Why AI Needs a “Nutrition Label” | Kasia Chmielinski | TED

翻译人员: Paris Yuan 校对人员: suya f.

Now, I haven't met most of you or really any of you,

现在，我还没有见过你们中的大多数人， 也没有见过你们中的任何人，

but I feel a really good vibe in the room.

但我感觉房间里的气氛非常好。

(Laughter)

（笑声）

And so I think I'd like to treat you all to a meal.

所以我想 请大家吃一顿饭。

What do you think?

你觉得怎么样？

Yes? Great, so many new friends.

是吗？太棒了，这么多新朋友。

So we're going to go to this cafe,

所以我们要去这家咖啡馆，

they serve sandwiches.

他们供应三明治。

And the sandwiches are really delicious.

而且三明治真的很好吃。

But I have to tell you that sometimes they make people really, really sick.

但我必须告诉你，有时候 它们会让人非常生病。

(Laughter)

（笑声）

And we don't know why.

而且我们不知道为什么。

Because the cafe won't tell us how they make the sandwich,

因为咖啡馆不会告诉我们他们是 怎么做三明治的，

they won't tell us about the ingredients.

所以他们不会告诉我们食材。

And then the authorities have no way to fix the problem.

然后当局没有办法解决这个问题。

But the offer still stands.

但是这个提议仍然有效。

So who wants to get a sandwich?

那么谁想吃三明治？

(Laughter)

（笑声）

Some brave souls, we can talk after.

一些勇敢的人，我们可以事后再谈。

But for the rest of you, I understand.

但是对于你们其余的人来说，我明白。

You don't have enough information

你没有足够的信息

to make good choices about your safety

来做出正确的安全选择，

or even fix the issue.

甚至无法解决问题。

Now, before I further the anxiety here, I'm not actually trying to make you sick,

现在，在我进一步阐述焦虑之前， 我并不是想让你感到恶心，

but this is an analogy to how we're currently making algorithmic systems,

但这可以类比我们 目前如何制造算法系统，

also known as artificial intelligence or AI.

也称为人工 智能或人工智能。

Now, for those who haven't thought about the relationship

现在，对于那些还没有 考虑过人工智能和三明治

between AI and sandwiches, don't worry about it,

之间关系的人， 不用担心，

I'm here for you, I'm going to explain.

我在这里等你，我会解释的。

You see, AI systems, they provide benefit to society.

你看，人工智能系统， 它们为社会带来好处。

They feed us,

它们养活我们，

but they're also inconsistently making us sick.

但也会不一致地 使我们生病。

And we don't have access to the ingredients that go into the AI.

而且我们无法获得 进入人工智能的要素。

And so we can't actually address the issues.

因此，我们实际上无法 解决这些问题。

We also can't stop eating AI

我们也无法停止吃人工智能，

like we can just stop eating a shady sandwich

就像停止吃阴凉的三明治一样，

because it's everywhere,

因为它无处不在，

and we often don't even know that we're encountering a system

而且我们通常甚至不知道我们遇到

that's algorithmically based.

的是一个基于算法的系统。

So today, I'm going to tell you about some of the AI trends that I see.

因此，今天，我将向大家 介绍我所看到的一些人工智能趋势。

I'm going to draw on my experience building these systems

我将借鉴我在过去二十年中 构建这些系统的

over the last two decades to tell you about the tools

经验， 向大家介绍我和 其他人为研究这些人工智

that I and others have built to look into these AI ingredients.

要素而开发的工具。

And finally, I'm going to leave you with three principles

最后，我要给你们讲三项原则

that I think will give us a healthier relationship

我认为这将使我们 与开发人工智能的公司

to the companies that build artificial intelligence.

建立更健康的关系。

I'm going to start with the question, how did we get here?

我首先要问一个问题，我们是 怎么来到这里的？

AI is not new.

人工智能并不是什么新鲜事物。

We have been living alongside AI for two decades.

