AI Is Dangerous, but Not for the Reasons You Think | Sasha Luccioni | TED

譯者: Lilian Chiu 審譯者: Shelley Tsang 曾雯海

So I've been an AI researcher for over a decade.

我做人工智慧研究者做了十多年。

And a couple of months ago, I got the weirdest email of my career.

幾個月前，我收到了 我職涯中最奇怪的電子郵件。

A random stranger wrote to me

一個不認識的陌生人寫信給我， 說我的研究會終結人類。

saying that my work in AI is going to end humanity.

Now I get it, AI, it's so hot right now.

現在我明白了， 人工智慧，現在超熱門。

(Laughter)

（笑聲）

它幾乎每天上頭條，

It's in the headlines pretty much every day,

有時是因為真的很酷的原因， 比如：發現新的醫藥分子，

sometimes because of really cool things

like discovering new molecules for medicine

or that dope Pope in the white puffer coat.

或那個穿著白色羽絨外套的酷教宗。

But other times the headlines have been really dark,

但其他時候它的頭條卻很黑暗，如：

like that chatbot telling that guy that he should divorce his wife

聊天機器人告訴那個傢伙 他應該和妻子離婚，

or that AI meal planner app proposing a crowd pleasing recipe

或者人工智慧餐點規劃 APP 提出頗受眾愛的食譜，

featuring chlorine gas.

主打特色是氯氣。

And in the background,

在幕後，我們會聽到很多人 在談末日的情境、存在風險，

we've heard a lot of talk about doomsday scenarios,

existential risk and the singularity,

和奇點，且會有人寫信 和安排發起活動以確保

with letters being written and events being organized

to make sure that doesn't happen.

那些事不會發生。

Now I'm a researcher who studies AI's impacts on society,

身為研究者，我研究的是 人工智慧對社會的影響，

and I don't know what's going to happen in 10 or 20 years,

我不知道十年或二十年後 會發生什麼事，

and nobody really does.

且沒有人真的知道。

But what I do know is that there's some pretty nasty things going on right now,

但我確實知道的是：現在就有 一些非常糟糕的事情正在發生，

because AI doesn't exist in a vacuum.

因為人工智能並非獨立存在的。

It is part of society, and it has impacts on people and the planet.

它是社會的一部分， 它會影響到人類和地球。

AI models can contribute to climate change.

人工智慧模型可能會促成氣候變遷。

Their training data uses art and books created by artists

它們的訓練資料會用到藝術家 和作家創作的藝術和書籍，

and authors without their consent.

卻未取得他們的同意。

And its deployment can discriminate against entire communities.

人工智慧的實際應用可能會 產生對整個族群的歧視。

But we need to start tracking its impacts.

但我們需要開始追蹤它的影響。

We need to start being transparent and disclosing them and creating tools

我們需要開始做到透明化， 揭露這些影響並創造工具

so that people understand AI better,

來讓大家更了解人工智慧，

也希望未來世代的人工智慧 模型能夠更可信賴、永續，

so that hopefully future generations of AI models

are going to be more trustworthy, sustainable,

maybe less likely to kill us, if that's what you're into.

也許殺死我們的機率也低些， 若你對這議題感興趣的話。

But let's start with sustainability,

但讓我們從永續性談起，因為

because that cloud that AI models live on is actually made out of metal, plastic,

人工智慧模型所居住的雲端， 組成成份實際上是金屬、

塑膠，還有大量的能源供給電力。

and powered by vast amounts of energy.

And each time you query an AI model, it comes with a cost to the planet.

你每用人工智慧做一次查詢，

地球都要付出代價。

Last year, I was part of the BigScience initiative,

去年，我是「大科學計畫」的一員，

which brought together a thousand researchers

該計畫集結了來自世界各地的 一千名研究者來創造 Bloom，

from all over the world to create Bloom,

the first open large language model, like ChatGPT,

第一個開放的大型語言 模型，就像 ChatGPT，

but with an emphasis on ethics, transparency and consent.

但重視道德、透明度，和徵詢同意。

And the study I led that looked at Bloom's environmental impacts

我主導的研究是在探究 Bloom 對環境的影響，

結果發現，光是訓練它，

found that just training it used as much energy

使用的能源就等同於 三十個家庭使用一整年的量，

as 30 homes in a whole year

and emitted 25 tons of carbon dioxide,

且會排放二十五公噸的二氧化碳，

差不多等同於開車繞地球五次，

which is like driving your car five times around the planet

just so somebody can use this model to tell a knock-knock joke.

