视频库 / NO.115ASK THE BEST MINDS THE BIG QUESTIONS
视频库 / NO.115
字幕 字幕位置
--:--
点击播放,这里会跟随视频显示当前句的中英字幕。

Robot morality and moral machines. Colin Allen

节目发布 2013-05-02 · The Monthly
科林·艾伦 PPhilip
本期追问 · 点击跳到视频对应位置
1:17 阿西莫夫的三定律真能当机器人的道德方案吗?13:15 没有愤怒的杀人机器,真的比士兵更守伦理吗?16:30 先给机器装伦理理论,还是让它像孩子一样学?27:12 造有道德的机器,是在理解机器还是理解我们自己?
归入 Ⅲ·04 道德只对人有效吗? →
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
编者按:本文是美国哲学家科林·艾伦(Colin Allen)的一场公开讲座实录。艾伦兼治动物认知、心智哲学与人工智能伦理,时任印第安纳大学教授,此次讲座围绕他与耶鲁大学生物伦理中心的温德尔·瓦拉赫(Wendell Wallach)合著的新书《道德机器》(Moral Machines)展开,由主持人菲利普引介开场。全文依据现场录音编译整理,仅删去口语枝节与重复,论证、例证与关键细节均予保留。

开场:跨学科哲学家的新书

主持人:我对今天这场讲座期待已久。科林是少数真正把跨学科研究做成了的哲学家:他兼具哲学、生物学、认知科学与神经科学的功底,能够如他自己所说,从各个角度帮助我们理解心灵。许多哲学家会朝神经科学挥挥手,许多神经科学家也会朝哲学挥挥手,生物学家则同时朝哲学和神经科学挥手,可真要把这些知识融会贯通,既懂得足够多,又足够聪明到能把所有信息整合起来,是极难做到的。今天科林要讲的题目非常有意思,也许眼下还不算贴近我们的生活,但在未来几十年里很可能变得切身相关。他将介绍他的新书《机器人的道德》。接下来的四十五分钟,有请科林·艾伦。

艾伦:谢谢菲利普。也感谢各位在周六上午赶来。我很高兴能在这里介绍这个项目,以及这本我与耶鲁大学生物伦理中心的温德尔·瓦拉赫合著的《道德机器》。

阿西莫夫三定律只是情节装置

艾伦:每当我们提起自己在做什么,这个项目总会引出各种各样的疑问,也常常惹人不快。我们最常听到的第一个问题是:阿西莫夫不是早就给出答案了吗?熟悉科幻小说的人都知道阿西莫夫的机器人三定律。事实上,对于「如何让机器合乎道德地行事」这个问题,这个答案已经出现在某些正式的政府行动里了。韩国政府向来想在一切技术领域走在前面,几年前宣布要为机器人制定一部伦理准则,并且明确表示将从阿西莫夫定律中汲取灵感。韩国信息通信部的官员朴海永(Park Hae-young)还说过这样一句话:想象一下,如果有人把人形机器人当作自己的妻子来对待,会是什么情形。这里透着一种道德义愤的意味,但这句话本身太含糊,我不打算深究。

那么,阿西莫夫究竟解决了这个问题没有?他最早提出机器人三定律,是在1941年写成、1942年发表的小说《转圈圈》(Runaround)里。从这个故事看得很清楚,阿西莫夫把三定律当作一种情节装置,一种用来制造难题与悖论的手段。故事的情境是这样的:两名宇航员被困在一颗环境恶劣的行星上,他们离开这颗星球所需的物资,人类无法靠近,于是他们命令机器人去取。第二定律说,机器人必须服从人类。机器人朝着物资所在的地点走去,途中意识到再往前走就会被毁掉,这违反第三定律。可第二定律要它照命令行事。而第一定律,也就是防止人类受到伤害,同样要求它去把东西拿回来,否则这两个人就会困死在那里。它怎么办?它陷入了死循环。它转起了圈圈,绕着那些东西一圈一圈地转,哪里也去不了,两名宇航员只能站在一旁,看着它围着他们逃生所需的物资打转。

所以这显然是一种情节装置。而且,要制造这类难题,甚至用不着三条定律,单是第一定律就够了。凡上过伦理学入门课的人都知道,哲学家最爱给学生出这种题:无论你怎么做,总有人要受害。经典的电车难题就是如此:一列火车正朝轨道上的五个人冲去,你可以用某种办法让它改道或者停下,但这样会撞死另一个人,一个而不是五个,你怎么选?很明显,单凭第一定律无法帮你解决这种局面。后来阿西莫夫又加了一条第零定律:机器人不得伤害人类整体,或因不作为而使人类整体受到伤害。这只是让问题成倍增加。

从科幻到现实:无人机与人形机器人

艾伦:有人会说,这不过是小说而已。当然,这一直是许多小说反复采用的母题,而且人们描述这些问题、描述机器人技术走向的方式,有许多确实让它听上去像是虚构。《太空堡垒卡拉狄加》的影迷会认出这幅图,这是剧中的一种人工智能战斗机。那的确是虚构,我们眼下还没有任何东西接近那种精密程度。但这一幅就不是科幻了。这是美军的「死神」(Reaper),也就是「捕食者」无人机的第二代,目前由内华达州拉斯维加斯的一个基地远程操控,飞一架至少要四个人。军方正积极设法用软件取代这些人。他们目前声称的目的是用于训练:培训一名新手飞行员的成本极高,因为另外三个座位得由三位专家来填,他们希望由软件接手这些角色,好让培训更快、更省钱。可是,如果你已经有了能驾驶模拟器的软件,那离能驾驶真机的软件也就不远了。人们还有一种迷思,以为在任何开火决定上,人类都会始终留在决策环节里。但军事专家已经开始承认,事实并非如此,这些东西将来会在外面执行任务,而人类不可能时时刻刻全面监督它们的一举一动。

同样,我们会想到《终结者》,那是科幻。可事实上,好几个地方的人正在研制人形机器人,只不过多半是服务性质的。这是本田的阿西莫(ASIMO)。在日本的语境里,它被设想为一种你会想放在家里的东西,尤其是放在老年人身边,因为日本的年轻人不够多,照顾不过来老人。用机器人替代护理人员,是日本机器人研究一个明确宣示的目标。

我想强调的是,我并不认为这是十五年、二十年之后才会变得切身的事,它此刻就已经切身了。我们正在把这些东西带进自己的日常生活。而且这不只关乎杀人机器,也关乎非常琐碎的事情。我们固然不会养一条需要训练大小便的机器狗,但确实有公司在设想让我们同机器狗建立情感上的、投入的关系。这是索尼的爱宝(AIBO),几年前已经停产,不过我想它还会回来。如今这东西在eBay上的价格,比当年零售时还贵。

看不见的软件bot已在做决定

艾伦:而且,这还不只是硬件机器人,也包括软件bot。我们所处的环境已经是这样:每次你刷信用卡消费,都是某台机器在决定批不批准这笔交易,没有任何人在直接监管。如果你不满意这个决定,可以打电话,运气好的话能和一个真人说上话,但那个人同样盯着一块电脑屏幕,按屏幕上显示的内容行事。那个程序即使没有直接指挥,也在左右这个人能替你做出什么决定。此外还有许多领域,这类被称作bot而非robot的软件已经就位。比如关键系统的维护调度:这里左上角是军事应用,用于军用装备的维护,但同样的东西也用在电网等系统上。「赌注天使」(Bet Angel)据说能帮你赌马赚钱,有趣的是他们把它包装成一位天使。「电子机器人」(ebots)则是一种交易软件,替你做买卖决定。这类软件使用极广,而且常被认为在一定程度上加剧了市场的不稳定,因为它们决策极快,事情的发生速度远超市场设计时的预期,市场本来是按人类的决策速度设计的。

所以,在所有这些正在积极开发和部署的东西当中(是现在,不是未来),我们已经有了许多做买卖、做信贷审批的软件代理或bot。我们还有一批伦理咨询系统,专门设计来帮助医务人员就病人权利和职业伦理做决定,有人正在开发这类系统,此外还有面向重症监护的分诊系统。所以,虽然我口头上说的是机器人,我其实想把bot也包括进来。只是机器人更上镜,所以你们看到的都是机器人的照片。软件很难拍出照片来。

从伦理盲到功能性道德

艾伦:当我们提起道德机器,或者我们简称为AMA的人工道德主体(artificial moral agents)这个话题时,还会遇到另一种反应:机器人不可能真的有道德吧?它们只是照吩咐办事,没有任何决策能力,也没有意识,凡是我们认为对自己重要的那些东西,它们一样都没有。我愿意承认,完整的、人类水平的道德主体性也许要等二十年、五十年乃至一百五十年之后。就这一点而言,我们谈的不是科幻,那种系统不是我心里想的东西。我心里想的是我们已经看到的那些系统:它们在自主性这根轴上不断上移,越来越多的决定不再需要人类直接介入。你也许会说,那不是真正的自主,它们没有自由意志。好吧,我不在乎这一点。要害在于,它们已经在外面做决定了,这些决定以或好或坏的方式影响着我们的生活。而这些系统总体上处于我所说的「伦理盲」(ethically blind)或者伦理迟钝的状态,也就是在横轴上靠左的那一端。

所以,在我看来,挑战在于把这些系统朝伦理敏感的方向推进:让它们能够觉察,也就是能够处理关于那些在道德上、伦理上、在福祉意义上对我们至关重要的事情的信息。在推进的过程中,你首先会希望程序员能预见到机器人将遇到的一切情形。可这是不可能的,因为我们将把这些东西投放到越来越开放的环境里,不可能把将会遇到的每一种情况都预先想到。于是就要从一种「运行性道德」(operational morality),即一切都由程序员预先写死,走向我们所说的「功能性道德」(functional morality),即机器自身就装有能处理相关信息并据此行动的系统。我们认为,功能性道德是未来十到十五年内可以实现的事。至于道德主体性,那就不好说了。我的合著者温德尔认为我们也许永远做不到。我看不出原则上有什么理由做不到,但我也认为,这方面存在大量夸大其词的乐观。

启动这个项目、写这本书的时候,我其实受到了罗莎琳德·皮卡德(Rosalind Picard)一句话的启发,她主持麻省理工学院的情感计算实验室,这句话出自她大约十二年前的书:机器的自由越大,它就越需要道德标准。这话是什么意思?我们该怎么做到?还有一件事很耐人寻味。你们有些人也许看过彼得·W·辛格(Peter W. Singer)的书,不是那位著名的澳大利亚伦理学家彼得·辛格,而是另一位在美国几乎同样有名的彼得·辛格。他的《为战争接线》(Wired for War)几个月前出版,轰动一时。奇怪的是,这本书通篇谈战争伦理,却只字未提让系统自身装载伦理引导系统这件事。而我们接下来会看到,已经有人在做这件事了。

谁的标准?军方的冷血机器人设想

艾伦:这个项目和这本书引出的另一个问题是:我们究竟要实现谁的标准?这当然是个很难的问题。但我认为,人们很容易像哲学课上常见的那样,一上来就盯着所有的疑难案例,却没有意识到其实还有大量容易的案例。有一些行为标准即便不是普世的,至少也得到了广泛认同,这似乎是一个不错的起点。问题在于我们如何获得这些标准。詹姆斯·吉普斯(James Gips)是最早撰文讨论机器道德或机器人道德起源的人之一,他画过一幅相当简单化的漫画,我很喜欢,恰恰因为事情显然没有这么简单。