我们与人工智能共处了二十年。

Every time that you apply for something online,

每当你在网上申请、

you open a bank account or you go through passport control,

开设银行账户 或通过护照检查时，

you're encountering an algorithmic system.

你都会遇到一个算法系统。

We've also been living with the negative repercussions of AI for 20 years,

20年来，我们也一直忍受着 人工智能的负面影响，

and this is how it makes us sick.

这就是它让我们生病的原因。

These systems get deployed on broad populations,

这些系统部署 在广泛的人群中，

and then certain subsets end up getting negatively disparately impacted,

然后某些子集最终会受到 不同的负面影响，

usually on the basis of race or gender or other characteristics.

通常是基于种族、 性别或其他特征。

We need to be able to understand the ingredients to these systems

我们需要能够了解这些 系统的组成部分，

so that we can address the issues.

这样我们才能解决问题。

So what are the ingredients to an AI system?

那么，人工智能系统的组成部分是什么？

Well, data fuels the AI.

好吧，数据为人工智能提供了动力。

The AI is going to look like the data that you gave it.

人工智能看起来会 像你给它的数据。

So for example,

因此，举例来说，

if I want to make a risk-assessment system for diabetes,

如果我想建立一个 糖尿病风险评估系统，

my training data set might be adults in a certain region.

我的训练数据集可能是某个地区 的成年人。

And so I'll build that system,

因此，我将建立这个系统，

it'll work really well for those adults in that region.

它将非常适合该地区 的成年人。

But it does not work for adults in other regions

但是它不适用于其他地区的成年人，

or maybe at all for children.

甚至根本不适用于儿童。

So you can imagine if we deploy this for all those populations,

因此，你可以想象，如果我们 对所有这些人群进行部署，

there are going to be a lot of people who are harmed.

将会有很多人受到伤害。

We need to be able to understand the quality of the data before we use it.

在使用数据之前， 我们需要能够了解数据的质量。

But I'm sorry to tell you that we currently live

但我很抱歉地 告诉你，我们目前生活

in what I call the Wild West of data.

在我所谓的数据狂野西部。

It's really hard to assess quality of data before you use it.

在使用数据之前， 真的很难评估数据的质量。

There are no global standards for data quality assessment,

没有数据质量评估 的全球标准，

and there are very few data regulations around how you can use data

关于如何使用数据

and what types of data you can use.

和使用哪些类型的数据法规也很少。

This is kind of like in the food safety realm.

这有点像食品安全领域。

If we couldn't understand where the ingredients were sourced,

如果我们无法理解 这些食材的来源，

we also had no idea whether they were safe for us to consume.

我们也不知道 它们是否可以安全食用。

We also tend to stitch data together,

我们也倾向于将数据拼接在一起，

and every time we stitch this data together,

每当我们将 这些数据拼接在一起时，

我们可能会在互联网上找到

which we might find on the internet, scrape, we might generate it,

这些数据，抓取， 我们可能会生成这些数据，

we could source it.

我们都可以获取这些数据。

We lose information about the quality of the data.

我们会丢失 有关数据质量的信息。

And the folks who are building the models

而且，构建模型

are not the ones that found the data.

的人并不是发现数据的人。

So there's further information that's lost.

因此，还有更多的信息丢失了。

Now, I've been asking myself a lot of questions

现在，我一直在问自己很多

about how can we understand the data quality before we use it.

关于在使用数据之前如何了解 数据质量的问题。

And this emerges from two decades of building these kinds of systems.

这源于二十年来 建造此类系统的历程。

The way I was trained to build systems is similar to how people do it today.

我接受系统构建培训的方式 与当今人们的操作方式类似。

You build for the middle of the distribution.

你在发行版 的中间进行构建。

That's your normal user.

那是你的普通用户。

So for me, a lot of my training data sets

因此，对我来说， 我的许多训练数据集

would include information about people from the Western world who speak English,

将包含有关来自西方世界的人的信息

这些人会说英语，

who have certain normative characteristics.

