How to Govern AI — Even If It’s Hard to Predict | Helen Toner | TED

譯者: 麗玲 辛 審譯者: C Leung

When I talk to people about artificial intelligence,

當我與人們談論人工智慧時，

something I hear a lot from non-experts is “I don’t understand AI.”

一般人經常說：「我不懂 AI。」

But when I talk to experts, a funny thing happens.

但是當我和專家交談時， 有趣的事情發生了。

They say, “I don’t understand AI, and neither does anyone else.”

他們說：「我不懂 AI， 其他人也都不懂。」

This is a pretty strange state of affairs.

這個情況很奇怪。

Normally, the people building a new technology

通常，開發新科技的人

understand how it works inside and out.

會對它的運作瞭如指掌。

But for AI, a technology that's radically reshaping the world around us,

但對於 AI 這種從根本重塑 我們周圍世界的科技來說，

that's not so.

情況並非如此。

Experts do know plenty about how to build and run AI systems, of course.

當然，專家們確實非常了解 如何建置和運行 AI 系統。

But when it comes to how they work on the inside,

但當談到它的內部運作方式時，

there are serious limits to how much we know.

我們所知卻非常有限。

And this matters because without deeply understanding AI,

這很要緊，因為 如果不深入了解 AI，

it's really difficult for us to know what it will be able to do next,

我們真的很難知道它接下來能做什麼，

or even what it can do now.

甚至它現在的能力到哪裡。

And the fact that we have such a hard time understanding

事實上，我們很難理解 這項科技目前的發展，

what's going on with the technology and predicting where it will go next,

也很難預測它下一步的走向，

is one of the biggest hurdles we face in figuring out how to govern AI.

這正是我們在思考如何治理 AI 時面臨的最大障礙之一。

But AI is already all around us,

但 AI 已無處不在，

so we can't just sit around and wait for things to become clearer.

因此我們不能只是坐等 事態變得更加清晰。

We have to forge some kind of path forward anyway.

無論如何，我們必須開闢 某種前進的道路。

I've been working on these AI policy and governance issues

我已經研究 AI 政策 和治理問題大約八年了，

for about eight years,

First in San Francisco, now in Washington, DC.

首先在舊金山，現在在華盛頓特區。

Along the way, I've gotten an inside look

一路走來，我深入了解

at how governments are working to manage this technology.

政府如何努力管理這項科技。

And inside the industry, I've seen a thing or two as well.

產業內部方面，我也略知一二。

So I'm going to share a couple of ideas

在此，我將分享

關於治理 AI 的途徑的一些想法。

for what our path to governing AI could look like.

But first, let's talk about what actually makes AI so hard to understand

首先，我們來談談

為何 AI 如此難以理解和預測。

and predict.

One huge challenge in building artificial "intelligence"

建構 AI 的一大挑戰是，

is that no one can agree on what it actually means

智能的真正定義並未達成共識。

to be intelligent.

This is a strange place to be in when building a new tech.

就建立一個新科技而言， 這是個不尋常的起點。

When the Wright brothers started experimenting with planes,

當萊特兄弟開始試驗建造飛機時，

they didn't know how to build one,

他們不知道如何做，

but everyone knew what it meant to fly.

但每個人都知道飛行的定義。

With AI on the other hand,

反觀 AI，

different experts have completely different intuitions

各個專家對於智能的核心

about what lies at the heart of intelligence.

有著完全不同的直觀想法。

Is it problem solving?

是解決問題？

Is it learning and adaptation,

是學習與適應能力？

are emotions,

還是要有情緒或有實體參與？

or having a physical body somehow involved?

We genuinely don't know.

我們真的不知道。

But different answers lead to radically different expectations

但不同的答案引領 人們對科技的發展方向

about where the technology is going and how fast it'll get there.

以及實現速度的預期截然不同。

An example of how we're confused is how we used to talk

一個困惑點的例子是 我們過去總是談論

about narrow versus general AI.

狹義 AI 與通用 AI。

For a long time, we talked in terms of two buckets.

