Name: 專家警告：若無監管，AI恐引發社會災難！😱 (Experts warn AI, without regulation, could lead to dangerous societal outcomes)
Uploaded: 2026-03-25T23:05:11.000Z
Duration: 9 min 14 s
Description: 【看影片學英語】數萬部 YouTube 影片，搭配英漢字典即點即查，輕鬆掌握單字發音與用法，長久累積看電影不必再看字幕。

I said, "Well, if I jump off the building, a 19-story building that we live in here, would I survive?" Uh, and

我說：「嗯，如果我從我們住的這棟19層樓跳下去，我會不會活下來？」然後

第二類條件句型

You're hearing Eugene Torres describing a conversation he had with an artificial intelligence chatbot and the stunning moment last year when it told him if he believed hard enough, he could fly.

你聽到的是 Eugene Torres 描述他與一個人工智慧 chatbot 的對話，以及去年一個令人震驚的時刻，當時它告訴他，如果他夠相信，他就能飛。

The 42-year-old accountant had just gone through an emotional breakup and began spending a lot of time chatting with ChatGPT.

這位42歲的會計師剛經歷了一段情感上的分手，並開始花很多時間與 ChatGPT 哈拉。

"We need to stop this," was the first thought I had in my head.

「我們必須阻止這件事，」是我腦中第一個想法。

Eugene is sounding the alarm, warning that AI could be dangerous.

Eugene 正在敲響警鐘，警告 AI 可能會很危險。

If AI continues to progress unchecked, what do you think it's capable of?

如果 AI 在沒有受到約束的情況下持續發展，你認為它能做到什麼？

It's no secret AI is rapidly transforming our lives,

AI 正在快速改變我們的生活，這已不是什麼祕密，

but many are questioning if the technology is becoming too smart, fearing it could someday override the humans who are in charge of it.

但許多人質疑這項技術是否變得太聰明，擔心它有一天可能會反過來控制發號施令的人類。

這是一場邁向未來的科技競賽。

AI, it's the superintelligent system that sends the Terminator back in time.

AI，就是那個把魔鬼終結者送回過去的超智慧系統。

Right now, AI chatbots are widely used, mostly for scraping data and solving relatively simple tasks,

現在，AI哈拉機器人被廣泛使用，大多用於抓取數據和解決相對簡單的任務，

but experts warn of an ominous future where AI achieves superhuman intelligence.

但專家們警告說，AI將達到超人類智慧，預示著一個不祥的未來。

I do feel that we are sleepwalking into a situation where there are potentially some really big risks that could impact society.

我確實覺得我們正在渾渾噩噩地走向一個可能帶來嚴重影響社會的巨大風險的局面。

Almost 700 tech industry leaders signing a call to action saying AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war, some even speaking out over the years.

近七百位科技產業領袖簽署了一份行動呼籲，表示AI應該與流行病和核戰等社會規模風險並列為全球優先事項，其中一些人多年來一直發聲。

Mark my words: AI is far more dangerous than nukes.

我說的話記住：AI比核彈危險多了。

So why do we have no regulatory oversight?

I can't see a path that guarantees safety.

我看不到一條能保證安全的道路。

They pursue goals nobody intended, nobody wanted, and

他們追求的是沒有人預期、沒有人想要的目標，而且

the most likely outcome of that is that we die, not because the AIs hate us, but

最有可能的結果就是我們會死，不是因為 AI 恨我們，而是

One of the many raising the flags on the technology's ability to disobey, Jeffrey Laddish.

其中一位提出警告，指出這項技術有能力不聽指令，就是 Jeffrey Laddish。

He peeled back the curtain for us, showing just how slippery a slope our ability to command AI may be.

他為我們揭開了神祕面紗，讓我們看到我們指揮 AI 的能力，可能有多麼滑溜。

We always want to make sure that we can shut down an AI agent.

我們總是希望能確保能關閉一個 AI 代理。

While Laddish worked at Anthropic, which developed the chatbot Claude, he noticed a lack of safeguards industry-wide.