具有某些规范特征。

And it took me an embarrassingly long amount of time

令人尴尬的是， 我花了很长时间才意识

to realize that I was not my own user.

到自己不是自己的用户。

So I identify as non-binary, as mixed race,

因此，我认定自己是非二进制， 作为混血儿

I wear a hearing aid

我戴着助听器,

and I just wasn't represented in the data sets that I was using.

只是我使用的数据集中没有我的代表。

And so I was building systems that literally didn't work for me.

所以我正在构建 对我根本不起作用的系统。

And for example, I once built a system that repeatedly told me

例如，我曾经建立过一个系统 ，它反复告诉我

that I was a white Eastern-European lady.

我是一位东欧白人女士。

This did a real number on my identity.

这是我的身份上的真实数字。

(Laughter)

（笑声）

But perhaps even more worrying,

但也许更令人担忧的是，

this was a system to be deployed in health care,

这是一个 要部署在医疗保健领域的系统，

where your background can determine things like risk scores for diseases.

你的背景可以决定诸如疾病 风险评分之类的东西。

And so I started to wonder,

于是我开始怀疑，我

can I build tools and work with others to do this

能否构建工具并 与其他人合作来做这件事，

so that I can look inside of a dataset before I use it?

这样我就可以在使用数据集之前 查看数据集的内部？

In 2018, I was part of a fellowship at Harvard and MIT,

2018年，我参加了哈佛和 麻省理工学院的协会，

and I, with some colleagues, decided to try to address this problem.

我和一些同事 决定尝试解决这个问题。

And so we launched the Data Nutrition Project,

因此，我们启动了数据营养项目，

which is a research group and also a nonprofit

这是一个研究小组 ， 也是为数据集建立营养标签的非

that builds nutrition labels for datasets.

营利组织。

So similar to food nutrition labels,

与食品营养标签非常相似，

the idea here is that you can look inside of a data set before you use it.

这里的想法是，在使用数据集之前， 你可以先查看数据集的内部。

You can understand the ingredients,

你可以了解成分，

see whether it's healthy for the things that you want to do.

看看它对你想做的事情是否有益。

Now this is a cartoonified version of the label.

现在这是该标签的卡通化版本。

The top part tells you about the completion of the label itself.

顶部告诉你标签本身的完成情况。

And underneath that you have information about the data,

其下方有关于数据、描述、

the description, the keywords, the tags,

关键字、标签的信息，

and importantly, on the right hand side,

更重要的是，在右边， 你应该和不应该

how you should and should not use the data.

如何使用这些数据。

If you could scroll on this cartoon,

如果你能在这部动画片上滚动，

you would see information about risks and mitigation strategies

你会看到 有关

across a number of vectors.

多个向量的风险和缓解策略的信息。

And we launched this with two audiences in mind.

我们推出这个游戏 时考虑了两个受众。

The first audience are folks who are building AI.

第一批受众是正在构建 AI 的人。

So they’re choosing datasets.

因此，他们在选择数据集。

We want to help them make a better choice.

我们想帮助他们做出更好的选择。

The second audience are folks who are building datasets.

第二个受众是正在构建数据集 的人。

And it turns out

事实证明

that when you tell someone they have to put a label on something,

，当你告诉别人 他们必须在某件事上贴标签时，

they think about the ingredients beforehand.

他们会事先考虑 食材。

The analogy here might be,

这里的比喻可能是，

if I want to make a sandwich and say that it’s gluten-free,

如果我想做一个三明治 并说它不含麸质，

I have to think about all the components as I make the sandwich,

我必须在做三明治、面包和食材、

the bread and the ingredients, the sauces.

酱汁时考虑所有成分。

I can't just put it on a sandwich and put it in front of you

我不能把它放在三明治上 然后放在你面前

and tell you it's gluten-free.

然后告诉你它不含麸质。

We're really proud of the work that we've done.