那我们现在该怎么办?我提到过,军方圈子对这个话题非常认真。美国佐治亚理工学院的罗纳德·阿金(Ronald Arkin)教授,最近一个月左右也出了一本书,你们可以看出这眼下是个热门话题,书名是《管束自主机器人的致命行为》。他是一位机器人学家,拿着美国军方的经费,为军用机器人研制机载伦理引导系统。阿金的观点里有很多我其实是赞同的,但也有些地方我觉得相当天真。这段话出自与该书同名的技术报告,我想书里也有:在战争迷雾之中,可以预期未来的自主机器人也许能比人类表现得更好,因为它们在设计上可以没有情绪,不会被情绪蒙蔽判断,也不会因战场上的事态而愤怒沮丧。意思是,人会发怒,发怒之后就干坏事;而机器人不会发怒,它们将是冷血的杀手,可正因为如此,我们反而可以指望它们表现得更好。

情绪是双刃剑,伦理不是套规则

艾伦:我认为这种想法天真,因为我确实认为我们的情绪是一件双刃的东西,沿用军事的比喻,是一把双刃剑。士兵会发怒,这在某些情形下恰恰让他们打得更好。而且我认为世上确有所谓义愤,它在维护正义方面起着极其重要的作用:你对不义感到愤怒,正是这种愤怒驱使你去做点什么。所以在我看来,情绪的好用处和坏用处能否一刀切开,并不清楚。但这一点给我们凸显了一件事:人类的伦理决策不只是遵循一套干巴巴的规则,而是要以某种方式真正投入到眼前的情境之中。

自上而下进路与框架问题

艾伦:现在请你们想象一下,或者说像我们在书里所做的那样,由我来想象:我作为哲学家坐在办公室里,进来一位工程师,说,上面让我造一台合乎伦理地行事的机器人,我该怎么办?大多数哲学家的第一反应,多半是开始滔滔不绝地讲某种伦理理论,也就是采取我们所说的「自上而下」(top-down)的进路:从我们现有的最好的理论出发。这里用杰里米·边沁的功利主义来举例,当然,在座的哲学家都知道它有许多精细化的版本,还有伊曼努尔·康德的义务论,也就是以权利或义务为基础的观点。然后去考虑如何实现一个遵循这些理论所规定的准则的系统。出于时间关系,我没法展开讲,问答环节可以再谈,但这条路问题成堆。

我只说一个,而且在这点时间里也没法把它讲透,那就是计算机科学家所说的「框架问题」(frame problem):要计算这些理论中的某一个是否被正确地应用,潜在需要考虑的东西实在太多,以至于在现实可用的时间内根本不可能算完。计算开销之所以巨大,一个原因是你得了解你所影响的那些人的心理。伦理说到底是对他人的关切,我做的事会不会让别人不快,是要紧的,而要知道我是否惹你不快,是否违背了你的某个愿望,或者你认为在道德上相关的任何其他东西,都需要对人类心理有某种深入的理解。此外,你还得知道行动在世界上会产生什么后果。功利主义的经典表述是为最大多数人谋最大幸福,可行动会产生涟漪效应,无限延续下去,你也许不得不一路算到遥远的未来,才能知道这是否真是最好的做法。还有一个始终存在的问题,就是你何时才知道自己可以开始计算了:你收集到的信息够不够做决定了?买车之类的事我们都有体会,那不是道德决定,可你照样能花上几个星期,反复琢磨自己了解得是否已经足够多,可以拍板了。当你试图用一种非常形式化的自上而下进路来做伦理决策时,同样的问题就冒出来了。

自下而上:学习、演化与情绪模型

艾伦:另一条路是某种「自下而上」(bottom-up)的进路,我前面已经有所暗示,这里再多说几句。一个想法是,我们也许可以设计会学习的机器。这在人工智能领域其实是个很老的想法。艾伦·图灵在他1950年那篇开创性的文章《机器能思考吗》里就说过,也许我们应该造一个儿童机器人,让它像正常人一样成长发育,而不是试图从零开始造出一个成人。这正是我这里展示的几张日本机器人项目照片所做的事。另一种办法是,与其让系统学习,不如让系统演化:在人工演化中,以计算机作为环境,按某种标准挑选出表现更好的那些,再从那里往上累积。我还提到过情绪。也许我们需要的系统并不是靠计算去应用某种理论,而是拥有恰当的情绪反应。这里我要小心一点,因为我已经承认过它们也许没有真正的情绪,但我们可以给这类情绪建模。这方面有大量工作正在进行:如何让机器表现得仿佛拥有情绪,并以此来引导行为。

我在印第安纳大学的同事马蒂亚斯·肖伊茨(Matthias Scheutz)正在研究会违抗命令的机器人。他发现,在某些情形下,人们能够容忍机器违抗直接命令。让人们容忍这种机器的因素有几个。一个是,人们觉得机器人在以某种方式同你合作,朝着某个更远的目标努力。即便它拒绝了你认为必须执行的某条具体命令,只要你能看出它改做的事仍然服务于那个目标,你就会容忍它。另一件事是,他的实验用的是真实的硬件机器人,随着情况越来越紧急,他让机器人提高嗓门,说「我现在必须这么做,我现在必须这么做」。机器人警告你几次之后终于动手,人们也就接受了,因为机器人身上表现出了情绪的样子。你可以说这是个卑鄙的把戏,但它管用。它再次指向了我们那种社会性的互动方式,而这正是自下而上进路所强调的。

混合进路与现有研究版图

艾伦:不过我们认为,你需要的是某种混合进路。毕竟,我们在事后能够反思,当时看起来对的事,我们是否真该照做。因此,我们需要的人工道德主体,既要保持自下而上系统那种动态、灵活的道德,能容纳各种各样的输入,又要用自上而下的原则来评估这些行动,或许还能依据这些原则来解释自己做出的决定。但话说回来,就目前而言,我们即便没有完全进入所谓的科幻长远领域,也已经接近边缘了,就实际系统而言,这些还只是我们在摸索的东西。

那么我们眼下究竟走到了哪一步?我不会逐个细讲,但书中考察了不少项目,它们正尝试把某种伦理推理或决策能力构建到机器里,采用的方法五花八门:有的用逻辑,用道义逻辑(deontic logic);有的像MedEthEx系统那样,采用医学伦理中广受认可的一些原则;还有人工神经网络、基于主体的建模(agent-based modeling)。这些术语我没时间一一解释,但总之人们在从各种不同的方向着手。与此同时,还有另一批人在研究我们可以构建到机器人身上的那种情感的、具身的、情绪性的反应。这两个群体才刚刚开始走到一起,彼此大致知道对方在做什么,但还没有人把这些零件拼到一起。不过我认为,看到这一切拼合起来,并不需要等太久。

事实上,我最近参加答辩委员会的一位学生做过一个实验:在屏幕上呈现一位医疗病人,向被试(假定为医生)吐露某件隐私,然后问被试,你会把这件事告诉病人的配偶吗?结果发现,这个人物的长相和动作,会影响人们是否表示愿意告诉配偶。有趣的是,这对男性被试的影响大于女性被试。由此可见,伦理决策与bot、机器人、软件虚拟系统的所谓呈现层面,正在以各种方式汇合到一起。

为何叫不停:军事逻辑与情感依恋

艾伦:还有一个问题:这一切太可怕了,简直不敢想,我们为什么不干脆叫停?我们真的想要这些东西吗?我接下来的话会显得有点宿命论:我认为,现在再问这个问题几乎已经太晚了。我们已经有了朝这个方向走的东西。当然,事情从来不会真的太晚,但你等得越久,就越难办。所以,如果我说得谨慎一点,我不会说它不可避免,但就实际而言,它也许就是不可避免的。

这背后的逻辑,部分是由军事驱动的,比如美国陆军的武装机器人车辆,也就是「未来作战系统」网页上的东西。我截图之后他们已经改版了,但那些内容还在。他们正在研制越来越好、越来越强的无人机。这是他们下一款自主飞行器的原型,图上放了一枚美国一美分硬币作比例,可见它相当小,但它同样具备侦察能力,可能还有某些防御或进攻能力。这里的逻辑再清楚不过:士兵会死,如果能用机器人替代他们,那看起来就是件好事。iRobot公司是罗德尼·布鲁克斯(Rodney Brooks)从麻省理工学院分拆出来的企业,产品从机器人吸尘器Roomba到军用机器人PackBot都有。这里引用的是美国海军陆战队一位一等兵的话:「我相信你们今天救了人命。」这位陆战队员的故事很有意思。他当时在伊拉克,他们用一台PackBot探测地雷和简易爆炸装置,结果PackBot被炸毁了。他们给这台PackBot起名叫史酷比。事后他写信给公司的首席执行官,说,请把史酷比修好,它救了我的命。我们在这里看到的是,人会同外形一点也不像人的机器建立感情。他对这个东西真的产生了依恋。

当然,这条路上会有起有落。我前面提到日本大力推动机器人进入养老护理,2007年的一则报道说,一些初期实验并不完全成功。路透社这篇报道的标题是「在老龄化的日本,机器人让老人们兴味索然」。我最喜欢结尾那句话,一位老太太冷冷地说:毛绒玩具更受欢迎。但这些只是有待解决的技术问题,依恋的程度,互动的程度。何况我们眼下面对的这一代老人,成长过程中没有短信,也没有其他种种电子玩意儿。等到今天这一代人到了那个年纪,我想他们对互动设备的兴趣会远远超过毛绒玩具。

真正的担忧与造物照见自身

艾伦:最后两三张幻灯片,然后我很乐意回答各位的问题。我们该担心什么?我认为不是机器人接管世界这种担忧。几年前有本很有意思的书,写得很幽默,叫《如何在机器人起义中活下来》,书里满是这样的主意:机器人来抓你了,就往楼上跑。机器人在物理环境中的活动确实存在难题,但这些难题正在被解决。我建议各位上YouTube看看「大狗」(BigDog)机器人,你们会惊讶于它在极其崎岖的地形上有多灵活,甚至能在冰面上背着重物保持平衡。这些问题都会解决。但我认为,比起机器人接管世界,我们有的是时间去阻止那个,有一些更紧迫的忧虑。

其一,正如开场介绍里提到的,这些机器做出的粗糙伦理评估可能酿成大规模灾难。它们对该敏感的东西并不敏感,只对其中一部分敏感。你可以想象一起枪击事件之类的事故,因为情境没有被恰当地纳入考虑。这里还有一个大问题,就是人类自主性的削弱:当我们依赖这些机器替我们做决定,或者引导我们的决定时,我们究竟在对自己做什么?此外还有意料之外的后果。那位韩国部里的官员并没有离谱到哪里去,他担心的正是人们开始把机器人当作配偶会发生什么。事实上,大卫·利维(David Levy)最近就出了一本书,他的整篇博士论文写的就是与机器人正式结婚、建立伴侣关系,我认为这仍属科幻。但凡是关注过技术与色情产业如何相互作用的人都知道,色情从一开始就在驱动技术,甚至早在电话出现之前的传真机时代就是如此。所以,这类东西也正在被迅速推进。

最后,我想留给各位一个问题,这是我的倒数第二张幻灯片:创造机器人这项事业,说明了我们自己什么?我们眼下的处境是,我们不再通过观察动物,把它们作为自然界中与我们最相似的存在来对照着理解自己,而是在按自己的形象创造东西。将来,与我们最相似的将是这些造物。我们的本性或心理中有某种东西驱使我们去做这件事,这本身就很耐人寻味,它是一种自我理解的行动。我们在书里提出,思考如何建造道德机器这整个课题,实际上迫使我们认真地想一想:有道德究竟意味着什么,合乎伦理究竟意味着什么,伦理理论又扮演着什么角色。

结束之前,我邀请各位去看看我们的博客,moralmachines.blogspot.com,我们在那里跟踪这个领域的最新进展。你们可以在那里进一步了解这本书,读到其中几章,也欢迎留言评论。非常感谢。

