人工智能的真相（2023）（3）

Warfare has many elements of chance.
敌方会欺骗 会使诈
The enemy tries to deceive, to be tricky.
众所周知 战局既迷局
There's the proverbial fog of war.
我们无法得知战争中会发生什么
We don't know what happens in warfare.
这就表示不可能有足够数据训练人工智能
That means an AI is not likely to have enough data to be trained
使其可靠 或者能被可靠使用
reliably or to be able to be used reliably,
也就是说会犯错误
which means mistakes happen.
这将是大问题
That's a big issue.
好 好的
Ok, ok.
在你所有的顾虑中
Now, of all of your concerns that you have,
对人工智能用于战争的最大担忧是什么
what's the single biggest concern about AI in warfare?
能想到的噩梦场景有很多 但其中之一
There are many nightmare scenarios I can think of, but one of the
令我们非常担心的就是所谓的闪电战
things that we're concerned about are so-called flash wars.
闪电战
Flash wars?
闪电战是指 一个人工智能系统
So a flash war would be when an AI
甄别出潜在威胁
system identifies a potential threat
然后在几秒之内
and then, within seconds,
向另一人工智能系统发出行动信♥号♥♥
pings off an action onto another AI system
使其启动武器平台
which activates perhaps a weapons platform.
随即由下一个人工智能进行反应
And then another AI reacts to that.
所有这些都瞬间发生 几秒之内
And all of that happens within a flash, within seconds,
速度之快 没人能及时监管
so much so that no human will have oversight
了解发生了什么
over what is actually happening,
发生原因 或者如何停止
why it is happening, or how to stop it.
即人工智能将事态升级到战争 却没人
So that's AI just escalating until we reach a war without humans
有机会进行干预 -没错
having the opportunity to intervene. - That's right.
好 请大家看大屏幕
Well, OK, this is on the screen in front of us.
这是斯坦尼斯拉夫·彼得罗夫空军中校
This is Lieutenant Colonel Stanislav Petrov.
我想大部分观众对他一无所知
Now, I think most of you won't have heard of him,
但要不是彼得罗夫 我们今天都不会存在
but if it weren't for Petrov, we might not be here today.
1983年 冷战正酣时 彼得罗夫当值
In 1983, at the height of the Cold War, Petrov was on duty
苏联的预警系统显示
when the Soviet Union's early warning systems indicated
美国的导弹可能即将来袭
a potential incoming missile strike from the United States.
彼得罗夫的职责是立刻上报该攻击
And Petrov's job was to immediately report the attack.
如果他当时上报了
And if he had done,
苏联就会反击 发射原♥子♥弹♥
a retaliatory nuclear strike could have been launched.
但他没有上报
But he didn't.
他介入了决策过程 认为这是误报
He intervened and decided it was a false alarm
并上报了系统的错误
and reported a fault in the system.
他是对的
And he was right.
之后的调查发现 原来是苏联卫星
And an investigation later revealed that Soviet satellites
误将云朵反射的太阳光
had misidentified sunlight glinting on clouds
识别为导弹的引擎
as the engines of missiles.
埃尔克 你觉得在这样的情况下
Elke, in a situation like that,
人工智能的判定会引发灾难吗
do you think an AI response might have led to a catastrophe?
肯定的
Absolutely.
人工智能会按照程序进行下去
And AI would have unfolded its programme.
只能在程序背景之下看问题
It would have stayed within the context of its operation.
而彼得罗夫花了一点时间
What Petrov did is he took a moment,
跳脱出程序背景
he thought beyond the context,
暂停了一下 做出了不一样的决定
and he took a pause and made the unobvious decision.
这是人工智能做不到的 -是的
An AI system cannot do that. - Ok.
这些都是重要的问题 也是各位观众
These are big issues and issues that our audience is going
将在未来设法解决的问题
to have to grapple with in the future.
埃尔克 谢谢你让我们对其有了更好的理解
Thanks for helping us to understand them, Elke.
非常感谢
Thanks very much.
当人工智能的决定会影响人类时
Deciding, deciding how much autonomy we want AI to have,
人工智能应该拥有多少自主♥权♥
where the decisions that it makes will affect human beings
这是接下来我们要解决的重要问题
is a fundamental question that we need to start addressing now.
有人认为 人工智能可以解放人类 比如你我
Now, some people think that AI might free us from the messy,
让我们不用去做那些
illogical decision-making that human beings,
棘手且不合逻辑的决策
like all of you and me, are prone to.
有谁没吃上午饭还能保持心情愉悦呢
So who amongst us hasn't got irritable when we skipped lunch?