我们为自己 所做的工作感到非常自豪。

We launched this as a design and then a prototype

我们将其作为设计推出， 然后是原型，

and ultimately a tool for others to make their own labels.

最终成为其他人 制作自己标签的工具。

And we've worked with experts at places like Microsoft Research,

我们还与微软研究院、联合国等地

the United Nations and professors globally

的专家以及全球的教授合作，

to integrate the label and the methodology

将标签和方法论整合到

into their work flows and into their curricula.

他们的工作流程和课程中。

But we know it only goes so far.

但是我们知道这只会持续很长时间。

And that's because it's actually really hard to get a label

那是因为实际上很难

on every single dataset.

在每个数据集上贴上标签。

And this comes down to the question

这归结为一个问题，

of why would you put a label on a dataset to begin with?

一开始你为什么要在数据集上加标签？

Well, the first reason is not rocket science.

好吧，第一个原因不是火箭科学。

It's that you have to.

这是你必须做的。

And this is, quite frankly, why food nutrition labels exist.

坦率地说，这就是食品营养 标签存在的原因。

It's because if they didn't put them on the boxes, it would be illegal.

这是因为如果他们不把它们 放在箱子里，那将是非法的。

However, we don't really have AI regulation.

但是，我们实际上 没有人工智能法规。

We don't have much regulation around the use of data.

我们对数据的使用没有太多规定。

Now there is some on the horizon.

现在有一些即将到来。

For example, the EU AI Act just passed this week.

例如，欧盟人工智能法案本周刚刚通过。

And although there are no requirements around making the training data available,

尽管对提供训练数据 没有要求，

they do have provisions for creating transparency labeling

但他们确实有创建 透明度标签的规定，

like the dataset nutrition label, data sheets, data statements.

例如数据集营养标签、 数据表、数据声明。

There are many in the space.

太空中有很多。

We think this is a really good first step.

我们认为这是一个非常好的第一步。

The second reason that you might have a label on a dataset

你可能在数据集上 贴上标签的第二个原因是因为

is because it is a best practice or a cultural norm.

这是一种最佳实践 或文化规范。

The example here might be how we're starting to see

这里的例子可能是我们 如何开始看到

more and more food packaging and menus at restaurants

越来越多的食品包装 和餐厅菜单中

include information about whether there's gluten.

包含有关是否含有麸质的信息。

This is not required by law,

这不是法律所要求的，

although if you do say it, it had better be true.

尽管如果你这么说， 那最好是真的。

And the reason that people are adding this to their menus

and their food packaging

和食品包装中，

is because there's an increased awareness of the sensitivity

是因为人们越来越 意识到这种过敏

and kind of the seriousness of that kind of an allergy or condition.

或疾病的敏感性和严重性。

So we're also seeing some movement in this area.

因此，我们在这一领域 也看到了一些动向。

Folks who are building datasets are starting to put nutrition labels,

正在建立数据集的人们 开始在他们的数据集上

data sheets on their datasets.

贴上营养标签和数据表。

And people who are using data are starting to request the information.

而使用数据的人 也开始索取这些信息。

This is really heartening.

这确实令人振奋。

And you might say, "Kasia, why are you up here?

然后你可能会说： “Kasia，你为什么在这里？

Everything seems to be going well, seems to be getting better."

一切似乎都进展顺利， 似乎越来越好。”

In some ways it is.

在某些方面确实如此。

But I'm also here to tell you that our relationship to data

但我在这里也要告诉你， 我们与数据的关系

is getting worse.

正在恶化。

Now the last few years have seen a supercharged interest

现在，在过去的几年中， 人们对收集数据集

in gathering datasets.

的兴趣激增。

Companies are scraping the web.

各公司正在抓取网页。

They're transcribing millions of hours of YouTube videos into text.

他们正在将数百万小时 的 YouTube 视频转录为文本。

By some estimates, they'll run out of information on the internet by 2026.