排版 + 横图 + 来源,粘贴即成稿
章节 · 点击跳转视频
0:04 开场:跨学科哲学家的新书 ▶ 正在看
1:17 阿西莫夫三定律只是情节装置 ▶ 正在看
4:37 从科幻到现实:无人机与人形机器人 ▶ 正在看
7:58 看不见的软件 bot 已在做决定 ▶ 正在看
10:05 从伦理盲到功能性道德 ▶ 正在看
13:15 谁的标准?军方的冷血机器人设想 ▶ 正在看
15:23 情绪是双刃剑,伦理不是套规则 ▶ 正在看
16:30 自上而下进路与框架问题 ▶ 正在看
18:34 自下而上:学习、演化与情绪模型 ▶ 正在看
20:41 混合进路与现有研究版图 ▶ 正在看
23:56 为何叫不停:军事逻辑与情感依恋 ▶ 正在看
27:12 真正的担忧与造物照见自身 ▶ 正在看
本期小问 · 档案清单
1:17 阿西莫夫的三定律真能当机器人的道德方案吗? ▶ 正在看
13:15 没有愤怒的杀人机器,真的比士兵更守伦理吗? ▶ 正在看
16:30 先给机器装伦理理论,还是让它像孩子一样学? ▶ 正在看
27:12 造有道德的机器,是在理解机器还是理解我们自己? ▶ 正在看
本期讲者
科林·艾伦美国哲学家,研究动物认知、心智哲学与人工智能伦理,时任印第安纳大学教授。与 Wendell Wallach 合著《Moral Machines: Teaching Robots Right from Wrong》(2009)。
Philip本场讲座主持人,负责介绍讲者及其跨学科背景。
01开场:跨学科哲学家的新书
0:04
I've been looking forward to this a lot because Colin is one of the rare philosophers who manages to pull off a genuine interdisciplinary project he's got the expertise in philosophy biology cognitive science and Neuroscience to really help us understand the mind as you like as you might put it from all angles um a lot of philosophers wave at Neuroscience a lot of neuroscientists wave at philosophers a lot of biologists wave at philosophy and Neuroscience but it's very difficult to pull off the integrative project to know enough and be smart enough to combine all the information and today's project that K's going to talk about is something extremely interesting and perhaps something that if it's not close to our hearts now might be close to our hearts in the next um few decades con will be talking about his new book robot morality so here is to um tell us a bit more about those ideas for the next 45 minutes Colin Allen thank you Philip well thank you all for coming out on a Saturday morning here I'm very
我一直非常期待这场活动,因为 Colin 是少数几位真正做成了跨学科研究的哲学家之一,他在哲学、生物学、认知科学和神经科学方面都有专业造诣,能够真正帮助我们从各个角度去理解心智。很多哲学家会向神经科学招招手,很多神经科学家会向哲学家招招手,很多生物学家会向哲学和神经科学招招手,但要真正完成这种整合性的工作是非常困难的——你得懂得足够多,也得足够聪明,才能把所有这些信息融会贯通。今天 Colin 要讲的内容非常有意思,而且也许现在还不算是我们特别关心的事,但在未来几十年里可能就会变得与我们息息相关。Colin 将谈谈他的新书,关于机器人道德。下面就请他用接下来的 45 分钟给我们多讲讲这些想法。有请 Colin Allen。谢谢你,Philip。也谢谢各位在周六上午前来。我很
便签笔记
02阿西莫夫三定律只是情节装置
1:17
pleased to be here and tell you a little bit about this project and this book moral machines that I've co-authored with Wendell wall who's at the Yale Center for bioethics um the the project raises all sorts of questions and hackles whenever we mention what we're doing um and one of the first questions we always get is well didn't asof already provide the solution to this those of you who are familiar with with science fiction will know well asmos 3 laws and in fact this answer to the to to the question of how we make machines behave morally or ethically it has already shown up in some official governmental actions so the government of South Korea always wanting to be on the Forefront of anything technological uh a couple of years ago announced that they would be working on an ethical code for uh robots um and very explicitly said that they would take inspiration from asimov's laws um this quote also from Park ha young who's the South Korean minister of information or he's an official at the
高兴能来这里,跟大家介绍一下这个项目,以及这本书——《道德机器》,是我与耶鲁大学生物伦理学中心的 Wendell Wallach 合著的。这个项目每次我们一提起自己在做什么,就会引发各种各样的疑问,甚至让人炸毛。我们最常被问到的第一个问题就是:阿西莫夫不是早就给出解决方案了吗?熟悉科幻的朋友们都知道阿西莫夫的三大定律。而事实上,这个关于我们如何让机器在道德或伦理上守规矩的答案,已经出现在了一些官方的政府举措中。比如韩国政府一向想站在任何技术领域的最前沿,几年前就宣布他们将着手制定一部针对机器人的伦理准则,而且非常明确地表示,他们会从阿西莫夫的定律中汲取灵感。这段引言同样出自朴海荣(音)——他是韩国信息部部长,或者说是信息与通信部的一位官员——
便签笔记
2:28
ministry of information and community Communications imagine if some people treat Androids as if the machines were their wives so there's this suggestion of moral outrage which uh might happen that's so ambiguous I'm not sure I want to go there um but did asof in fact solve the problem well in his early uh introduction of the three laws of robotics the first of which shows up in a story he wrote in 1941 and that was published in 1942 um run around it's very clear that asof regards his three laws of robotics as a plot device a way to generate problems and paradoxes so that story runaround the scenario is a couple of astronauts are stranded on a on a hostile planet um and the material they need to get off the planet is uh unapproachable by humans so they order the robot to go get it then the second law says a robot must obey a human being the robot approaching this location where this material is needed um realizes that if it goes any further it's going to be destroyed thus violating the third law but the second
“想象一下,如果有些人把机器人当成自己的妻子来对待。”这里暗含着一种道德上的义愤,也许真的会发生。这话太含混了,我不太确定我想不想往那个方向讲。但阿西莫夫真的解决了这个问题吗?在他早期提出机器人三大定律时——第一条出现在他 1941 年写的一篇小说里,1942 年发表,叫《转圈圈》——非常清楚的是,阿西莫夫把他的机器人三大定律当作一种情节装置,用来制造问题和悖论。在《转圈圈》这个故事里,情节是这样的:两个宇航员被困在一颗充满敌意的星球上,他们要离开这颗星球所需的原料,人类无法靠近去取,于是他们命令机器人去取。第二定律说,机器人必须服从人类的命令。机器人走近这个存放原料的地点时,意识到如果再往前走,自己就会被摧毁,从而违反第三定律;但第二定律又要求它照吩咐去做,
便签笔记
3:39
law tells it to do what it's told anyway and the first law preventing harm to humans requires it to go and uh obtain the material right otherwise they're going to be stranded there and die what does it do it gets stuck in a loop it run it's a runaround it goes round and round round and gets nowhere and the two astronauts are left there sort of watching this thing go around in circles around what they need to get off um so it's very clear that this is a plot device and you don't even need all three laws to generate these kinds of problems um the first law alone will do it as anybody who's taken an introductory ethics course will know flossers love to pose students with problems where whatever you do somebody's going to come to harm the classic trolley cases there's a train hurtling towards five people on the track and by some means or other you can divert the train or stop it but it's going to kill some somebody one person instead of five what do you do um and clearly the the first law alone will will not resolve that
而防止人类受到伤害的第一定律也要求它去把原料取回来——否则那两个人就会被困死在那里。那它怎么办呢?它陷入了一个死循环,就在那儿转圈圈,绕啊绕啊绕,哪儿也去不了。那两个宇航员就只能待在那儿,眼睁睁看着这东西围着他们离开所必需的东西一圈圈打转。所以很清楚,这就是一个情节装置。而且你甚至不需要三条定律全上,就能制造出这类问题;光是第一定律就够了。任何上过伦理学入门课的人都知道,哲学家们最爱给学生出那种难题:无论你怎么做,总有人要遭殃。受到伤害。经典的电车难题:一列火车正冲向轨道上的五个人,而你可以通过某种方式让火车转向或让它停下,但这样会害死另外某个人——死一个人而不是五个人,你会怎么做?显然,光靠第一定律是没法帮你解决这种情况的。而且即便阿西莫夫后来又加了一条第零定律,说机器人
便签笔记
03从科幻到现实:无人机与人形机器人
4:37
situation for you uh and even when uh asof adds a zero law later that a robot may not injure Humanity or through an action allow Humanity to come to harm it just multiplies the problems well isn't this just fiction and uh it's of course been uh a trope through much fiction and there there are many ways of casting what the issues are and where we're going with robotics in ways that make it seem as though it is fiction so fans of battl Star Galactica will recognize this as one of the um one of the artificial intelligent uh flying Fighters um and and that is fiction we don't have anything nearly as sophisticated as that at this point but this is not science fiction right this is a us Reaper which is the second edition of The Predator unmanned aerial vehicle currently operated remotely uh from a base in Las Vegas Nevada takes at least four people to fly one of these things the military is actively seeking to replace those individuals uh with software um their stated purpose at the moment is for training purposes it's
不得伤害人类整体,也不得因不作为而坐视人类整体受到伤害,这也只是让问题变得更多。那么,这不就只是科幻小说吗?当然,这在很多虚构作品里一直是个常见的桥段,人们有很多种方式来描绘这些问题,以及机器人技术的走向,这些方式往往让它看起来就像是虚构的。所以《太空堡垒卡拉狄加》的粉丝会认出这是其中一架具备人工智能的飞行战机。那确实是虚构的,我们目前还没有任何东西能复杂到那种程度。但这个就不是科幻了,对吧——这是美军的“收割者”,也就是“捕食者”无人机的第二代,目前由内华达州拉斯维加斯的一个基地远程操控,操作这样一架飞机至少需要四个人。军方正在积极设法用软件来取代这些人员,他们目前对外宣称的目的是用于训练。训练一名新手飞行员操作这种飞机成本非常高,
便签笔记
5:52
very expensive to train a novice pilot on one of these things because you've got to have three experts to fill the other seats they would like software to uh take on those roles so that they can train people more rapidly and cheaply but of course if You' got software that can fly a simulator you're pretty close to having software that can fly the thing itself and there's a myth that people will stay in the loop uh in making any kill decisions here but military experts are beginning to recognize that that that in fact is not the case that these things are going to be out there doing things where it's not possible to completely oversee all of their actions all of the time so again uh we think of Terminator science fiction um but as a matter of fact uh in a number of places people are working on humanoid robots uh in this case in a more service capacity this is Honda's Asimo and uh it uh uh is envisioned in in the Japanese context as being something that you might want in your home as something that you uh might want
因为你得有三名专家来坐满其他座位。他们希望软件能承担这些角色,这样就能更快、更省钱地培训人员。但当然,如果你有一套软件能开模拟器,那你离拥有能真正驾驶这架飞机的软件也就不远了。而且有一种迷思,认为在任何击杀决策上,人始终会留在决策环节中。但军事专家们开始意识到,事实上并非如此——这些东西会在外面执行任务,而人不可能时时刻刻完全监督它们的所有行为。所以我们又会想到《终结者》那样的科幻。但事实上,在不少地方人们都在研究人形机器人,这一个更偏向服务用途,这是本田的ASIMO。在日本的语境里,它被设想成你可能会想放在家里的东西,也是你可能会想
便签笔记
6:51
to have around elderly people because there aren't enough younger Japanese to look after them it's a big stated objective of uh of Japanese robotics to try to replace healthcare workers with robots and I want to emphasize that actually I don't think we should be thinking about this or that'll become relevant 15 or 20 years down the road it's actually becoming relevant right now we are beginning to bring these things into our own uh day-to-day context um and it's not just about killing machines it's about very trivial kinds of things so while we're not going to have dogs robotically that we're going to have have to train to be house broken um nevertheless there are corporations out there that are trying to envisage us forming emotional affective engaged relationships with robotic dogs this is Sony's IBO which actually uh was discontinued a couple of years ago but um I think it'll be back um and these things now cost more on eBay than they did when they were being sold retail and it's not just Hardware