長久以來，我們都是 用這兩個分類來談。

A lot of people thought we should just be dividing between narrow AI,

很多人認為我們應該區分 狹義 AI 和通用 AI，

trained for one specific task,

狹義 AI 是針對一項特定任務

like recommending the next YouTube video,

而進行訓練的， 例如推薦下個 YouTube 影片。

versus artificial general intelligence, or AGI,

而通用 AI 或 AGI，

that could do everything a human could do.

則可以完成人類能做的所有事情。

We thought of this distinction, narrow versus general,

我們認為這種狹義與通用之間的區別

as a core divide between what we could build in practice

是我們在實際上能打造的科技

and what would actually be intelligent.

與智能的可能樣貌之間的核心鴻溝。

But then a year or two ago, along came ChatGPT.

但是一兩年前， ChatGPT 出現了。

If you think about it,

仔細想想，

you know, is it narrow AI, trained for one specific task?

它是狹義 AI， 訓練來完成某項特定任務嗎？

Or is it AGI and can do everything a human can do?

或者它是通用 AI ， 可以做人類能做的一切？

Clearly the answer is neither.

顯然答案都不是。

It's certainly general purpose.

它當然是通用的。

It can code, write poetry,

它可以編碼，寫詩，

analyze business problems, help you fix your car.

分析業務問題，幫助你修車。

But it's a far cry from being able to do everything

但這比起你我能把所有事情 做好的程度，還相去甚遠。

as well as you or I could do it.

So it turns out this idea of generality

事實證明，通用性的這個概念

doesn't actually seem to be the right dividing line

實際上似乎並不是智能與 非智能之間的正確分界線。

between intelligent and not.

And this kind of thing

而這種狀況

is a huge challenge for the whole field of AI right now.

對於現在整個 AI 領域來說 是個巨大的挑戰。

We don't have any agreement on what we're trying to build

我們對於正在試圖建造的科技

or on what the road map looks like from here.

或未來的發展路線沒有任何共識。

We don't even clearly understand the AI systems that we have today.

我們甚至也不了解 現今擁有的 AI 系統。

Why is that?

為什麼會這樣？

Researchers sometimes describe deep neural networks,

研究人員有時會將深度神經網路，

the main kind of AI being built today,

即當今主要建構的 AI 類型

as a black box.

比喻為黑盒子。

But what they mean by that is not that it's inherently mysterious

但他們的意思並不是說 它本質上是神秘的，

and we have no way of looking inside the box.

或是說我們無法看到盒子的內部。

The problem is that when we do look inside,

問題是，當我們仔細觀察內部時，

what we find are millions,

我們發現數以百萬計、

billions or even trillions of numbers

數十億甚至數萬億的數字

that get added and multiplied together in a particular way.

以特定的方式相加和相乘。

What makes it hard for experts to know what's going on

專家之所以很難知道怎麼回事，

is basically just, there are too many numbers,

基本上只是因為數字太多，

and we don't yet have good ways of teasing apart what they're all doing.

而我們還沒有好方法來套取資訊， 了解它們在做什麼。

There's a little bit more to it than that, but not a lot.

事實上，不只如此，但也差不多了。

So how do we govern this technology

那我們要如何管理

that we struggle to understand and predict?

這項我們難以理解和預測的科技呢？

I'm going to share two ideas.

我要分享兩個想法。

One for all of us and one for policymakers.

一個給我們所有人， 一個給政策制定者。

First, don't be intimidated.

首先，不要害怕。

Either by the technology itself

無論科技本身，

or by the people and companies building it.

或者建構這科技的人和公司。

On the technology,

在科技方面，

AI can be confusing, but it's not magical.

AI 可能令人困惑，但它並不神奇。

There are some parts of AI systems we do already understand well,

AI 系統的某些部分 我們已經非常了解，

and even the parts we don't understand won't be opaque forever.

甚至我們現在不了解的部分 也不會永遠難以理解。

An area of research known as “AI interpretability”

過去幾年，一個被稱 「 AI 可解釋性」的研究領域

has made quite a lot of progress in the last few years

在理解數十億個數字的意義方面 取得了很大進展。

in making sense of what all those billions of numbers are doing.