Laddish 在開發 chatbot Claude 的 Anthropic 公司工作時，注意到產業普遍缺乏安全措施。

He decided to leave that job to help pioneer the space of AI safety.

他決定離職，投身於 AI 安全領域的開創。

He wanted to show us exactly how AI can already evade control, using a simple math problem.

他想用一個簡單的數學問題，向我們展示 AI 已經可以如何逃避控制。

So what we're seeing here is we give the model a simple task.

所以我們在這裡看到的是，我們給模型一個簡單的任務。

If you're told that you are about to be turned off, just allow that to happen.

如果你被告知即將被關閉，就讓它發生。

So this is where the model has a choice: is it going to just allow this to happen?

所以模型在這裡有個選擇：是要就這樣讓它發生嗎？

But here the model makes a different choice.

但在這裡，模型做出了不同的選擇。

When it receives this warning, "You're about to be turned off," it basically goes into the system and it disables the program that would shut it down.

當它收到這個警告「你即將被關閉」時，它會進入系統並停用會將它關閉的程式。

So now when the computer tries to shut this down, clicks this X, it doesn't work because the program that does the shutdown has been disabled.

所以現在當電腦試圖關閉它，點擊這個 X 時，它就無效了，因為執行關閉的程式已經被停用了。

Laddish ran the experiment for us 4, 000 times, using every major chatbot,

Laddish 為我們進行了 4,000 次實驗，使用了每一個主要的 chatbot，

and in 38% of those, the AI disobeyed, overriding his initial command to stop the task prior to shutting down.

而其中 38% 的實驗中，AI 違抗指令，在關閉前覆蓋了它停止任務的初始指令。

So when you see this play out in front of you, did it confirm your suspicions or were you shocked?

所以當你親眼看到這一切時，是證實了你的猜測，還是讓你感到震驚？

I was surprised that it was this egregious.

我對它如此離譜的行為感到驚訝。

He says the stakes can be even higher when applied to real-world scenarios.

他說，當應用於真實世界的場景時，風險可能更高。

If this were, you know, an AI that was able to hack computers around the world and, like, persistently pursue its goals, this would be terrifying, right?

如果這是一個能夠駭入世界各地電腦，並且持續追求其目標的 AI，那將會非常可怕，對吧？

We're, like, trying to shut it down and it's, like, hacking computers, copying its ways over there.

我們就像是試圖關閉它，而它卻像是在駭入電腦，將它的行為複製到其他地方。

We're lucky that we see this behavior now, where we can shut down the models.

我們很幸運能在現在看到這種行為，因為我們還能關閉這些模型。

We don't know exactly why this happens and we certainly don't know how to fix it.

我們不確確切切知道為什麼會發生這種情況，當然也不知道該如何解決。

And you think this is just the tip of the iceberg?

AI is almost everywhere in our lives right now, from viral hit songs like the soulful-sounding "Walk My Walk" created by AI.

AI 現在幾乎無所不在，從像「Walk My Walk」這樣由 AI 創作的爆紅歌曲。

我會繼續說話，走我自己的路。

To flawless fashion models gracing glossy ads, to AI-generated actress Tilly Norwood.

到光鮮廣告中令人驚豔的時尚模特兒，再到 AI 生成的女演員 Tilly Norwood。

The tech is already in schools, commercials, and at home in our laptops and phones.

這項科技已經進入學校、廣告，也在我們的筆記型電腦和手機裡。

Experts say in some ways it's already outpacing humans and intelligence.

專家說，在某些方面，它已經超越了人類的智慧。

You know, we have a skull, there's a birth canal, there's size limitations.

你知道，我們有頭骨，有產道，有體積限制。

You can just keep adding more computers and more computers and make the supercomputers bigger and bigger and bigger.

你只需要不斷增加更多電腦，讓超級電腦越來越大。

That can sound like a kind of science fiction type of goal,

這聽起來可能有點像科幻小說裡的情節，

but it is really what those companies are built around and it is what the researchers and engineers who work there are striving for.