据估计，到 2026 年， 他们将在互联网上耗尽信息。

They're even considering buying publishing houses

他们甚至在考虑收购出版社

so they can get access to printed text and books.

这样他们就可以获得 印刷的课本和书籍。

So why are they gathering this information?

那么他们为什么要收集这些信息呢？

Well, they need more and more information

好吧，他们需要 越来越多的信息

to train a new technique called generative AI.

来训练一种名为 生成式人工智能的新技术。

I want to tell you about the size of these datasets.

我想告诉你这些数据集 的大小。

If you look at GPT-3, which is a model that launched in 2020,

如果你看一下在 2020 年 推出的模型 GPT-3，

the training dataset included 300 billion words, or parts of words.

训练数据集包含 3,000 亿个 单词或部分单词。

Now for context, the English language contains less than a million words.

现在，就上下文而言，英语 包含的单词不到一百万个。

Just three years later, DBRX was launched,

仅仅三年后，DBRX问世，

which was trained on eight trillion words.

它接受了八万亿字的训练。

So 300 billion to eight trillion in three years.

因此，三年内将达到 3000 亿至 8 万亿美元。

And the datasets are getting bigger.

而且数据集越来越大。

Now with each successive model launch,

现在，随着模型的连续推出，

the datasets are actually less and less transparent.

数据集的透明度实际上越来越低。

And even we have access to the information,

而且，即使我们有机会获得信息，

it's so big, it's so hard to look inside without any kind of transparency tooling.

信息量如此之大，

如果没有任何透明度工具， 就很难了解内部。

And the generative AI itself is also causing some worries.

生成式人工智能本身 也引起了一些担忧。

And you've probably encountered this technique through ChatGPT.

而且你可能已经通过 ChatGPT 遇到了这种技术。

I don't need to know what you do on the internet,

我不需要知道你在互联网上做什么，

that's between you and the internet,

那是在你和互联网之间，

but you probably know, just like I do,

但你可能知道， 就像我一样，使用 ChatGPT

how easy it is to create information using ChatGPT

和其他生成式 AI 技术

and other generative AI technologies

创建信息并将其发布到网络上是

and to put that out onto the web.

多么容易。

And so we're looking at a situation

因此，我们正在研究这样一种情况

in which we're going to encounter lots of information

我们将遇到大量通过算法生成的信息

that's algorithmically generated but we won't know it

但我们不知道

and we won't know whether it's true.

也不会知道这是否属实。

And this increases the scale of the potential risks and harms from AI.

这增加了人工智能潜 在风险和危害的规模。

Not only that, I'm sorry,

对不起，不仅如此，

but the models themselves are getting controlled

模型本身也

by a smaller and smaller number of private actors in US tech firms.

由越来越少的美国科技 公司私人行为者控制。

So this is the models that were launched last year, in 2023.

因此，这是去年， 即 2023 年推出的车型。

And you can see most of them are pink, meaning they came out of industry.

你可以看到它们中 的大多数是粉红色的，

这意味着它们是从工业中脱颖而出的。

And if you look at this over time, more and more are coming out of industry

如果你随着时间的推移来看看， 越来越多的人走出工业

and fewer and fewer are coming out of all the other sectors combined,

而来自所有其他领域的总和越来越少，

including academia and government,

包括学术界和政府，

where technology is often launched in a way

在这些领域

that's more easy to be scrutinized.

技术的推出往往更容易受到审查。

So if we go back to our cafe analogy,

因此，如果我们回到 我们的咖啡馆打个比方，

this is like you have a small number of private actors

就好像 有少数私人演员

who own all the ingredients,

拥有所有食材，

they make all the sandwiches globally,

他们在全球范围内制作所有三明治，

and there's not a lot of regulation.

而且没有太多的监管。

And so at this point you're probably scared

因此，此时 你可能会感到害怕

and maybe feeling a little uncomfortable.

可能会感到有点不舒服。

Which is ironic because a few minutes ago, I was going to get you all sandwiches

这很具有讽刺意味，因为几分钟前， 我本来要给你们买所有的三明治

and you said yes.