One team of researchers, for example,

例如，某個研究小組找到了一種方法

found a way to identify different parts of a neural network

來識別神經網路的不同部分，

that they could dial up or dial down

他們可以調高或調低參數，

to make the AI's answers happier or angrier,

讓 AI 的答案開心或憤怒一點、

more honest,

更誠實、

more Machiavellian, and so on.

更狡猾等等。

If we can push forward this kind of research further,

如果我們能進一步推進這種研究，

then five or 10 years from now,

那麼五年、十年後，

we might have a much clearer understanding of what's going on

我們可能會對所謂黑盒子 內部發生的事情

inside the so-called black box.

有更清晰的了解。

And when it comes to those building the technology,

至於那些建構科技的人，

科技專家有時會表現得好像

technologists sometimes act as though

if you're not elbows deep in the technical details,

如果你不深入了解科技細節，

then you're not entitled to an opinion on what we should do with it.

那麼你就無權發表意見， 談我們應該如何處理它。

Expertise has its place, of course,

當然，專業知識有其一席之地，

but history shows us how important it is

但歷史告訴我們，

that the people affected by a new technology

在形塑新科技的使用時，

get to play a role in shaping how we use it.

受新科技影響的人們 參與其中有多麼重要。

Like the factory workers in the 20th century who fought for factory safety,

就像 20 世紀為工廠安全 奮鬥的工廠工人，

or the disability advocates

或是確保全球資訊網無障礙的倡議者。

who made sure the world wide web was accessible.

You don't have to be a scientist or engineer to have a voice.

你不必成為科學家或工程師， 才能擁有發言權。

(Applause)

（掌聲）

Second, we need to focus on adaptability, not certainty.

第二，我們要注重適應性， 而不是確定性。

A lot of conversations about how to make policy for AI

許多關於如何制定 AI 政策的討論

get bogged down in fights between, on the one side,

都陷入爭論，

這一方的人說：

people saying, "We have to regulate AI really hard right now

「我們現在必須非常嚴格地 監管 AI，因為風險太大。」

because it's so risky."

And on the other side, people saying,

另一方的人則說：

「但監管會扼殺創新， 而且這些風險都是揑造出來的。」

“But regulation will kill innovation, and those risks are made up anyway.”

But the way I see it,

但在我看來，

it’s not just a choice between slamming on the brakes

我們不是只能在踩煞車

or hitting the gas.

還是踩油門之間選擇。

If you're driving down a road with unexpected twists and turns,

如果你在無法預知的曲折道路上行駛，

then two things that will help you a lot

那麼，有兩個東西會很有幫助，

are having a clear view out the windshield

那就是擋風玻璃外的清晰視野

and an excellent steering system.

和出色的操控系統。

就 AI 而言，這意味著

In AI, this means having a clear picture of where the technology is

要清楚了解科技的現狀和發展方向，

and where it's going,

and having plans in place for what to do in different scenarios.

並制定在不同狀況的行動計畫。

Concretely, this means things like investing in our ability to measure

具體來說，這意味著

要投資於我們衡量 AI 系統功能的能力。

what AI systems can do.

This sounds nerdy, but it really matters.

這聽起來很無趣乏味， 但這確實很重要。

Right now, if we want to figure out

現在，如果我們想弄清楚

whether an AI can do something concerning,

AI 是否有能力 做某些令人擔憂的事，

like hack critical infrastructure

例如入侵關鍵基礎設施

or persuade someone to change their political beliefs,

或說服人們改變他們的政治信仰，

our methods of measuring that are rudimentary.

我們的衡量方法還很初級。

We need better.

我們需要更好的方法。

We should also be requiring AI companies,

我們也應該要求 AI 公司，

especially the companies building the most advanced AI systems,

尤其是建立最先進的 AI 系統的公司，

to share information about what they're building,

分享有關他們正在建造的內容、

what their systems can do

他們的系統有那些能力，

and how they're managing risks.

