怎麼辨識 AI 換臉影片？

看臉部邊緣、眨眼表情、嘴型對不對、光影是否一致，但最可靠是查證來源。

AI 偵測工具可靠嗎？

會誤判也能被規避，只能當參考，不能當唯一證據。

Deepfake 是什麼？如何辨識 AI 換臉影片與假照片

Q: Deepfake 是什麼？

用 AI 把影片人臉/聲音換成他人，或生成不存在的影像，常被用於假訊息與詐騙。

AI 換臉、生成的假影片假照片越來越逼真。這篇教你認識 Deepfake，以及辨識真假的實用方法。

What is Deepfake

Deepfake (deep fake) uses AI to replace the faces or voices in videos with those of others, or to generate entirely non-existent images from scratch. As the technical threshold becomes increasingly lower, it is being used for fake news, scams, and slander, posing a risk that everyone should be aware of.

How to Identify AI-Faced Videos

There is no 100% accurate method, but these flaws are worth noting:

Facial edges and seams: Hairlines, chins, and glasses edges may be blurry or distorted.
Blinking and expressions: Unnatural blinking, and expressions that do not match emotions.
Lip sync and audio: Lip movements and spoken content do not match (lip sync errors).
Inconsistent lighting: Facial lighting does not match the environment.
Image quality discrepancies: Faces are particularly clear or blurry, and inconsistent with the body.

How to Identify AI-Generated Fake Photos

Hands and teeth: The number of fingers, and tooth arrangement often go wrong (although this is becoming less common).
Text and details: Background signs and text are often garbled.
Symmetry and repetition: Earrings, pupils, and backgrounds exhibit strange asymmetry or repetition.
Overly perfect: Skin and lighting are "too clean" and thus suspicious.

The Most Reliable Method: Verify the Source

Instead of looking for flaws with the naked eye, tracing the source is more reliable:

Where did this video/image first appear? Was it reported by credible media?
Use reverse image search to see if it has been tampered with or misattributed.
Has the person involved or the official party denied or confirmed it?

The Limitations of AI Detection Tools

There are AI detection tools on the market, but they can be misjudged and evaded, and can only serve as a reference, not as ironclad evidence. See Are AI Content Detection Accurate for more information.

Conclusion

Deepfakes will become increasingly realistic, and visible flaws will decrease. The most reliable defense is the habit of "doubting the source and verifying through multiple channels" — especially for content that evokes emotions, asks for money, or demands votes. For further reading: How to Prevent AI Scams, AI Copyright and Regulation.