我们的心理和身体状态
Our mental and our physical state
会随时影响我们的决策
affects our decisions all the time.
而且我们所有人都持有偏见
And we, all of us, have biases.
有的偏见是我们能意识到的
Some of those biases we're conscious and aware of,
但很多偏见是潜意识的
but many are subconscious.
我们甚至不知道它们的存在
We're not even aware of them.
最糟糕的是 人工智能会放大人类的偏见
And the worst thing is that AI can actually amplify human bias.
这是如何发生的呢
To help us understand how this can happen,
请欢迎计算机科学教授 苏·布莱克
please welcome Professor of Computer Science Sue Black.
欢迎 苏 谢谢你的到来
Welcome, Sue. Thanks for joining us.
你来帮我们理解人工智能中的偏见
And you're going to help us to understand bias in AI.
是的 你好 麦克
Yes, absolutely. Hi, Mike.
大家好
And hi, everybody.
我想做一项实验 来解释数据集中
I'd like to do an experiment which might shed some light
可能出现的偏见 -好
on the potential for bias in data sets. - Ok.
早些时候 我们随机挑选了六名观众
Now, earlier on, we selected six people at random from the audience,
让他们提前准备好来参与接下来的实验
and we prepared them for the following experience.
有请这六位观众上台
So can those six people that we that we identified, please come down now?
请稍微挪一挪 好了
Hey, just a little bit. OK, that's great.
很好
Fantastic.
这就是我们可爱的六名志愿者
We've got six fine-looking volunteers here,
苏 他们需要做什么呢
Sue, what are they going to do?
我们的志愿者将成为数据集
So our volunteers are going to be our data set and our audience,
在座的其他观众将成为人工智能
you all are going to be the AI.
我们将用志愿者的数据训练大家
We're going to train you on our volunteers and create a rule
创造一条决策规则
that you can use to make decisions.
请仔细观察志愿者
Take a good look at our volunteers.
他们最显眼的共同特征是什么
What's the most distinctive feature that unites all of them?
他们所有人身上都有的最明显的特征
The most obvious thing that you can notice about everybody here.
仔细观察
OK, take a good look at them
我们花点时间想想
and let's just think about it for a moment.
让我来听听大家的想法
And I'm going to come to some of you in the audience.
你先来吧 -圣诞毛衣
I'll come to you first. - Christmas jumpers.
圣诞毛衣 好的
Christmas jumpers. OK.
我觉得这个答案不错 苏
I think that's a pretty good answer, Sue.
他们都穿着圣诞毛衣
They're all wearing Christmas jumpers.
没错
Yes, absolutely.
那么 根据这个 你就得到了一条规则
So, based on that, you've got a rule
能让你做出决定
that you can use to make decisions.
这条规则就是
And the rule is that all visitors to
所有来到圣诞讲座的观众都穿着圣诞毛衣
the Royal Institution Christmas lectures wear Christmas jumpers.
那么 根据该规则 扮演人工智能的观众们
Now, audience, as our AI, and according to that rule,
你们能否告诉我
can you tell me, would this visitor...
请看大屏幕 你们能否告诉我
Oh, have a look at our screen. ..would this visitor
这位观众是否能参加圣诞讲座
get to come to the Christmas lectures.
是或否 根据我们的规则
Yes or no? According to our rule?
是的话请举手
Hands up if you think yes.
他穿了圣诞毛衣 对吧 -很好
Yeah, they're wearing a Christmas jumper, right? - Great.
她能参加吗
And how about this person?
是或否
Yes or no?
没有穿圣诞毛衣
No Christmas jumper there.
不能
No.
那么它呢
And how about this visitor?
很好 能来
Very good. Yes.
非常感谢我们的志愿者们
Thank you very much, volunteers.
请回到你们的座位
You can take your seats again.
苏 尽管如此 没穿圣诞毛衣的人
But Sue, surely people who don't wear Christmas jumpers
也该有资格来参加我们的圣诞讲座
should nevertheless be allowed to come to the Christmas lectures.
是的 没错
Yes, absolutely.
不过观众们从接收的数据集中得到的信息是
But our audience learnt from the data set that they were given,
来参加圣诞讲座的人全都穿着圣诞毛衣
people that come to the Royal Institution only wear Christmas jumpers.
我们扮演人工智能的观众
Ok. So the audience, the AI audience that
从数据中发现了一个规律
you have, recognised a pattern
他们也没有错
in our data and the AI wasn't wrong.
问题在于给到他们的数据
It's just the problem with the data that it was given.
是的 我们给人工智能的数据并不准确
Yes, the data we gave the AI was skewed.
并不能代表一般的人群
It's not representative of the general population.
所以 这就是一个简单的例子