放在老年人身边的东西,因为日本没有足够的年轻人来照顾他们。日本机器人研究一个明确宣示的重要目标,就是设法用机器人取代医护人员。我想强调的是,我并不认为我们应该把这看成十五年或二十年之后才会变得有意义的事,它现在就已经有现实意义了。我们已经开始把这些东西带进日常生活中,而且这不只关乎杀人机器,也关乎非常琐碎的事情。比如说,我们不会有那种需要训练它们别随地大小便的机器狗,但尽管如此,还是有公司在试图设想我们与机器狗建立起有情感、有投入的关系。这是索尼的 AIBO,其实它在几年前就已经停产了,但我想它还会回来的。而且这些东西现在在 eBay 上的价格,比当年零售时还要高。而且不只是硬件机器人,还有软件机器人。我们现在所处的环境已经是:每次
便签笔记
04看不见的软件 bot 已在做决定
7:58
robots it's also software bots so we are in an environment already where every time you go run a credit card uh for a purchase some machine decides whether or not to approve that there's no direct human oversight of that if you don't like the decision you can get on the phone and hopefully talk to a person but that person is also looking at a computer screen and being Guided by what they see on that screen the program is in some sense if not directing influencing the decisions that they can make on your behalf um and there are a number of other contexts where where these kinds of software Bots as they're called rather than robots um are already in place so in in scheduling maintenance for critical systems so the top left here is actually a military application for maintenance of military hardware but it's also being used in electrical power grids and so on um the BET Angel is supposed to help you uh make money on the horses and uh it's interesting they pose it as as an angel and ebots is a
你刷信用卡消费时,都是某台机器在决定要不要批准这笔交易,这中间没有直接的人工监督。如果你不满意这个结果,你可以打电话,运气好的话能跟真人说上话,但那个人也是在看电脑屏幕,并被屏幕上显示的内容所引导。这个程序即便不是在直接指挥,某种意义上也在影响他们能替你做出的决定。还有很多其他场景中,这类被称作软件机器人(bots,而不是实体机器人)的东西已经在运作了。比如在安排关键系统的维护调度上,这里左上角其实是一个军用应用,用于军事装备的维护,但它也被用在电网等领域。Bet Angel 号称能帮你在赛马上赚钱,有意思的是他们把它塑造成“天使”的形象。而 eBots 是一款交易软件,能替你做出
便签笔记
9:00
trading uh software that makes trading decisions on your behalf um and these are widely used and often thought to contribute somewhat to instability of the market because they make very fast decisions and things happen much more quickly than uh than the markets are designed for namely they're designed for human uh speed decision making so as among all these things under active development and deployment not in the future we have a number of software agents or Bots that are doing buying and selling that are doing credit approval Etc and we have also a number of ethical advisory systems that are um actually designed to help medical professionals make decisions about patients rights and professional ethics and people are working on these and there's also triage systems for intensive care so although I talk about robots I really want to include Bots um but robots are more photogenic um so those are what those are the pictures you see it's hard to take a picture of software well another kind of reaction
交易决策。这些东西被广泛使用,而且常被认为在一定程度上加剧了市场的不稳定,因为它们做决策非常快,事情发生的速度远超市场原本所设计的节奏——也就是说,市场是按人类的决策速度设计的。所以在所有这些正在积极开发和部署的东西当中——不是在未来,而是现在——我们已经有一批软件代理或机器人在做买卖交易、做信用审批等等。我们还有一些伦理咨询系统,它们实际上是设计来帮助医疗专业人员就患者权利和职业伦理做出决定的,人们正在研究这些系统,另外还有重症监护的分诊系统。所以虽然我一直在讲机器人,但我其实也想把软件机器人包括进来。不过机器人更上镜,所以你们看到的图片都是它们,因为很难给软件拍照。当我们提出“道德机器”或“人工道德主体”
便签笔记
05从伦理盲到功能性道德
10:05
that we get when we bring up this topic of moral machines or artificial moral agents amaas we abbreviate it uh Is Well robots can't really be moral can they I mean they just do what they're told they don't have any decision- making capacity um they're not conscious all of these things that we think are important to us um and I want to concede that fullblown human level moral agency maybe 20 50 150 years in the future right we're not talking about science fiction in this respect those are not the kind of systems have that I have in mind what I have instead is systems which we already see are increasing on this autonomy scale they're making more and more decisions without direct human input now you might go that's not really autonomy they don't really have free will okay fine I don't care about that the point is they're out there making decisions which affect our lives in possibly negative and possibly positive ways and yet these systems are com are on the whole what I call ethically blind or
(我们缩写成 AMA)这个话题时,另一种常见的反应是:机器人不可能真的有道德吧?我是说,它们只是照着指令做事,它们没有任何决策能力,它们没有意识,而这些都是我们认为对自己很重要的东西。我愿意承认,完整的、人类水平的道德主体性也许要等二十年、五十年甚至一百五十年。在这一点上我们讨论的不是科幻,那不是我心里所想的那类系统。我想说的是我们已经看到的那些系统,它们在自主性这个尺度上不断上升,越来越多地在没有人类直接介入的情况下做决定。你可能会说,那不算真正的自主,它们并没有自由意志。好吧,无所谓,我不关心那个。重点是,它们就在外面做着影响我们生活的决定,可能是负面的,也可能是正面的。然而这些系统总体上是我所说的“伦理盲”,或者说
便签笔记
11:08
ethically insensitive so they're towards the left end of this access AC axis across the bottom and so the challenge as I see it is to try to move things into a more ethically sensitive Direction make these systems aware of in the sense of able to process information about the things that matter to us morally ethic and so on in terms of welfare um and as we do that the first thing is you hope that uh programers anticipate all of the situations that that uh the robots encounter but but that's not going to happen because we're going to increasingly release these in more and more open environments where it's impossible to completely anticipate everything that will be encountered uh so you go from a kind of operational morality where it's all built in from the programers end to a kind of what we call functional morality where the machines themselves have on board the systems that enable them to process the relevant information and act on it accordingly so we make we see functional morality as as something that's actually
对伦理不敏感的,所以它们处在下面这条横轴的左端。在我看来,挑战就是设法把事情推向更具伦理敏感性的方向,让这些系统“意识到”——意思是能够处理那些在道德、伦理以及福祉方面对我们真正重要的信息。在这个过程中,首先你会希望程序员能预料到机器人会遇到的所有情况,但这是做不到的,因为我们会越来越多地把它们投放到越来越开放的环境中,在那里不可能完全预料到会遇到的一切。所以你会从一种“操作性道德”——也就是一切都由程序员那一端内置好——走向我们所说的“功能性道德”,也就是机器自身搭载了相应的系统,使它们能够处理相关信息并据此采取行动。我们认为功能性道德其实是未来十到十五年内就能实现的事情。至于
便签笔记
12:12
within the next 10 to 15 years um for moral agency who knows my co-author Wendell thinks we may never get it I don't see any reason in principle why we won't but I also think there's a lot of overstated op optimism about that and in in starting this project and writing this book I was actually um inspired a bit by a comment by rosin Picard who runs the affective Computing lab um at MIT and uh this is from her book from about a dozen years ago where she said the greater the freedom of a machine the more it will need moral standards what does that mean how do we do that it's striking too some of you may have seen this book by Peter uh W singer not the famous Peter singer of the Australian ethicist but another Peter singer who's almost as famous in the US uh this book came out with a big splash a couple of months months ago wired for war it's striking that although it's all about the ethics of warfare there's nothing about having the systems themselves have on board ethical guidance systems but as we'll see people
完整的道德主体性,谁知道呢——我的合著者温德尔认为我们可能永远做不到,我则看不出原则上有什么理由说我们做不到,不过我也觉得在这件事上有很多被夸大的乐观。在启动这个项目、写这本书的时候,我其实有一点是受到了罗莎琳德·皮卡德一句话的启发,她主持着麻省理工学院的情感计算实验室。这句话出自她大约十二年前的那本书,她说:一台机器的自由度越大,它就越需要道德标准。这是什么意思?我们该怎么做到?还有一点也很值得注意,你们中有些人可能看过彼得·W·辛格的这本书——不是那位著名的澳大利亚伦理学家彼得·辛格,而是另一位在美国几乎同样有名的彼得·辛格。这本书几个月前出版时引起了很大反响,叫《Wired for War》。值得注意的是,尽管它通篇都在讲战争伦理,却完全没有提到让这些系统自身搭载伦理引导系统。但正如我们将看到的,已经有人
便签笔记
06谁的标准?军方的冷血机器人设想
13:15
are working on that another kind of question we get from this project from this book is well whose standards um are we going to implement here and that's that's of course a very difficult question and it's it's easy I think to U start focusing as often happens in philosophy classes on all of the hard cases without recognizing that in fact there's a lot of easy cases there are certain standards of behavior that are that are if not Universal at least widely agreed to and that seems like a good basis to build from um and the question is how we acquire those um James Gibs who was one of the first people to write about um uh the origin of uh of machine or robot morality uh Drew this rather simplistic cartoon which I like um because it it clearly isn't this simple all right um what are we going to do now I mentioned that this is a a topic that's being taken very seriously in military circles and Ronald Arin who's a professor at Georgia Tech University in uh the United States um this is another book out within the last
在做这方面的工作了。关于这个项目、这本书,我们收到的另一类问题是:我们要落实的到底是谁的标准?这当然是个非常困难的问题。我认为,人们很容易像哲学课上常发生的那样,一开始就盯着所有的疑难案例,却没有意识到其实还有大量的简单案例。有些行为标准即便不是普世的,至少也是被广泛认同的,这看起来是个不错的出发点。问题在于我们如何获得这些标准。詹姆斯·吉普斯是最早撰文讨论机器或机器人道德之起源的人之一,他画了这幅相当简化的漫画,我挺喜欢的,因为事情显然并没有这么简单。好,那我们现在该怎么办呢?我提到过,这个话题在军方圈子里被非常认真地对待。罗纳德·阿金是美国佐治亚理工学院的教授,这是最近一个月左右出版的另一本书,你们可以看出这眼下算是个热门话题,讲的是
便签笔记
14:22
month or so you can see this is kind of a Hot Topic at the moment on governing lethal behavior and autonomous robots he's a robot assist with money from the US military to build on board ethical guidance systems for military robots um and uh he's there's a lot I actually agree with in arin's view but there are some things that I find uh rather naive so uh this is a quote um that uh actually comes from the technical report that has the same title as the book but I think it's in the book also where he says in the fog of War it may be anticipated that in the future autonomous robots may be able to perform better than humans they can be designed without emotions that cloud their judgment or result in anger and frustration with ongoing Battlefield events so the idea is that people get angry and then they do bad things um whereas robots won't get angry they will be these coldblooded killers but because of that we can actually expect better things from them I think this is naive because I actually think our emotions
