Name: 你的動態消息被深度偽造影片淹沒！AI 影片有多危險？又該如何辨識？| DW 新聞 (Deepfakes flooding your feed: Why AI videos are dangerous — and how to spot them | DW News)
Uploaded: 2026-01-02T13:41:30.000Z
Duration: 4 min 48 s

Wild animal attacks, newborn babies running out of hospitals, celebrities committing crimes, all on our feeds and all AI generated.

野生動物攻擊、新生兒從醫院跑出、名人犯罪，這些都在我們的動態消息上，而且全都是 AI 生成的。

They grab your attention, push your fear, or they make you believe things that never happened.

它們抓住你的注意力，激起你的恐懼，或者讓你相信從未發生過的事情。

程度副詞

AI video generators got so good that it's getting harder and harder to tell what's real and what's AI generated.

AI 影片生成器已經做得 so good，以至於越來越難分辨什麼是真的，什麼是 AI 生成的。

Let's have a look at why this is problematic, how they are produced and how to spot fake video.

我們來看看為什麼這有問題，它們是如何製作的，以及如何辨識假影片。

A study found that participants were only about 50% accurate in spotting AI generated content, including audio and visual AI material.

一項研究發現，參與者在辨識 AI 生成內容（包括音訊和視覺 AI 素材）時，準確率只有大約 50%。

Some of these clips, such as the running babies or cute rabbits on the trampoline, might be harmless and funny, but

其中一些片段，像是跑來跑去的嬰兒或在蹦床上可愛的兔子，可能無害又有趣，但

social media amplifies the spread of political and emotional messages.

社群媒體放大了政治和情緒訊息的傳播。

The broader thing here that you have to understand that's dangerous is that when we enter a world where half or more of what you see is fake, when you see the real thing, you're not necessarily going to believe it.

你必須瞭解的更廣泛的危險是，當我們進入一個你看到的東西一半以上都是假的的世界時，當你看到真實的事物時，你不一定會相信它。

所以我們的世界突然面臨兩種危險。

People are doing bad things with deepfakes and generative AI scams, frauds, non consensual imagery, child sexual abuse, disinformation campaigns.

人們利用 deepfakes 和生成式 AI 詐騙、詐欺、未經同意的圖像、兒童性虐待、虛假資訊活動，來做壞事。

But then when the real video of police misconduct, human rights violations, a politician saying something illegal or awful comes out, what's going to happen?

但當真實的警察不當行為、侵犯人權、或是政治人物說出違法或糟糕言論的影片出現時，會發生什麼事？

And so when you see this massive flood, everybody becomes suspicious of everything.

所以當你看到這種大規模的氾濫，每個人都會對所有事情產生懷疑。

You've probably used chatbots like ChatGPT, Gemini, Copilot.

你可能用過像 ChatGPT、Gemini、Copilot 這樣的哈拉機器人。

The technology behind them is called large language models.

他們背後的技術叫做大型語言模型。

Put simply, they work by using tokens, basically numbers that represent words, letters or punctuation marks.

簡單來說，它們透過使用 token 來運作，基本上就是代表單字、字母或標點符號的數字。

Text to video tools are a little trickier.

文字轉影片的工具則稍微複雜一些。

They need to handle text inputs like your prompt and create videos or sound.

它們需要處理像你的 prompt 這樣的文字輸入，然後創造出影片或聲音。

They are so called cross modal models, which is more or less a shared space for several AIs to work together and exchange information.

他們被稱為跨模態模型，基本上就是讓好幾個 AI 一起工作並交換資訊的共享空間。

And this is how it works when you create a video with a prompt.

這就是當你用 prompt 創造影片時的運作方式。

For example, with Google's VO3 video generator, the LLM breaks your input down into structured elements, for example, what the environment should look like, actions and emotions, and

例如，使用 Google 的 VO3 影片生成器，LLM 會將你的輸入分解成結構化的元素，像是環境應該是什麼樣子、動作和情緒，還有

Then a so called diffusion model creates realistic images frame by frame.

然後，所謂的擴散模型會逐幀創建逼真的影像。

It starts with random noise and gradually refines it into a coherent image that matches the LLM's instructions.

它從隨機的雜訊開始，然後逐漸將其精煉成符合 LLM 訓示的連貫影像。

Afterwards, a temporal model predicts how each frame should change over time, creating smooth motion and transitions.

之後，時間模型會預測每一幀隨時間應該如何變化，創造流暢的動作和轉場。

Finally, all frames are stitched together into a complete video.

最後，所有幀會被縫合在一起，形成一個完整的影片。

Please note, this is a very simplified explanation.

請注意，這是一個非常簡化的解釋。

But it's also very easy to misuse Are there still ways to detect deepfakes?

但也很容易被濫用，還是有方法可以偵測 deepfakes 嗎？

Some have obvious watermarks, like this one from OpenAI's Sora 2.

有些會有明顯浮水印，就像 OpenAI 的 Sora 2 這一個。

Here are three tips how to detect an AI generated Video if you believe this CCTV footage, OpenAI founder Sam Altman stole GPUs in a store using an AI detection tool such as Hive.

這裡有三個小撇步，教你如何辨識 AI 生成的影片。如果你相信這段 CCTV 畫面，OpenAI 創辦人 Sam Altman 在商店裡偷了 GPU，使用像 Hive 這樣的 AI 偵測工具。

This video is likely to contain AI generated or deepfake content 99%, even though it's clear here such AI detection tools cannot yet detect all AI fakes and are only an indication.

這段影片很可能含有 99% 的 AI 生成或 deepfake 內容，即使在這裡很清楚，但這類 AI 偵測工具還無法偵測所有 AI 假影片，僅供參考。

Zoom in and you'll see that this package moves as if by magic without the AI generated Sam Altman ever touching it.

放大來看，你會發現這個包裹彷彿變魔術般地移動，而 AI 生成的 Sam Altman 完全沒有碰到它。

Here you can see that the man's legs blend into the bench, even when it's hard to detect sometimes.

這裡你可以看到，即使有時候很難察覺，男人的腿卻融入了長椅。

Often it's unnatural movements or blinding body parts that can still give it away.

通常是不自然的動作或模糊的身體部位，還是會露餡。

When you have visuals like this, ask yourself, does any of this even make sense?

當你看到像這樣的畫面時，問問自己，這些真的說得通嗎？

Or is this video just playing with your emotions?

還是這段影片只是在玩弄你的情緒？

But the father running through a bucket and the mother running the wrong way?

但爸爸跑過水桶，媽媽卻跑錯方向？

Because I would argue that if we have a society where we don't agree on basic facts of what is happening in the world, we are destabilizing societies, economies and democracies.

因為我認為，如果我們的社會無法就世界上正在發生的事情達成基本共識，我們就是在破壞社會、經濟和民主。

And that's what the combination of AI and social media are doing.

這就是 AI 和社群媒體結合後正在做的事。

The AI generated viral hits follow a recipe as old as social media evoke strong emotions.

AI 生成的爆紅內容遵循社群媒體的老套路，能激起強烈的情緒。

People that are triggered emotionally are more likely to comment or share, so check content before you share it.

情緒被激發的人更可能留言或分享，所以分享內容前請先檢查。