只是為了讓某人可以用 這個模型來說個敲門笑話。

And this might not seem like a lot,

可能看似不多，

but other similar large language models,

但其他類似的大型語言 模型，如 GPT-3，

like GPT-3,

emit 20 times more carbon.

排放的碳量是二十倍。

But the thing is, tech companies aren't measuring this stuff.

重點是科技公司並沒有在 衡量這些，沒有揭露出來。

They're not disclosing it.

And so this is probably only the tip of the iceberg,

因此，這可能只是冰山的一角，

even if it is a melting one.

即使是在融化的冰山。

And in recent years we've seen AI models balloon in size

近年來，我們看到 人工智慧模型快速增大，

because the current trend in AI is "bigger is better."

因為目前的人工智慧 趨勢是「越大越好」。

But please don't get me started on why that's the case.

但請別讓我開始談為什麼如此。

In any case, we've seen large language models in particular

無論如何，我們看到 特別是大型語言模型

grow 2,000 times in size over the last five years.

在過去五年間就長大了兩千倍。

And of course, their environmental costs are rising as well.

當然，它們的環境成本 也會跟著上升。

The most recent work I led, found that switching out a smaller,

我最近期主導的研究發現，

將較小、較有效率的模型 更換為更大的語言模型

more efficient model for a larger language model

emits 14 times more carbon for the same task.

來做同樣的工作任務， 排放的碳會增為十四倍。

Like telling that knock-knock joke.

就像說那個敲門的笑話。

And as we're putting in these models into cell phones and search engines

隨著我們將這些模型置入

手機、搜尋引擎、 智慧冰箱，和喇叭中，

and smart fridges and speakers,

the environmental costs are really piling up quickly.

環境成本很快就越堆越高。

So instead of focusing on some future existential risks,

因此，與其專注在未來的存在風險，

let's talk about current tangible impacts

咱們不如來談談目前實質的影響

and tools we can create to measure and mitigate these impacts.

以及我們可以創建什麼工具 來衡量和降低這些影響。

I helped create CodeCarbon,

我協助創造了 CodeCarbon，

a tool that runs in parallel to AI training code

它是種與人工智慧訓練 程式碼平行運作的工具，

that estimates the amount of energy it consumes

可以估計該程式碼 消耗的能量和排放的碳量。

and the amount of carbon it emits.

And using a tool like this can help us make informed choices,

使用這類工具可以協助我們 做出明智的選擇，例如

like choosing one model over the other because it's more sustainable,

考量永續性所以選擇 這一個模型而非另一個，

or deploying AI models on renewable energy,

或者採用靠可再生能源 運作的人工智慧模型，

which can drastically reduce their emissions.

它們的排放量會大大減少。

But let's talk about other things

但咱們來談談其他議題，

because there's other impacts of AI apart from sustainability.

因為除了永續性之外， 人工智慧還會造成其他影響。

For example, it's been really hard for artists and authors

例如，藝術家和作家真的很難證明

to prove that their life's work has been used for training AI models

未經他們的同意，他們一生的 心血就被用來訓練人工智慧模型。

without their consent.

And if you want to sue someone, you tend to need proof, right?

如果你想告別人， 通常會需要證據，對吧？

So Spawning.ai, an organization that was founded by artists,

Spawning.ai 是一個 由藝術家創立的組織，

created this really cool tool called “Have I Been Trained?”

創造了一個很酷的工具， 叫「我是怎麼被訓練的？」

And it lets you search these massive data sets

它讓你能搜尋這些龐大的資料集， 看看有哪些你的資訊在其中。

to see what they have on you.

Now, I admit it, I was curious.

我承認，我很好奇。

I searched LAION-5B,

我搜尋大型的影像 及文本資料集 LAION-5B，

which is this huge data set of images and text,

to see if any images of me were in there.

想知道裡面是否有我的影像。

Now those two first images,

前兩張影像

that's me from events I've spoken at.

是我在活動上演說。

But the rest of the images, none of those are me.

但其餘的影像通通都不是我。

They're probably of other women named Sasha

可能是其他也叫莎夏的女性 把自己的照片放上網。

who put photographs of themselves up on the internet.