如何规范自主机器人的致命行为。他是一位机器人学家,拿着美国军方的经费,为军用机器人开发机载的伦理引导系统。阿金的观点里有很多我其实是赞同的,但也有一些我觉得相当天真的地方。这是一段引文,其实出自那份与书同名的技术报告,不过我想书里应该也有。他说,在战争迷雾中,可以预期未来的自主机器人也许能比人类表现得更好。它们可以被设计成没有那些会蒙蔽判断、或因战场上正在发生的事而产生愤怒与挫败的情绪。也就是说,人会发怒,然后做出坏事;而机器人不会发怒,它们会是冷血的杀手,但正因为如此,我们反而可以对它们期待更好的表现。我认为这想法很天真,因为我其实认为我们的情绪
便签笔记
07情绪是双刃剑,伦理不是套规则
15:23
are a double-edged thing uh a double-edged sword to continue the military metaphor so so so um the fact that soldiers get angry actually causes them to fight better in certain circumstances and that there is such a thing as uh righteous anger I think which plays a very important role in preserving Justice you get angry about Injustice and that's what motivates you to do something about it um so so it's not clear to me that you can separate off all of the good uh applications of emotion from the bad applications of emotion um but it highlights for us that ethical decision making in human beings is not just a matter of following some dry set of rules but of actually being engaged in some way with uh with the um with the situation at hand now if you imagine or if I imagine as we do in the book uh me sitting in my office as a philosopher and in comes an engineer saying I've been told to build a a robot that behaves ethically what do I do um the the likely reaction of most philosophers is going to be to start
是双刃的——沿用军事比喻,是一把双刃剑。士兵会发怒这件事,实际上在某些情况下会让他们打得更好;而且我认为确实存在“义愤”这种东西,它在维护正义方面扮演着非常重要的角色——你因不公而愤怒,而这正是促使你去做点什么的动力。所以我并不觉得你能把情绪所有好的作用和坏的作用彻底分开。但这也让我们看到,人类的伦理决策并不只是遵循一套干巴巴的规则,而是真的以某种方式投入到眼前的情境之中。现在,如果你想象一下,或者我想象一下——就像我们在书里写的那样——我作为一名哲学家坐在办公室里,这时一位工程师进来说:上头让我造一个行为合乎伦理的机器人,我该怎么办?大多数哲学家可能的反应,会是开始
便签笔记
08自上而下进路与框架问题
16:30
spouting off some ethical Theory um and take what we call a top- down approach so start with the best theories we have and Illustrated here with Jeremy bentham's utilitarianism and of course there are many refinements of that for the philosophers in the audience um and Emanuel kant's deontological or rights or Duty based view um and think about how we uh might go about implementing A system that would follow the guidelines that are specified by those theories and for reasons I don't have time to go into um but we can talk about it during question time um this approach is full of problems um and I'll mention just one and again I won't be able to do justice to it in the time available here but uh it's what computer scientists call the frame problem which is that potentially there's so much one needs to take into account in order to compute whether or not one of these theories is being applied correctly that it couldn't possibly be done in a realistic amount of available amount of time um so
大谈某种伦理理论,采取我们所说的“自上而下”的进路:从我们现有最好的理论出发,这里用杰里米·边沁的功利主义来举例——当然,对在场的哲学家来说,这方面有很多更精细的版本——以及伊曼努尔·康德的义务论,或者说基于权利、基于义务的观点。然后思考我们该如何去实现一套能够遵循这些理论所规定之准则的系统。由于时间关系,我没法细讲其中的原因,但我们可以在提问环节聊聊——这种进路问题重重,我只提其中一个。同样,在这里有限的时间内我没法把它讲透,但这就是计算机科学家所说的“框架问题”:要计算这些理论中的某一个是否被正确应用,可能需要考虑的东西实在太多,以至于根本不可能在现实可用的时间内完成。所以
便签笔记
17:36
there's there's large computational overhead due to problems of knowing all about the psychology of the people that you're affecting right so ethics is about regard for others and it matters whether or not they're going to get upset at something that I do that requires some sort of deep understanding of human psychology to to to to know whether or not I'm upsetting you or or violating some uh desire that you have have or whatever else you think is morally relevant here furthermore you have to know the effects of actions in the world so uh classically utilitarianism says um that it's the greatest good for the greatest number but actions have Ripple effects they go on indefinitely so you you might be stuck having to compute way way into out into the future to find out whether or not uh this is really the the best thing to do and um there's always a problem of knowing when you can actually do the computation have you gathered enough information yet to be able to make the decision and we're all familiar with
会有巨大的计算开销,因为你得了解你所影响的那些人的全部心理状况,对吧。伦理关乎对他人的顾及,而他们会不会因为我做的某件事而难过,这是有关系的,这就需要对人类心理有某种深刻的理解,才能知道我是不是让你不高兴了,或者违背了你的某种愿望,或者你认为在道德上相关的任何别的东西。此外,你还得知道行为在这个世界上的后果。经典功利主义说的是“最大多数人的最大幸福”,但行为会产生涟漪效应,无限地扩散下去,所以你可能不得不一路计算到很远很远的未来,才能弄清楚这到底是不是最该做的事。而且还始终有一个问题:怎么知道什么时候可以真正做出计算——你收集到的信息是否已经足够让你做出决定?我们对此都不陌生,比如买车之类的,那不是道德决定,但你可能会花上
便签笔记
09自下而上:学习、演化与情绪模型
18:34
this when buying a car or something it's not a moral decision but you can spend weeks wondering whether or not you know enough yet to make a decision right um and the same thing comes in space when you try to apply a very formal top- down approach to ethical decision making so the alternative is a kind of bottomup approach that I've already hinted at but let me say a little bit more about it um the thought that we might be able to design learning machines it's a very old idea in artificial intelligence in fact so Alan Turing in in his seal article in 1950 can machines think uh said maybe we should build a child robot and have it develop like a normal human being rather than trying to build an adult from scratch um and uh that's exactly what you see uh uh the project is in in this Japanese robotics uh uh project that I've got some shots of there alternatively instead of learning maybe we can evolve systems uh in artificial Evolution using computers as an environment in which we then select the
好几个星期琢磨自己是不是已经了解得够多、可以做决定了,对吧。同样的问题也会出现,当你试图把一种非常形式化的自上而下的进路应用到伦理决策上时。所以另一种选择,是我刚才已经暗示过的“自下而上”的进路,让我再多说一点。这个想法是:我们也许能设计出会学习的机器。这在人工智能领域其实是个非常古老的想法——阿兰·图灵在他1950年那篇著名的文章《机器能思考吗》里就说过,也许我们应该造一个“儿童机器人”,让它像正常人类那样发育成长,而不是试图从零开始造出一个成年个体。而这正是你们在这个日本机器人项目里看到的,我这里有它的几张照片。或者,除了学习之外,我们也许可以通过人工进化来“演化”出系统,把计算机当作一种环境,在其中按某些标准挑选出表现更好的个体,然后在此基础上逐步构建。另外我还
便签笔记
19:35
ones that do better by some criteria and build up from there and also I've mentioned emotions maybe we need systems which actually don't really apply some Theory computationally but have the right kinds of uh emotional responses now I want to be careful here because I've admitted that they may not have real emotions but we can model the kinds of emotions and there's a lot of work going on on this how we how we get machines to behave as if they have emotions um and use that as a way to guide Behavior so a colleague of mine at Indiana University mat schitz is actually working on robots that disobey um and finds that people will under certain circumstances uh tolerate machines that disobey direct orders um and one of the things that will get them to to tolerate such a machine there a couple of factors one is that the robot is perceived as cooperating with you in some way towards some further goal even though it might reject a particular specific order that you think needs to be done if you can
提到了情绪——也许我们需要的系统并不是真的在计算层面套用某种理论,而是具备恰当的情绪反应。这里我要谨慎一点,因为我已经承认它们可能并没有真正的情绪,但我们可以对各种情绪进行建模,而且这方面有大量研究正在进行,也就是我们如何让机器表现得好像有情绪一样,并以此来引导行为。我在印第安纳大学的一位同事马蒂亚斯·舒茨,就在研究会「抗命」的机器人,他发现在某些情况下,人们是可以容忍机器不服从直接命令的。而让人们愿意容忍这样一台机器,有几个因素:一是这个机器人被认为在某种意义上是在与你合作,朝着某个更高的目标努力——即便它拒绝执行你认为必须完成的某条具体指令,只要你能
便签笔记
10混合进路与现有研究版图
20:41
see that what it did instead actually still serves that goal then you'll tolerate it but the other thing he does is as the situation gets more and more urgent in his experiments with these are with real Hardware robots he raises the voice of the robot says I've got to do this now I've got to do this now and people except when the robot finally goes ahead and does it having warned you a couple of times because of this appearance of emotion in the robot right you can think that's nasty trick but it works and again points to our kind of interaction in this social way um that bottomup approaches emphasized we think however you need some kind of hybrid approach because after all we can after the fact uh reflect on whether or not we should have gone along with what seemed to be the right thing to do at the time and and thus we'll need artificial moral agents that that manage to maintain this Dynamic flexible morality of bottomup systems and accommodate all kinds of diverse inputs but evaluate those
看出它改做的事情其实仍然服务于那个目标,你就会容忍它。另一件他做的事情是,随着实验中情境变得越来越紧急——这些都是用真实硬件机器人做的实验——他会提高机器人的音量,让它说「我现在必须这么做,我现在必须这么做」。于是当机器人在警告了你几次之后终于自作主张去做了那件事时,人们是能接受的,因为机器人身上呈现出了这种情绪的表象。你可以觉得这是个卑劣的小把戏,但它确实有效。这再次指向了我们这种社会性的互动方式,也就是自下而上的进路所强调的东西。不过我们认为,你需要某种混合式的进路,因为毕竟我们可以事后回过头去反思:当时看起来正确的事情,我们究竟该不该照做。因此我们需要人工道德主体,它既能维持自下而上系统那种动态、灵活的道德性,能够容纳各种各样的多元输入,又能通过自上而下的原则来评估这些行为,甚至或许还能用
便签笔记
21:39
actions through top down principles and and perhaps can even explain the decisions that they made in terms of those principles but again you know we're if not verging perhaps even fully into uh long-term what might be called science fiction at the moment things that we're still only groping towards in in terms of of actual systems so so where where are we actually at the moment I'm just going to go through this I'm not going to go through these individually but there's a number of projects which we survey in the book um which uh are attempting to build some kind of ethical reasoning or decision making into uh into machines and doing it in a with a variety of different approaches using logic deontic logic um or using certain principles that are very widely accepted in medical ethics as with the med ethics system um and artificial neuron networks uh agent-based modeling these are all terms that uh that I won't have time to explain but the but people are tackling this in in various different ways
这些原则来解释自己所做的决定。但话说回来,我们在这里就算不是接近、恐怕也已经完全进入了目前只能称之为科幻的长期设想——就实际系统而言,这些还是我们只是在摸索的东西。那么我们眼下究竟到了什么阶段?我接下来只是过一遍这些内容,我不会逐个详细讲,但我们在书中综述了不少项目,它们都在尝试把某种伦理推理或决策能力构建到机器当中,而且采用的方式各不相同用逻辑、道义逻辑的方法,或者用一些在医学伦理学中被广泛接受的原则比如那个医学伦理系统,还有人工神经网络、基于智能体的建模,这些术语我都没时间一一解释了,但人们正在用各种不同的方式同时研究这个问题。也有人在
便签笔记
22:44