以及他們如何管理風險等資訊。

And they should have to let in external AI auditors to scrutinize their work

他們應該讓外部 AI 稽核員 來審查他們的工作，

so that the companies aren't just grading their own homework.

這樣，這些公司就不會自吹自擂。

(Applause)

（掌聲）

A final example of what this can look like

最後一個可能的例子是

is setting up incident reporting mechanisms,

建立事件報告機制，

so that when things do go wrong in the real world,

這樣，當現實世界中真的出現問題時，

we have a way to collect data on what happened

我們就有辦法收集 事件的相關數據

and how we can fix it next time.

以及下次如何修復的數據。

Just like the data we collect on plane crashes and cyber attacks.

就像我們收集的飛機失事 和網路攻擊資料一樣。

None of these ideas are mine,

這些都不是我的想法，

and some of them are already starting to be implemented in places like Brussels,

其中一些想法已經開始在布魯塞爾、

London, even Washington.

倫敦、甚至華盛頓等地方實施。

But the reason I'm highlighting these ideas,

但我之所以強調這些想法，

measurement, disclosure, incident reporting,

衡量系統、充分披露、事件報告，

is that they help us navigate progress in AI

是因為這會給我們擋風玻璃外 更清楚的視野，

by giving us a clearer view out the windshield.

從而幫助我們確定 AI 的進步方向。

If AI is progressing fast in dangerous directions,

如果 AI 正朝著 危險的方向快速發展，

these policies will help us see that.

這些政策將幫助我們看到這一點。

And if everything is going smoothly, they'll show us that too,

如果一切進展順利， 它們也會讓我們知道，

and we can respond accordingly.

我們就可以做出相應的反應。

What I want to leave you with

我想告訴你們的是，

is that it's both true that there's a ton of uncertainty

AI 領域確實存在著

大量的不確定性和分歧。

and disagreement in the field of AI.

And that companies are already building and deploying AI

而且各公司已經 在各地建置、部署 AI，

all over the place anyway in ways that affect all of us.

無論如何都會影響我們所有人。

Left to their own devices,

如果放任他們自行決定，

it looks like AI companies might go in a similar direction

AI 公司可能會走向

to social media companies,

與社群媒體公司類似的方向，

spending most of their resources on building web apps

將大部分資源花在建立網路應用程式

and for users' attention.

和吸引用戶注意力上。

And by default, it looks like the enormous power of more advanced AI systems

在目前預設情況下， 更先進的 AI 系統的巨大力量

might stay concentrated in the hands of a small number of companies,

可能會集中在少數公司，

or even a small number of individuals.

甚至在少數人的手中。

But AI's potential goes so far beyond that.

但 AI 的潛力遠不止於此。

AI already lets us leap over language barriers

AI 已經讓我們跨越語言障礙，

and predict protein structures.

預測蛋白質結構。

More advanced systems could unlock clean, limitless fusion energy

更先進的系統可以釋放 乾淨、無限的聚變能源，

or revolutionize how we grow food

徹底改變我們種植糧食的方式

or 1,000 other things.

或更多無數事情。

And we each have a voice in what happens.

我們每個人對事情的發展都有發言權。

We're not just data sources,

我們不僅是資料來源，

we are users,

我們是用戶，

we're workers,

我們是員工，

we're citizens.

我們是公民。

So as tempting as it might be,

因此，儘管我們很想這麼做，

we can't wait for clarity or expert consensus

但我們不能坐等事態明朗， 或者專家共識，

to figure out what we want to happen with AI.

才弄清楚我們要 AI 實現什麼目標。

AI is already happening to us.

AI 已在我們週遭。

What we can do is put policies in place

我們能做的是讓政策到位，

to give us as clear a picture as we can get

盡可能清楚地了解科技正在如何變化，

of how the technology is changing,

and then we can get in the arena and push for futures we actually want.

然後我們可以進入競技場， 並推動我們真正想要的未來。

Thank you.

謝謝。

(Applause)

（掌聲）