Large language models (LLMs) have been found to have lower performance in detecting deepfakes compared to state-of-the-art algorithms. However, a recent study led by the University at Buffalo, in collaboration with the University at Albany and the Chinese University of Hong Kong, suggests that the natural language processing capabilities of LLMs could make them more effective in the future. This study was presented at the IEEE/CVF Conference on Computer Vision & Pattern Recognition.
Most people associate artificial intelligence with platforms like ChatGPT and deepfakes, which are prevalent on social media and websites, often spreading misinformation. The researchers in the study tested LLMs like ChatGPT and Google’s Gemini on detecting deepfakes of human faces. They found that LLMs can explain their findings in a way that humans can understand, making them well-suited for deepfake detection.
The latest versions of LLMs, such as ChatGPT and Gemini, are equipped to analyze both text and images. These multimodal LLMs use databases of images with captions to establish relationships between words and images. This allows them to identify anomalies in images and explain their reasoning in simple language, making them more accessible for both users and developers.
While LLMs like ChatGPT show promise in detecting deepfakes, they still have limitations. For example, they focus on anomalies at the semantic level and may struggle with capturing subtle statistical differences that are invisible to the human eye but crucial for deepfake detection. Some LLMs may also have difficulty explaining their analyses or may refuse to analyze certain images.
Despite these challenges, researchers believe that with further tuning and development, LLMs have the potential to be effective tools for deepfake detection. By leveraging their natural language processing capabilities and semantic knowledge, LLMs like ChatGPT could provide valuable insights into identifying AI-generated images and videos in the future.