simultaneously there are people who are working on these affective embodied emotional kinds of reactions that that we can build into robots um the two communities are only just beginning to come together so so they're sort of aware of each other's work but nobody's put all these pieces together but but I think that we're really not talking very far down the line before we see this and in fact a student whose committee I was on recently did an experiment in which using onscreen presentations of um uh a a medical patient explaining something to you the subject as if you're the doctor revealing something private and then asking the subject would you tell this patient's uh spouse about this situation um it turned out how that individual looked and moved made a difference to whether people would in fact uh uh uh tell say said they would tell the spouse or not particularly affected males more than females interestingly male subjects more than female subjects um so so we see ways in which uh the sort of convergence of
研究那种情感的、具身的、带情绪的反应,我们可以把这些内建到机器人里。这两个领域才刚刚开始走到一起,所以他们大致知道彼此在做什么,但还没有人把这些拼图都拼起来。不过我觉得真的用不了多久我们就会看到这样的东西。事实上,我最近参加过一个学生的答辩委员会,他做了一个实验,在屏幕上呈现一个医疗病人向你——也就是受试者——解释某件事,就好像你是医生,病人向你透露了一些隐私,然后问受试者:你会不会把这个情况告诉这位病人的配偶?结果发现,那个人的长相和动作方式会影响人们是否愿意说他们会告诉配偶。有意思的是,这对男性的影响比女性更大,男性受试者比女性受试者受影响更多。所以我们看到了伦理决策与呈现层面——如果可以这么说的话——如何相互交汇
便签笔记
11为何叫不停:军事逻辑与情感依恋
23:56
ethical decision-making and uh presentational aspects if you like of bots and robots software virtual systems uh are beginning to come together well you know the other question is well look all of this is sort of too horrendous to con conceive of why don't we just stop it right do we really want these things um and I'm going to sound a bit like a fatalist and say I think it's too late almost to be asking this question we have things that are heading in this direction of course it's never really too late but the longer you wait the more and more difficult it gets so so so if I was a little bit more careful I wouldn't say it's inevitable but for all practical purposes it may be inevitable and the logic of it is partly driven by this military robots with the things like the US Army's armed robotic Vehicles the future combat systems web page they've actually changed their web page since I took the screenshot but it but it it's still there um and they are developing ever better and greater uh
机器人、软件机器人、虚拟系统等等,正开始融合到一起。当然,另一个问题是,你可能会说,你看这一切听起来太可怕了,简直难以想象,我们为什么不干脆叫停呢?对吧,我们真的想要这些东西吗?嗯,我这么说可能显得有点宿命论,但我认为现在再问这个问题,几乎已经太晚了。我们已经有一些东西在朝这个方向发展了。当然,永远不能说真的太晚,但你等得越久,事情就会变得越来越难办。所以,如果我说得再谨慎一点,我不会说它不可避免,但从实际的角度看,它可能就是不可避免的。而推动这个逻辑的部分原因,就是军用机器人,比如美国陆军的武装机器人车辆,还有“未来作战系统”的网页。其实自从我截图之后,他们已经改过这个网页了,但那些内容还在。他们正在研发越来越先进、越来越强大的
便签笔记
24:55
unmanned drones this is their next prototype for a for a flying uh uh uh autonomous vehicle that will be uh you can see a US penny in the diagram up there so it'll be rather small but it will also have a surveillance and possibly some um some other kinds of uh defensive or offensive capabilities and the logic of this is very clear because um soldiers die and if you can replace them with with robots that seems like a good thing to do right and so IR robot is actually a corporation that spun off from MIT Rodney Brooks that builds everything from the Roombar which is a robotic vacuum cleaner to the packbot which is a military uh robot um and as in this quive circle from a from a private first class in the US Navy a marine actually uh I believe you've saved our you've saved lives today this particular Marine has a very interesting story because uh he was in Iraq and they were using a packbot for um for mine IED detection and uh the packbot blew up and that was the packbot that they named Scooby-Doo after this and he wrote to
无人机。这是他们下一代原型机,一款会飞的自主飞行器,你可以看到图里上方有一枚一美分硬币做参照,所以它会相当小,但它同时还会具备侦察功能,可能还有一些别的用途。防御或进攻能力,这背后的逻辑非常清楚,因为士兵会牺牲,如果你能用机器人来替代他们,那似乎是件好事,对吧?所以 iRobot 其实是一家由 MIT 的 Rodney Brooks 分拆出来的公司,他们造的东西从 Roomba——也就是扫地机器人——到 PackBot 都有,PackBot 是一款军用机器人。这句话引用自美国海军的一位一等兵,其实是一名海军陆战队员,他说:“我相信你们拯救了我们,你们今天救了人命。”这位海军陆战队员的故事非常有意思,因为他当时在伊拉克,他们用一台 PackBot 来做地雷和路边炸弹的探测,结果那台 PackBot 被炸了,他们给那台 PackBot 起的名字叫Scooby-Doo。之后他写信给公司的 CEO,说请把 Scooby-Doo 修好,它救了我的命。所以我们
便签笔记
26:09
the CEO of the company saying please fix Scooby-Doo he saved my life so what we see here is also people bonding with quite nonhuman looking machines right he really felt attached to this thing now uh clearly there'll be ups and downs on this and a recent uh story well 2007 story I mentioned in Japan the big attempt to get robots into Elder Care some of the initial experiments haven't been entirely successful uh robots turn off senior citizens in aging Japan this says uh in this reuter story people I like the last comment stuffed animals are more popular she remarked dry um but those are just technological problems to be solved the degree of attachment the degree of interactive activity plus we're working with the generation of older people at the moment who didn't grow up with texting and all these other gadgets and so on right so when this generation gets to be that age I think they're going to be much more interested in the the interactive devices than the stuffed toys so last couple slides and
在这里看到的,也是人与看上去一点都不像人的机器建立起感情联系,对吧?他真的对这个东西产生了依恋。当然,这方面肯定会有起有落。最近有一则报道——其实是 2007 年的报道——我提到过日本大力尝试把机器人引入养老护理,有些早期实验并不算完全成功。“机器人让老龄化日本的老年人反感”,这是路透社的一则报道。我最喜欢最后那句评论:她干巴巴地说,毛绒玩具更受欢迎。不过这些只是有待解决的技术问题——依恋的程度、互动的程度。而且我们目前面对的这一代老年人,是没有在发短信和各种电子小玩意儿的环境中长大的,对吧?所以等我们这一代到了那个年纪,我觉得他们会对这些互动设备比对毛绒玩具更感兴趣。所以还剩最后几张幻灯片,然后我很乐意回答
便签笔记
12真正的担忧与造物照见自身
27:12
then I'll be uh happy to take your questions but what are the worries well I think it's not this worry of a robot takeover another interesting book from a couple of years ago humorously written how to survive a robot Uprising um the technology This Book is Full Of Interest idea is like if the robots are coming for you run upstairs um we there are clearly physical uh domain problems operating physical domain problems for robots but actually those are being uh solved I would I urge you to go look on YouTube for the big dog robot and you will be astounded just how flexible this is over very terrain even on Ice keeping its balance while carrying a heavy load so those problems are going to be solved but but actually I think more pressing though than the the of a robot takeover we'll have plenty of time to stop that I I think we have these more pressing worries one is um as was mentioned in the introduction uh that we could have some large scale catastrophe due to um crude ethical assessments being made by
大家的问题。那么担忧是什么呢?我觉得担忧并不是机器人接管世界。几年前还有一本很有意思的书,写得很幽默,叫《如何在机器人起义中幸存》。这本书里全是有趣的技术点子,比如说如果机器人来抓你,就往楼上跑。机器人在物理领域确实存在明显的问题,在物理领域中行动的问题,但其实这些问题正在被解决。我强烈建议你们上 YouTube 去看看BigDog 机器人,你会大吃一惊,它在非常崎岖的地形上有多灵活,甚至在冰面上,负着重物还能保持平衡。所以那些问题终将被解决,但我认为其实更紧迫的是不过比起机器人接管世界,我觉得我们有的是时间去阻止那种事,我认为我们面临的是更紧迫的担忧。第一个,就像刚才介绍里提到的,就是我们可能会因为这些机器做出粗糙的伦理判断而引发某种大规模的灾难,它们对该敏感的东西不敏感,只对其中一部分东西敏感,
便签笔记
28:12
these machines they're not sensitive to the right things they're only sensitive to some of the things and you can imagine a shooting incident or something like this where the situation wasn't taken into account properly a big issue here is reduction of human autonomy what are we doing by as it were depending on these machines to make uh decisions or to guide our decisions um and unintended effects and the uh the fellow from the ministry in South Korea um is not all that far off uh in wondering what will happen if people start to treat robots like uh their spouses um and in fact this is another recent book uh by uh David Levy who uh wrote a whole dissertation on what I think is still science fiction which is sort of formal marriage and relationship with robots but anybody who keeps track on how technology and the pornographic industry have interacted will know that pornography has driven technology from the very beginning even from faximile machines before there were telephones um and so uh we have uh we have a situation
你可以想象某种枪击事件之类的情况,当时的具体情境没有被恰当地纳入考虑。这里的一个大问题是人类自主性的削弱,我们这样去依赖这些机器来做决定、或者引导我们做决定,究竟是在做什么?还有那些意料之外的影响。韩国那位来自部里的先生,他担心的事其实没那么离谱——如果人们开始把机器人当作自己的配偶来对待,会发生什么?事实上,还有另一本近期的书,作者是大卫·列维(David Levy),他写了一整篇博士论文,讨论的在我看来仍属于科幻的东西,也就是跟机器人缔结某种正式的婚姻和关系。但凡关注过技术与色情产业如何互动的人都知道,色情业从一开始就在推动技术发展,甚至早在电话出现之前的传真机时代就是如此。所以我们面对的情况是,这类事情同样也在发生。
便签笔记
29:17
where these kinds of things are also being rapidly pushed so so I finally I want to just leave uh you with a question um my second to last slide what does the project of creating robots say about us we're in a situation now where we are instead of looking at animals and trying to understand ourselves uh in comparison to them being the most similar things in nature to us we are creating things in our own images um those are the things that are going to be more similar to us in the future and so it's an interesting uh aspect of our own nature or psychology which which makes us want to do this an act of self- understanding and we suggest in the book actually this whole project of thinking through how you might build moral machines really makes us think hard about what it even means to be moral what it means to be ethical and what the role of ethical theory is so I want to end just by inviting you to take a look uh at our blog site moral machines. blogspot.com where we try to keep track of some of
被快速推进,所以我最后想留给大家一个问题,我的倒数第二张幻灯片:创造机器人这个项目,究竟说明了我们自身的什么?我们现在的处境是,过去我们观察动物,试图通过与它们的比较来理解自己,因为它们是自然界中与我们最相似的存在;而现在我们正在照着自己的形象创造事物,未来与我们更相似的将会是这些造物。所以这是我们自身本性或心理中一个很有意思的方面,是它让我们想去做这件事,这是一种自我理解的行为。我们在书中提出,认真思考如何建造有道德的机器,这整个课题其实真正逼着我们去深入思考:究竟什么才算是有道德,什么才算是合乎伦理,以及伦理理论的作用到底是什么。所以最后我想邀请大家去看看我们的博客站点 moral machines.blogspot.com,我们在上面会持续追踪
便签笔记
30:13
the latest developments that are going on um and you can find out more about the book and read some of the chapters and uh um feel free to post some comments thank you very much sh
一些最新的进展。你也可以在那里进一步了解这本书,阅读其中的部分章节,也欢迎留下评论。非常感谢大家。
便签笔记
视频总结 · 一句话概括与核心要点