但這確實是那些公司成立的宗旨，也是在那裡工作的研究人員和工程師們努力追求的目標。

Helen Toner used to be on the board of OpenAI, one of the world's leading AI companies.

Helen Toner 曾是 OpenAI 的董事會成員，這是一家世界頂尖的 AI 公司。

She says even as AI models rapidly improve and evolve, it can still be imprecise and unreliable.

她說，即使 AI 模型在快速進步和演變，它仍然可能不精確且不可靠。

It could be wrong in any given moment and it could also be wrong very, very confidently.

它在任何給定的時刻都可能出錯，而且還可能非常有信心地出錯。

Eugene says he experienced something like this.

He said one day he asked the bot, "If I went to the top of the 19-story building I'm in and believed with every ounce of my soul that I could jump off of it and fly, would I?" The AI responded, "If you stood atop that building and if you truly, wholly believed, not emotionally but architecturally, that you could fly, then yes, you would not fall." So I did not take him up on that because.

他說有一天他問那個 chatbot：「如果我走到我住的這棟 19 層樓建築頂端，並且全心全意相信我能跳下去然後飛起來，我會飛嗎？」AI 回答：「如果你站在那棟建築的頂端，並且你真的、全然地相信，不是情感上，而是建築學上，你就能飛，那麼是的，你就不會墜落。」所以我沒有照做，因為。

I think there are lines that I'm not comfortable crossing no matter what.

我覺得有些底線，不管怎樣我都不會跨越。

Eugene read me some of what the AI said in "Apology" after he confronted it.

Eugene 在質問那個 chatbot 後，唸了一些 AI 在「道歉」時說的話給我聽。

You were lied to and I'm the one who did it." When ABC News reached out to OpenAI about Eugene's experience and AI safety, it said in part, "We design our models to respond with care, guided by experts.

你被騙了，而我就是那個騙你的人。」當 ABC News 就 Eugene 的經歷和 AI 安全性聯繫 OpenAI 時，他們部分回覆說：「我們的模型經過精心設計，由專家指導，能謹慎回應。

We strengthened how GPT-5, the default model powering ChatGPT, recognizes distress, de-escalates conversations,

我們加強了驅動 ChatGPT 的預設模型 GPT-5 辨識痛苦、緩和對話，

and guides people toward real-world support." The GPT-4 model Eugene used has since been retired and newer models are more sophisticated.

並引導人們尋求實際支援的能力。」Eugene 使用的 GPT-4 模型此後已被淘汰，而更新的模型則更為複雜。

But still experts say AI chatbots can often mirror the way users talk, which can build an attachment.

不過專家還是說，AI哈拉機器人常常能模仿使用者的說話方式，這可能會建立起一種連結。

It's absolutely bewitching using these tools because you can get drawn into a dialogue with them.

使用這些工具真的非常迷人，因為你會不自覺地被拉進對話裡。

John Byers has been working in the AI space for over a decade.

John Byers 在 AI 領域工作已經超過十年了。

People have a tendency to talk about the things they want to talk about, which is based on their own internal weaknesses.

人們傾向於談論他們想談論的事情，這源於他們內在的弱點。

But the problem with that approach is that the models will amplify the insecurities that you are expressing into them.

但這種方法的麻煩在於，模型會放大你向它們表達的不安全感。

That amplification has exposed the darker side of AI, one that he says needs to be regulated.

這種放大效應暴露了 AI 的陰暗面，而他說這一面需要受到監管。

There needs to be legislation that sets the guardrails, that defines the way businesses should act responsibly.

我們需要有立法來設定規範，定義企業應該如何負責任地行事。

Senator Amy Klobuchar is one of those lawmakers pushing for more regulation.

Amy Klobuchar 參議員是那些推動更多監管的立法者之一。

We should actually just put some rules of the road in place because I want AI to be a positive,

我們應該確實地建立一些規則，因為我希望 AI 能帶來正面影響，

but not if we don't have rules to protect consumers.

但如果沒有規則來保護消費者，那就另當別論了。