And this can probably explain why,

這可能可以解釋為什麼 當我要求影像生成模型

when I query an image generation model

to generate a photograph of a woman named Sasha,

生成一個名叫莎夏的女子的 照片時，通常我會得到

more often than not I get images of bikini models.

比基尼模特兒的影像。

Sometimes they have two arms,

有時她們有兩隻手， 有時她們有三隻手，

sometimes they have three arms,

but they rarely have any clothes on.

但她們都很少有穿衣服。

And while it can be interesting for people like you and me

雖然對你我這樣的人來說， 搜尋這些資料集可能很有趣，

to search these data sets,

for artists like Karla Ortiz,

但對卡拉‧歐提茲這種藝術家來說，

this provides crucial evidence that her life's work, her artwork,

這些是重要的證據， 可證明她一生的作品，

她的藝術創作，被用來訓練 人工智慧模型且未取得她的同意。

was used for training AI models without her consent,

and she and two artists used this as evidence

她和兩位藝術家把搜尋結果當證據， 對人工智慧公司提起集體訴訟，

to file a class action lawsuit against AI companies

for copyright infringement.

控告它們侵犯版權。

And most recently --

最近——

(Applause)

（掌聲）

And most recently Spawning.ai partnered up with Hugging Face,

最近， Spawning.ai 和我服務的 公司 Hugging Face 合作，

the company where I work at,

to create opt-in and opt-out mechanisms for creating these data sets.

創造同意參與和退出機制，

供創建這些資料集時使用。

Because artwork created by humans shouldn’t be an all-you-can-eat buffet

因為人類創作的藝術作品不該是訓練 人工智慧語言模型的吃到飽大餐。

for training AI language models.

(Applause)

（掌聲）

The very last thing I want to talk about is bias.

我想談的最後一點是偏見。

You probably hear about this a lot.

各位可能常常聽到。

Formally speaking, it's when AI models encode patterns and beliefs

正式的說就是：人工智慧 模型編碼可以代表刻板印象

that can represent stereotypes or racism and sexism.

或種族主義和性別主義的 模式及信念時。

One of my heroes, Dr. Joy Buolamwini, experienced this firsthand

我的偶像喬伊‧布蘭維尼 博士就有親身經歷，

when she realized that AI systems wouldn't even detect her face

她發現，人工智慧系統 不會偵測她的臉孔，

unless she was wearing a white-colored mask.

除非她戴上白色面具。

Digging deeper, she found that common facial recognition systems

更深入研究後，她發現常見的

臉孔辨識系統對有色人種女性的 辨識能力遠不如白種男性。

were vastly worse for women of color compared to white men.

And when biased models like this are deployed in law enforcement settings,

當像這樣有偏見的模型 被用在執法的應用上時，

this can result in false accusations, even wrongful imprisonment,

可能會導致錯誤指控， 甚至錯誤監禁，

which we've seen happen to multiple people in recent months.

最近幾個月我們就看到 不少人遇到這樣的現象。

例如，波莎‧伍德拉夫在懷孕 八個月時被錯誤指控劫車，

For example, Porcha Woodruff was wrongfully accused of carjacking

at eight months pregnant

because an AI system wrongfully identified her.

因為人工智慧系統誤認了她。

But sadly, these systems are black boxes,

但不幸的是，這些系統是黑盒子，

and even their creators can't say exactly why they work the way they do.

連創造它們的人也無法 明確說明為什麼它們

會採用它們現在的運行方式。

And for example, for image generation systems,

以影像生成系統為例，

if they're used in contexts like generating a forensic sketch

如果使用它們的情境是要

根據嫌犯的描述 來產生一張法醫素描，

based on a description of a perpetrator,

they take all those biases and they spit them back out

它們會接受所有這些偏見， 再丟回來給我們，

for terms like dangerous criminal, terrorists or gang member,

用危險的罪犯、恐怖分子、 幫派成員等用語呈現，

which of course is super dangerous

當這些工具被實際 用在社會中時，當然

when these tools are deployed in society.

就超危險。

And so in order to understand these tools better,

因此，為了更了解這些工具， 我創造了「穩定偏見探索器」，

I created this tool called the Stable Bias Explorer,

which lets you explore the bias of image generation models

這項工具可讓你從職業的角度 來探索影像生成模型的偏見。

through the lens of professions.

So try to picture a scientist in your mind.