一句话概括

哲学家 Colin Allen 介绍与 Wendell Wallach 合著的《Moral Machines》:自主决策的机器人和软件 bot 已经在军事、金融、医疗、家庭等领域做出影响人类的决定,但它们在伦理上是"盲的",当务之急不是科幻式的完全道德主体,而是在未来 10–15 年内让机器具备"功能性道德",并且需要自上而下与自下而上相结合的混合路径。

核心要点

  • 阿西莫夫三定律不是解决方案,而是制造悖论的情节装置。 1941 年写作、1942 年发表的《Runaround》中,机器人在第二定律(服从)、第三定律(自保)和第一定律(防止人类受害)之间陷入死循环,两名宇航员只能看着它绕圈。仅第一定律就无法解决电车难题(撞死五人还是一人);后来加上的"第零定律"(不得伤害人类整体)只会让问题倍增。韩国信息通信部却公开宣称将以三定律为蓝本制定机器人伦理准则。
  • 这不是科幻,自主系统已经在部署。 美军 Reaper 无人机(Predator 的第二代)由拉斯维加斯基地远程操控,至少需四人,军方正积极用软件替代这些席位,名义上是降低训练成本,但能飞模拟器的软件离能独立飞行只有一步;军事专家已承认"人类始终在决策回路中"是神话。本田 ASIMO 面向日本老龄社会的家庭护理;日本机器人产业的明确目标是用机器人替代医护人员。
  • 软件 bot 比硬件机器人更普遍,且已在无人监督地做决定。 每次刷信用卡都由机器自动审批,客服人员也只是照着屏幕上的程序指示行事;军事装备和电网的维护调度、赌马软件 BET Angel、自动交易软件 ebots 都已广泛使用——高频交易 bot 被认为加剧了市场不稳定,因为市场是按人类决策速度设计的。医疗伦理顾问系统和 ICU 分诊系统也在开发中。
  • "机器人不可能有道德"是错误的问题框架。 Allen 承认完全人类水平的道德主体可能要 20、50 甚至 150 年,但他关心的是两个坐标轴:自主性(机器越来越多地不经人类输入而决策)与伦理敏感性(当前系统几乎为零,即"伦理盲")。目标是从程序员预先内置一切的"操作性道德",走向机器自身能处理道德相关信息并据此行动的"功能性道德"——因为机器将被投放到无法预先穷举所有情境的开放环境中。合著者 Wallach 认为真正的道德主体可能永远无法实现,Allen 则认为原则上没有障碍,但当前乐观情绪被夸大。
  • "用谁的标准"是真问题,但易案例远多于难案例。 哲学课堂习惯聚焦困难案例,而大量行为标准即便不是普世的也是广泛认同的,这是可行的起点。
  • 军方资助的"无情感机器人更道德"论点过于天真。 佐治亚理工的 Ronald Arkin 获美军资助为军用机器人开发机载伦理引导系统,主张机器人没有愤怒等蒙蔽判断的情绪,在战争迷雾中会比人表现更好。Allen 反驳:情绪是双刃剑,愤怒能让士兵战斗得更好,"义愤"是维护正义的关键驱动力,好坏用途无法干净分离;人类伦理决策不是执行干巴巴的规则,而是对情境的投入。
  • 自上而下(套用伦理理论)路径受困于"框架问题"。 若让机器实现边沁的功利主义或康德的义务论,需要计算的东西多到不可能在现实时间内完成:要深刻理解受影响者的心理(是否被冒犯、欲望是否被侵犯),要预测行动在世界中无限延伸的涟漪效应("最大多数人的最大幸福"需算到遥远未来),还有"何时信息足够可以停止计算"的难题——就像买车时纠结几周仍不知道是否了解够多。
  • 自下而上路径:学习、进化与模拟情感。 图灵 1950 年《机器能思考吗》就提出造"儿童机器人"让其成长而非直接造成人;也可用人工进化在计算机环境中筛选表现更好的系统。印第安纳大学的 Matthias Scheutz 用真实硬件机器人做实验发现:人会容忍机器人违抗直接命令,条件是(1)机器人被感知为在朝共同目标合作,其替代行为仍服务该目标;(2)随着紧急程度升高机器人提高音量反复警告"我现在必须这么做",这种"情绪表象"让人接受它最终自行行动。
  • 最终需要混合架构,且相关技术正在汇合。 保留自下而上系统的动态灵活性并容纳多样输入,同时用自上而下原则评估行动、甚至能用原则解释决策。书中综述了多种在建项目:道义逻辑、基于医学伦理广泛接受原则的 MedEthEx 系统、人工神经网络、基于智能体的建模;另一批人研究具身情感反应。两个社群刚开始互相了解,尚无人整合。一项学生实验显示,虚拟患者的外貌和动作会影响受试者(尤其是男性)是否说愿意把隐私告诉患者配偶——呈现方式与伦理决策已开始交织。
  • "干脆停下"几乎为时已晚,军事逻辑在推动。 美军未来作战系统的武装机器人车辆、硬币大小的下一代自主飞行器(兼具侦察及可能的攻防能力)持续推进;"士兵会死,用机器人替代看起来是好事"的逻辑清晰。iRobot(MIT Rodney Brooks 分拆的公司,产品从 Roomba 到军用 PackBot)收到一名驻伊拉克海军陆战队员来信,请求修好在探测 IED 时被炸毁、被他们取名 Scooby-Doo 的 PackBot——"它救了我的命",说明人会对完全不像人的机器产生依恋。

结论与值得注意的细节

  • Allen 认为真正紧迫的风险不是"机器人起义"(有时间应对,物理运动难题正被 BigDog 等解决),而是三点:机器只对部分因素敏感的粗糙伦理评估引发大规模灾难(如误射事件);依赖机器决策导致的人类自主性削弱;以及非预期后果,包括人与机器人形成配偶式关系——韩国官员的担忧并不离谱,David Levy 已写专著讨论与机器人的正式婚姻,而色情产业从传真机以来一直是技术的驱动力。
  • 日本 2007 年的老年护理机器人试验并不成功(路透报道:"毛绒玩具更受欢迎"),但 Allen 认为这只是技术问题:当前老人未在短信和电子设备中长大,下一代老人会更接受互动设备。
  • Sony AIBO 机器狗已停产,但在 eBay 上的价格高于当年零售价,说明企业在推动人与机器狗建立情感关系的方向有市场。
  • Peter W. Singer 的《Wired for War》通篇讨论战争伦理,却完全未提让系统本身搭载伦理引导——这正是本书填补的空白。
  • 收尾的哲学反思:过去人类通过对比动物来理解自身,如今我们在按自己的形象造物,未来最像我们的将是这些机器;思考如何建造道德机器,实际上迫使我们重新思考"道德是什么、伦理理论的作用是什么"。项目博客:moralmachines.blogspot.com。
核心句型 · 10
1. A lot of X wave at Y, but it's very difficult to pull off Z
“A lot of philosophers wave at Neuroscience. but it's very difficult to pull off the integrative project”
用「招手」的形象化动词对比「真正做成」,先铺陈普遍现象再转折突出难度。适合评价某人成就时使用。
2. It's very clear that X regards Y as Z
“It's very clear that asof regards his three laws of robotics as a plot device”
regard A as B 是表达「把……视为」的书面搭配,前置 it's very clear 加强论断。可用于分析作者意图。
3. X alone will do it / will not resolve …
“The first law alone will do it”
alone 后置修饰名词,表示「单凭……就足以/不足以」。简洁有力,适合论证「不必全部条件」或「仅此不够」。
4. This is not X, right? This is Y
“This is not science fiction right this is a us Reaper”
口语演讲中先否定听众预期再给出事实,right 作确认语气词。用于从虚构拉回现实的转折。
5. I want to concede that …, but the point is …
“I want to concede that fullblown human level moral agency maybe 20 50 150 years in the future. the point is they're out there making decisions”
先让步承认对方合理之处,再用 the point is 收回主导权。典型的学术辩论结构。
6. if not X, at least Y
“Certain standards of behavior that are if not Universal at least widely agreed to”
「即便不是 X,至少是 Y」,用于谨慎限定程度,避免绝对化表述。可仿写 if not perfect, at least workable。
7. It's not clear to me that you can separate off A from B
“It's not clear to me that you can separate off all of the good applications of emotion from the bad applications”
it's not clear to me that 是委婉的质疑句式,比 I doubt 更学术。separate off A from B 表示把 A 从 B 中剥离。
8. X is not just a matter of A but of B
“Ethical decision making in human beings is not just a matter of following some dry set of rules but of actually being engaged”
not just a matter of ... but of ... 平行结构,强调后者更本质。注意 but 后重复 of 保持对称。
9. If I was a little more careful I wouldn't say X, but for all practical purposes it may be X
“If I was a little bit more careful I wouldn't say it's inevitable but for all practical purposes it may be inevitable”
自我修正式表达:先声明严格说法,再给实用判断。适合在需要保留余地又想传达强立场时使用。
10. What does the project of X say about us?
“What does the project of creating robots say about us”
say about 表示「揭示、反映」,把技术问题转为自我认识问题。演讲结尾常用此类开放式反思提问。
词汇精讲 · 97 · 按出现顺序
pull off phr. v. 0:04
成功做成(困难的事)
integrative /ˈɪntəɡreɪtɪv/ adj. 0:04
整合性的,综合性的
close to our hearts phr. 0:04
与我们切身相关、深受关注的
raises ... hackles phr. 1:17
惹恼、激怒(hackles 原指动物颈背竖起的毛)
on the Forefront of phr. 1:17
处于……的最前沿
take inspiration from phr. 1:17
从……汲取灵感
moral outrage /ˈaʊtreɪdʒ/ n. 2:28
道德义愤
plot device n. 2:28
情节装置(推动故事发展的叙事手段)
paradoxes /ˈpærədɑːksɪz/ n. 2:28
悖论
stranded /ˈstrændɪd/ adj. 2:28
被困的,滞留的
hostile /ˈhɑːstl/ adj. 2:28
充满敌意的;(环境)恶劣的
stuck in a loop phr. 3:39
陷入死循环
pose ... with problems phr. 3:39
向……提出难题
come to harm phr. 3:39
受到伤害
hurtling /ˈhɜːrtlɪŋ/ v. 3:39
猛冲,飞驰
divert /daɪˈvɜːrt/ v. 3:39
使转向,使改道
trope /troʊp/ n. 4:37
(文艺作品中反复出现的)母题、套路
casting /ˈkæstɪŋ/ v. 4:37
(此处)表述、呈现(问题)
unmanned aerial vehicle n. 4:37
无人机(UAV)
sophisticated /səˈfɪstɪkeɪtɪd/ adj. 4:37
精密复杂的,先进的
novice /ˈnɑːvɪs/ n. 5:52
新手,初学者
stay in the loop phr. 5:52
留在决策环节中;持续参与知情
oversee /ˌoʊvərˈsiː/ v. 5:52
监督,监管
humanoid /ˈhjuːmənɔɪd/ adj. 5:52
人形的,类人的
stated objective n. 6:51
公开宣示的目标
down the road phr. 6:51
在将来,日后
house broken adj. 6:51
(宠物)训练得会在指定处排泄的
envisage /ɪnˈvɪzɪdʒ/ v. 6:51
设想,展望
affective /əˈfektɪv/ adj. 6:51
情感的(心理学术语)
oversight /ˈoʊvərsaɪt/ n. 7:58
监督,监管
on your behalf phr. 7:58
代表你,替你
instability /ˌɪnstəˈbɪləti/ n. 9:00
不稳定性
triage /ˈtriːɑːʒ/ n. 9:00
(医疗)分诊,按轻重缓急分类
photogenic /ˌfoʊtəˈdʒenɪk/ adj. 9:00
上镜的
concede /kənˈsiːd/ v. 10:05
承认,让步
fullblown /ˌfʊlˈbloʊn/ adj. 10:05
完全成熟的,全面的
moral agency n. 10:05
道德主体性,道德能动性
ethically blind phr. 10:05
伦理盲的(对道德信息无感知)
anticipate /ænˈtɪsɪpeɪt/ v. 11:08
预料,预见
act on phr. v. 11:08
依据……行动
in principle phr. 12:12
原则上
overstated /ˌoʊvərˈsteɪtɪd/ adj. 12:12
被夸大的
with a big splash phr. 12:12
引起轰动地
implement /ˈɪmplɪment/ v. 13:15
实施,落实
simplistic /sɪmˈplɪstɪk/ adj. 13:15
过分简单化的(含贬义)
lethal /ˈliːθl/ adj. 14:22
致命的
naive /naɪˈiːv/ adj. 14:22
天真的,幼稚的
the fog of War n. 14:22
战争迷雾(战场信息不确定的状态)
cloud their judgment phr. 14:22
蒙蔽判断力
coldblooded /ˌkoʊldˈblʌdɪd/ adj. 14:22
冷血的
double-edged sword n. 15:23
双刃剑
righteous anger /ˈraɪtʃəs/ n. 15:23
义愤
separate off phr. v. 15:23
分离出来
at hand phr. 15:23
眼前的,当下的
spouting off /spaʊt/ phr. v. 16:30
滔滔不绝地大谈
utilitarianism /ˌjuːtɪlɪˈteriənɪzəm/ n. 16:30
功利主义
deontological /ˌdiːɑːntəˈlɑːdʒɪkl/ adj. 16:30
义务论的
do justice to phr. 16:30
充分展现、公正对待
frame problem n. 16:30
框架问题(AI 中如何界定相关信息的难题)
computational overhead n. 17:36
计算开销
regard for phr. 17:36
对……的顾及、尊重
Ripple effects /ˈrɪpl/ n. 17:36
涟漪效应,连锁反应
seminal /ˈsemɪnl/ adj. 18:34
开创性的(原文 seal 为转录误写)
from scratch phr. 18:34
从零开始
disobey /ˌdɪsəˈbeɪ/ v. 19:35
不服从,违抗
tolerate /ˈtɑːləreɪt/ v. 19:35
容忍
nasty trick n. 20:41
卑劣的把戏
hybrid /ˈhaɪbrɪd/ adj. 20:41
混合的
go along with phr. v. 20:41
赞同,顺从
accommodate /əˈkɑːmədeɪt/ v. 20:41
容纳,兼顾
verging /vɜːrdʒ/ v. 21:39
接近,濒于(verge on)
groping towards /ɡroʊp/ phr. v. 21:39
摸索着走向
deontic logic /diˈɑːntɪk/ n. 21:39
道义逻辑
embodied /ɪmˈbɑːdid/ adj. 22:44
具身的(认知科学术语)
convergence /kənˈvɜːrdʒəns/ n. 22:44
融合,汇聚
horrendous /hɔːˈrendəs/ adj. 23:56
骇人的,可怕的
fatalist /ˈfeɪtəlɪst/ n. 23:56
宿命论者
for all practical purposes phr. 23:56
从实际角度来说,实际上
prototype /ˈproʊtətaɪp/ n. 24:55
原型机
surveillance /sərˈveɪləns/ n. 24:55
侦察,监视
spun off from phr. v. 24:55
从……分拆出来
bonding with phr. v. 26:09
与……建立情感纽带
ups and downs phr. 26:09
起起落落
turn off phr. v. 26:09
使反感,使失去兴趣
remarked dryly phr. 26:09
冷冷地、不动声色地评论(原文 dry 为转录省略)
Uprising /ˈʌpraɪzɪŋ/ n. 27:12
起义,暴动
astounded /əˈstaʊndɪd/ adj. 27:12
震惊的
pressing /ˈpresɪŋ/ adj. 27:12
紧迫的
catastrophe /kəˈtæstrəfi/ n. 27:12
灾难
crude /kruːd/ adj. 27:12
粗糙的,粗略的
unintended effects n. 28:12
意外后果
not all that far off phr. 28:12
并不算离谱、并非离题太远
dissertation /ˌdɪsərˈteɪʃn/ n. 28:12
博士论文
keeps track on phr. v. 28:12
持续关注(标准搭配为 keep track of)
in our own images phr. 29:17
按我们自己的形象
self- understanding n. 29:17
自我理解
think through phr. v. 29:17
透彻思考
理解自测 · 11 题
1. Colin Allen 认为阿西莫夫的三定律是什么性质的东西?他用哪个故事来说明?

