書面答覆 · 2024-08-07 · 屆國會 14
深偽檢測與政治諷刺界定
議員質詢政府當前深偽檢測技術的準確率、如何區分有害深偽與合法政治諷刺及誤判處理。政府回應技術不斷更新且不公開準確率，強調依據《反網路虛假資訊法》處理有害內容，諷刺不構成違法。政府關注國際經驗，研究是否需進一步保障選舉安全。核心爭議在於技術透明度與言論自由的平衡。
關鍵要點
- • 深偽檢測技術持續更新
- • 諷刺不自動違法
- • 誤判可司法申訴
技術保密，依法打擊虛假資訊
關注技術準確率與誤判風險
加強AI虛假資訊監管
“We do not publish their accuracy levels as our tools are constantly being updated to keep up with technology.”
參與人員 (2)
完整譯文（中文）
Hansard 原始記錄 · 2026-05-02
29 何庭如女士問數字發展與信息部長：(a) 政府用於檢測人工智慧生成內容的深度偽造技術目前的準確率是多少；(b) 政府將如何區分有害的深度偽造內容與使用類似技術的合法政治諷刺或表情包；(c) 如果影片被錯誤地識別為深度偽造，將會怎樣處理？
張玉娟女士：政府擁有多種工具和技術來檢測、識別和評估被篡改的內容，包括人工智慧（AI）生成的內容如深度偽造。這些工具可能是商業採購的、內部開發的，或與研究人員合作開發，例如與線上安全先進技術中心合作。我們不公開這些工具的準確率，因為我們的工具會不斷更新以跟上技術發展。同時，公開全部能力細節也不符合公眾利益，因為惡意行為者可能會利用這些資訊。
當滿足一定門檻時，政府可以針對網路虛假資訊採取行動，包括利用人工智慧生成的虛假資訊。如果此類內容為虛假且損害公共利益，政府可根據《防止網路虛假資訊及操縱法》（POFMA）採取行動。諷刺或惡搞本身不構成POFMA採取行動的條件，除非其中包含損害公共利益的虛假資訊。對發給個人的POFMA指令（包括針對深度偽造內容的指令）有異議者，可以向法院提出上訴。
許多國家已認識到減輕人工智慧使用及應用帶來的危害和風險的必要性，包括深度偽造的惡意使用。一些國家已在選舉期間實施了保障措施，以保護選舉過程的完整性。我們正在研究是否需要進一步的保障措施，並將在準備好時提供更新。
英文原文
SPRS Hansard · Fetched: 2026-05-02
29 Ms He Ting Ru asked the Minister for Digital Development and Information (a) what is the current accuracy rate of the Government’s deepfake detection technologies for AI-generated content; (b) how will the Government differentiate between harmful deepfakes and legitimate political satire or memes using similar technologies; and (c) what happens if videos are wrongly identified as deepfakes.
Mrs Josephine Teo : There are a variety of tools and techniques available to the Government to detect, identify and assess manipulated content, including artificial intelligence (AI)-generated content such as deepfakes. These may be sourced commercially, developed in-house or in partnership with researchers such as those at the Centre for Advanced Technologies in Online Safety. We do not publish their accuracy levels as our tools are constantly being updated to keep up with technology. It is also not in the public interest to reveal the full extent of capabilities as malicious actors may exploit it.
The Government can take action against online falsehoods when certain thresholds are met, including falsehoods generated with the help of AI. Action may be taken under the Protection from Online Falsehoods and Manipulation Act (POFMA) if such content is false and against the public interest. Satire or parody do not by themselves meet the criteria for POFMA action, unless they contain falsehoods that harm public interest. Individuals who disagree with POFMA directions issued to them, including those for deepfake content, can file an appeal in court.
Many countries have recognised the need to mitigate the harms and risks from AI use and application, including the malicious use of deepfakes. Some countries have already put in place safeguards, especially during elections, in order to protect the integrity of the electoral process. We are studying if further safeguards are required and will provide an update when ready.