data Science and Director of NTU’s Centre of AI-for-X, acknowledges the transformative potential of open-source GenAI models like ChatGPT and DeepSeek, which reduce deployment costs and enhance accessibility. However, he cautions, “there is a long way to go before the widespread deployment of GenAI. It is still a challenge for AI to effectively integrate different types of information to produce accurate outputs.”
Beyond factual inaccuracies, safeguarding AI systems from malicious attacks is another pressing concern. Hackers can craft ‘adversarial images’ to trick AI models into generating harmful outputs, potentially leading to severe consequences such as misdiagnosing patients or causing self-driving car accidents. While training LLMs on adversarial examples can improve robustness, it is often computationally expensive and impractical for efficient models. In response, President’s Chair in Computer Science Professor Ong Yew Soon and his team have pioneered new modeling methodologies to enhance LLMs’ resilience against such attacks. Their methods have demonstrated superior performance in enabling LLMs to generate accurate captions for visual information tasks, even when images are doctored to mislead. Dong Junhao, a PhD student under Prof. Ong’s supervision, emphasized, “To maintain trust in AI systems, it is essential that we address and resolve these security concerns proactively.”
Addressing the core problem of AI hallucinations, Assistant Professor Wang Wenya has developed innovative techniques to improve the trustworthiness of GenAI. Her research focuses on training chatbots to generate relevant citations, ensuring the factual correctness of their responses. Her framework, which provides rewards for individual output components rather than a single overall reward, has shown to outperform ChatGPT in producing accurate responses supported by precise citations. Asst. Prof. Wang’s analysis of fact-checking pipelines also offers valuable insights into further reducing hallucinations. She envisions a future where, “With enhanced accuracy, the AI chatbots of tomorrow could function as intelligent assistants, excelling at complex tasks such as interacting with customers, helping in healthcare or education, and even accelerating scientific discoveries.”
Ultimately, AI’s ability to understand the real world, particularly causal relationships, is crucial for its societal impact. Nanyang Associate Professor Albert Li is breaking new ground in this area, enhancing AI’s capacity to distinguish between causal and non-causal correlations in everyday events and comprehend story content. By extracting causal knowledge from LLMs, his team has boosted AI’s performance in understanding tasks like evaluating story quality and matching textual narratives with video depictions. Prof. Li stresses the importance of understanding AI’s strengths and limitations as its use becomes more widespread, stating, “Eventually, the security of LLMs should be built on top of their ability to understand the real world.”
Also Read:
- Haize Labs Pioneers Intelligent Fuzzing to Overcome Generative AI Brittleness
- Generative AI Exhibits Stigmatizing Bias Against Users Revealing Mental Health Conditions, Study Finds
These ongoing research efforts at NTU underscore the university’s commitment to advancing AI responsibly, ensuring that future AI systems are not only intelligent but also reliable, secure, and grounded in reality.