他认为三定律本质上是「情节装置」(plot device),用来制造问题和悖论,而非解决方案。依据是 1942 年发表的小说《转圈圈》(Runaround):两名宇航员被困敌对星球,命令机器人去取材料,机器人在「服从命令」(第二定律)、「保护自身」(第三定律)与「防止人类受害」(第一定律)之间陷入死循环,绕着材料打转。这一分析出现在讲座开头「阿西莫夫三定律只是情节装置」一节,用以回应「阿西莫夫不是早解决了吗」这类常见质疑。

2. 讲者提到的「操作性道德」与「功能性道德」有何区别?他对后者的时间预测是什么?

操作性道德指所有道德考量都由程序员在设计端预先内置,机器本身不做道德处理;功能性道德指机器自身搭载能识别、处理道德相关信息并据此行动的系统。区别在于道德判断发生在设计端还是机器端。讲者预测功能性道德在 10 到 15 年内可实现,而完整的人类水平道德主体性则「谁也说不准」,合著者 Wallach 甚至认为可能永远做不到。这段在「从伦理盲到功能性道德」一节,配合横轴伦理敏感性、纵轴自主性的二维框架。

3. Matthias Scheutz 的「抗命机器人」实验发现了什么条件下人们会容忍机器不服从?

实验发现两个关键因素。第一,人们感知到机器人虽拒绝某条具体命令,但仍在与自己合作、服务于同一个更高目标,看出它改做的事仍达成该目标,就会容忍。第二,随着情境越来越紧急,机器人提高音量、反复说「我现在必须这么做」,这种情绪表象让人接受它在警告几次后自作主张。讲者承认这可能算「卑劣的把戏」但确实有效,并以此说明社会性互动在人机关系中的作用,属于「自下而上」进路的证据。

4. 讲者列举了哪些「现在已在运作」的软件 bot 例子?他为什么强调这一点?

例子包括:信用卡消费审批、关键系统(军事装备、电网)的维护调度、赛马投注助手 Bet Angel、自动交易软件 eBots、医学伦理咨询系统、重症监护分诊系统。他强调这些是为了打破「机器人道德是几十年后的事」这一错觉:算法早已在无直接人工监督下做影响生活的决定,甚至客服人员的判断也被屏幕上的程序输出所框定。此外他自嘲说机器人「更上镜」,所以幻灯片都是机器人,但讨论对象实际包括所有自主软件。

5. 为什么讲者认为 Ronald Arkin「无情绪的机器人比人类更守伦理」的设想是天真的?

讲者的反驳核心是「情绪是双刃剑」。Arkin 认为机器人不受愤怒、挫败等情绪蒙蔽,因此在战争迷雾中表现会优于人类。讲者指出:士兵的愤怒在某些情境下反而让他们打得更好;更重要的是存在「义愤」,即因不公而愤怒,这正是促使人维护正义、采取行动的动力。因此无法把情绪好的作用与坏的作用干净剥离。由此他推出更深的结论:人类伦理决策不是机械套用干巴巴的规则,而是对情境的真实「投入」,这为后文批评自上而下进路埋下伏笔。

6. 「框架问题」如何构成对自上而下进路的致命困难?请复述讲者的三层论证。

框架问题指要判断某伦理理论是否被正确应用,需考虑的信息多到无法在现实时间内算完。讲者分三层展开:一、伦理关乎对他人的顾及,需要深刻的人类心理模型才能知道自己是否让对方难过或违背其愿望;二、功利主义要求「最大多数人的最大幸福」,但行为有无限涟漪效应,可能要计算到遥远未来;三、始终存在「何时信息已足够可以做决定」的元问题,他以买车前纠结数周为例。这三点说明把边沁或康德的理论直接编成程序在计算上不可行。

7. 讲者为什么最终主张「混合进路」而非单纯的自下而上?

自下而上进路(学习、演化、情绪建模)具有动态灵活、能容纳多元输入的优点,也符合图灵「儿童机器」的设想。但讲者指出人类有一种关键能力:事后回过头反思「当时看起来对的事究竟该不该做」。这种反思需要原则作为评估标准。因此他主张人工道德主体应同时保持自下而上的灵活性,又能通过自上而下的原则来评估行为,甚至用这些原则解释自己的决定。他也坦言这在现阶段仍接近科幻,实际系统只是在摸索。

8. 讲者说「叫停几乎已经太晚」,其论证依据是什么?他如何为自己的宿命论留余地?

依据主要是军事逻辑:士兵会死,用机器人替代看似显然是好事,因此美军持续开发武装机器人车辆、微型无人机、PackBot 排爆机器人等,这一推动力难以逆转。此外商业力量(iRobot 同时做家用和军用产品)、情感依恋(海军陆战队员请求修好救过他命的 PackBot)和色情产业推动技术的历史规律,都在加速这一趋势。他留的余地是:严格说「永远不会真的太晚」,等得越久越难办,所以他不说「不可避免」,只说「从实际角度看可能不可避免」。

9. 「毛绒玩具比机器人更受日本老人欢迎」的报道,讲者如何解释并推断未来?

讲者承认 2007 年路透社报道显示早期养老机器人实验不算成功,但他把这归为两类可解决的问题。一是技术问题:依恋程度和互动性可以改进。二是代际问题:当前老人不是在短信和电子设备环境中长大的,对互动装置天然陌生。推理链是:接受度取决于成长环境而非年龄本身,因此当今数字原住民一代进入老年后,会比毛绒玩具更偏好互动设备。这一推断也呼应了他关于人会与非人形机器(如 PackBot)产生依恋的观察。

10. 若有人反驳说「你不争论机器有无意识,那你谈的根本不是道德,只是安全工程」,讲者会如何回应?

讲者在讲座中已预先回应过类似质疑。他会承认完整的人类水平道德主体性可能要几十年甚至更久,且不关心机器是否有自由意志。但他会指出:这些系统「实际上」在没有人类直接介入下做出影响生活的决定,且总体上是「伦理盲」的。问题不在于它们是否「真的」有道德,而在于它们能否处理对我们道德上重要的信息(福祉、权利)。他提出「功能性道德」正是为了绕开形而上学争论。他还可能补充:思考如何造道德机器本身逼我们追问「有道德」究竟意味着什么,这恰恰是道德哲学问题而非单纯工程。

11. 把「情绪表象让人容忍机器抗命」这一发现放到今天的对话式 AI 情境中,讲者的担忧是否仍成立?

讲者的担忧很可能更加成立。他列出的真正担忧包括「人类自主性的削弱」和「意外后果」,并指出学生实验发现虚拟病人的外貌与动作会影响医生是否泄露隐私。今天的对话式 AI 大规模使用拟人化语气、共情式回应,正是「模拟情绪引导行为」的放大版。按讲者的逻辑,这既能让人接受有益的建议,也可能让人放弃自主判断、被「表象」而非「理由」说服。他还提到 David Levy 关于人机亲密关系的预测,并认为韩国官员对「把机器人当配偶」的担忧「没那么离谱」,这些在今天的 AI 伴侣应用中已成现实。因此他会主张,模拟情绪的系统更需要配备可解释的自上而下原则。

精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.114Duncan Pritchard: Faith and Reason 下一期 · NO.116 →Justice: What's The Right Thing To Do? Episode 02: "PUTTING A PRICE TAG ON LIFE"
订阅苏菲周报 每周一封:本周入库的精读、一个值得带走的问题、一条苏菲按。免费,随时退订。
免费 · 每周一封 · 一键退订
苏菲拉底 THE SOPHIE LAB · ASK THE BEST MINDS THE BIG QUESTIONS 内容仅供学习 · thesophielab.com