而你说好的。

This is why you should not accept food from strangers.

这就是为什么 你不应该接受陌生人的食物。

But I wouldn't be up here if I weren't also optimistic.

但是，如果我不是很乐观 的话，我就不会来这里。

And that's because I think we have momentum

那是因为我认为 我们在监管和文化变革

behind the regulation and the culture changes.

背后势头强劲。

Especially if we align ourselves with three basic principles

特别是如果我们坚持

about how corporations should engage with data.

关于企业应如何处理 数据的三项基本原则。

The first principle is that companies that gather data should tell us

第一个原则是，收集数据的公司

what they're gathering.

应该告诉我们他们在收集什么数据。

This would allow us to ask questions like, is it copyrighted material?

这将允许我们提出一些问题， 比如，它是受版权保护的材料吗？

Is that information private?

这些信息是私密的吗？

Could you please stop?

你能停下来吗？

It also opens up the data to scientific inquiry.

它还为科学研究开辟了数据。

The second principle is that companies that are gathering our data should tell us

第二个原则是，收集我们数据的公司 应该在对据做任何事情之前告诉我们

what they're going to do with it before they do anything with it.

他们将如何处理这些数据。

And by requiring that companies tell us their plan,

而且，要求公司 告诉我们他们的计划，

this means that they have to have a plan,

这意味着他们必须制定计划，

which would be a great first step.

这将是很好的第一步。

It also probably would lead to the minimization of data capture,

这也可能会最大限度地减少数据采集，

because they wouldn't be able to capture data

因为

if they didn't know what they were already going to do with it.

如果他们不知道自己 已经打算用数据做什么，

就无法捕获数据。

And finally, principle three,

最后，原则三，构建人工智能的

companies that build AI should tell us about the data

公司应该告诉我们他们用来

that they use to train the AI.

训练人工智能的数据。

And this is where dataset nutrition labels

这就是数据集营养标签

and other transparency labeling comes into play.

和其他透明度标签发挥作用的地方。

You know, in the case where the data itself won't be made available,

你知道，在数据本身无法提供的情况下，

which is most of the time, probably,

在大多数情况下，

the labeling is critical for us to be able to investigate the ingredients

标签对于我们能够研究成分

and start to find solutions.

并开始寻找解决方案至关重要。

So I want to leave you with the good news,

因此，我想给你们留个好消息，

and that is that the data nutrition projects and other projects

那就是数据 营养项目和其他项目

are just a small part of a global movement

只是全球人工智能问责运动

towards AI accountability.

的一小部分。

Dataset Nutrition Label and other projects are just a first step.

数据集营养标签 和其他项目只是第一步。

Regulation's on the horizon,

监管即将到来，

the cultural norms are shifting,

文化规范正在发生变化，

especially if we align with these three basic principles

特别是如果我们 符合这三项基本原则，

that companies should tell us what they're gathering,

即公司应该告诉我们他们在收集什么，

tell us what they're going to do with it before they do anything with it,

在对数据做任何事情之前 告诉我们他们将要用这个数据做什么，

and that companies that are building AI

以及正在构建人工智能的公司

should explain the data that they're using to build the system.

应该解释他们用来构建系统的数据。

We need to hold these organizations accountable

我们需要让这些组织

for the AI that they're building

对他们正在构建

by asking them, just like we do with the food industry,

的人工智能负责， 就像我们对待食品行业一样，

问他们内部有什么， 你是如何制造出来的？

what's inside and how did you make it?

Only then can we mitigate the issues before they occur,

只有这样，我们才能 在问题发生之前，

as opposed to after they occur.

而不是在问题发生之后缓解问题。

And in doing so, create an integrated algorithmic internet

通过这样做，可以创建一个 对所有人来说都更健康

that is healthier for everyone.

的集成算法互联网。

Thank you.

谢谢。

(Applause)

（掌声）