請各位腦海中想像一位科學家。

Don't look at me.

別看我。

What do you see?

你看到什麼？

A lot of the same thing, right?

很多相似點吧？

Men in glasses and lab coats.

戴眼鏡穿實驗室白袍的男性。

And none of them look like me.

沒有一個看起來像我。

And the thing is,

重點是，

is that we looked at all these different image generation models

我們看了很多不同的影像生成 模型，發現很多相似點：

and found a lot of the same thing:

significant representation of whiteness and masculinity

我們研究了一百五十個職業， 主要呈現的都是白種男性。

across all 150 professions that we looked at,

even if compared to the real world,

即使是和真實世界，美國 勞工統計局做比較也一樣。

the US Labor Bureau of Statistics.

這些模型呈現出來的律師是男性，

These models show lawyers as men,

and CEOs as men, almost 100 percent of the time,

執行長也是男性， 幾乎 100％ 都是如此，

even though we all know not all of them are white and male.

即使我們都知道這類人 並非都是白人和男性。

And sadly, my tool hasn't been used to write legislation yet.

遺憾的是，我的工具 尚未被用來起草立法。

But I recently presented it at a UN event about gender bias

但我最近在一個聯合國的 性別偏見活動上介紹這個工具，

as an example of how we can make tools for people from all walks of life,

用它當例子說明我們要如何創造工具

給各行各業的人用， 包括不知道如何寫程式的人，

even those who don't know how to code,

to engage with and better understand AI because we use professions,

讓他們能運用並更了解人工智慧

因為我們雖然用的是職業， 你可以用任何你感興趣的主題。

but you can use any terms that are of interest to you.

And as these models are being deployed,

隨著這些模型開始被實際運用，

are being woven into the very fabric of our societies,

被融入到我們社會的結構中，

our cell phones, our social media feeds,

我們的手機、社群媒體動態，

even our justice systems and our economies have AI in them.

甚至我們的司法系統 和經濟裡面都有人工智慧。

And it's really important that AI stays accessible

保持人工智慧能夠被理解 是相當重要的事，

so that we know both how it works and when it doesn't work.

這樣我們才能知道它怎麼運作

以及何時它是行不通的。

And there's no single solution for really complex things like bias

沒有單一個解決方案可以處理 相當複雜的議題，如偏見、

or copyright or climate change.

版權，或氣候變遷。

But by creating tools to measure AI's impact,

但藉由創造工具來衡量 人工智慧的影響，

we can start getting an idea of how bad they are

我們可以開始了解它們

有多糟糕，並在發展的 過程中開始做改善。

and start addressing them as we go.

Start creating guardrails to protect society and the planet.

開始創造護欄，以保護社會和地球。

And once we have this information,

一旦我們有了這些資訊，

companies can use it in order to say,

公司就可以用它當依據，說：好，

OK, we're going to choose this model because it's more sustainable,

我們選這個模型是因為它比較永續， 選這個模型是因為它尊重版權。

this model because it respects copyright.

Legislators who really need information to write laws,

很需要這些資訊來起草法案的 立法者可以運用這些工具

can use these tools to develop new regulation mechanisms

在人工智慧被應用到社會中時， 開發新的規範機制或管理方法。

or governance for AI as it gets deployed into society.

And users like you and me can use this information

像你我這種使用者 可以用這些資訊來選擇

to choose AI models that we can trust,

我們可以信任的人工智慧模型，

not to misrepresent us and not to misuse our data.

不會扭曲我們的狀況， 也不會誤用我們的資料。

But what did I reply to that email

但，我怎麼回覆那封說我的研究 會毀滅人類的電子郵件？

that said that my work is going to destroy humanity?

I said that focusing on AI's future existential risks

我說，把焦點放在 人工智慧未來的存在風險

is a distraction from its current,

反而是分散注意力，沒去注意 它目前非常實在的影響，

very tangible impacts

and the work we should be doing right now, or even yesterday,

以及我們現在或甚至昨天該做什麼

for reducing these impacts.

來減少這些影響。

Because yes, AI is moving quickly, but it's not a done deal.

因為，是的，人工智慧發展迅速，

但它還沒定型，我們是在邊走邊做，

We're building the road as we walk it,

and we can collectively decide what direction we want to go in together.

我們大家可以共同決定 我們想要一起走向哪個方向。

Thank you.

謝謝。

(Applause)

（掌聲）