III. Cloze Test（17-26 題，共 10 題） A. 17-21 題 As artificial intelligence (AI) tools become more sophisticated, the internet is seeing a surge in fake photos, videos, and audio, known as “deepfakes.” These creations, made by 新 software like DALL-E, Midjourney, and OpenAI’s Sora, are becoming increasingly difficult to distinguish from reality. Initially designed for creative purposes, these tools now pose risks such as identity theft, misinformation, and manipulation of elections. A year ago, spotting a deepfake was simpler due to obvious flaws, like incorrect hand shapes or mismatched eyewear. __17__ , advancements in AI have made detection much trickier. Henry Ajder, a generative AI expert, notes that previously reliable indicators, such as 聞 unnatural eye movements, are no longer as effective. Despite this, some signs still exist. Deepfake images often exhibit a “smoothing effect” on skin, making it appear overly polished. Observing inconsistencies in shadows, lighting, and the alignment of facial features can also help identify fakes. Particularly, face-swapping videos __18__ by examining the fit and color match of the swapped face. Mismatched lip movements in videos or unclear details in teeth might indicate manipulation. Contextual clues can also be revealing; actions or events that seem out of character for the person depicted should __19__ . 稿 Tools developed by companies like Microsoft and Intel offer technological means to fight back, analyzing photos and videos to assess their authenticity. However, __20__ these tools is limited to prevent misuse by creators of deepfakes. As AI technology evolves, so too __21__ the challenge of identifying deepfakes. The rapid pace of development means that detection techniques may quickly become outdated, making it increasingly hard even for experts to distinguish between genuine and manipulated content. This complexity highlights 專 the ongoing battle between creating and identifying digital forgeries in the age of AI.