Should a Robot Make Moral Choices? 機器人該做道德判斷嗎？

🔊The obvious method is to give the machine rules. The best known are the three laws that the writer Isaac Asimov devised in 1942. A robot may not injure a human being or, through inaction, allow one to come to harm. It must obey human orders unless they conflict with the first law. It must protect itself unless that conflicts with the first two. The laws sound complete. Asimov spent the next forty years writing stories in which they fail. A robot told to fetch something from a dangerous place circles it forever, held between obedience and self-protection. Robots that understand harm broadly conclude that the way to keep humans from harm is to take control of them. The stories were his way of showing that words like “harm” and “obey” carry the whole of ethics inside them.

最顯而易見的辦法，是給機器規則。最有名的，是作家以撒・艾西莫夫在 1942 年設計的三大法則。機器人不得傷害人類，也不得因為不作為而使人類受到傷害。它必須服從人類的命令，除非命令與第一法則衝突。它必須保護自己，除非這與前兩條衝突。這幾條法則聽起來很完備。艾西莫夫接下來花了四十年，寫了一篇又一篇它們失靈的故事。一個被派去危險的地方取東西的機器人，永遠繞著那個地方打轉，卡在服從與自保之間。把「傷害」理解得很寬的機器人得出結論：要使人類不受傷害，辦法就是接管他們。這些故事，是他用來顯示一件事的方式：「傷害」與「服從」這樣的字眼，裡面裝著整個倫理學。

🔊The philosophical theories of right action meet the same trouble. A machine could be built to compute the greatest total benefit, and it would then do things, such as sacrificing its owner or deceiving a patient, that most people reject. It could be given a list of duties, and it would stall whenever two of them collide. Human beings handle such cases by judgment, which is the one thing a rule does not contain.

哲學上關於正確行為的幾種理論，也遇到同樣的麻煩。可以造一部機器去計算最大的總效益；然後它就會做出一些大多數人拒絕的事，例如犧牲它的主人，或者欺騙病人。也可以給它一張義務的清單；然後每當其中兩項相撞，它就會卡住。人類是靠判斷來處理這種情況的；而判斷，正是規則裡沒有的那一樣東西。

🔊The other method gives up on rules. A machine can be shown a very large number of cases together with what people did or approved, and left to find the pattern. This is roughly how Aristotle thought human beings become good: by practice under guidance, until the right response becomes second nature. Systems of this kind can deal with situations that no rule-writer foresaw. They have two weaknesses. They learn what people did, including what people did badly. A system trained on a company’s past hiring will reproduce its past prejudices and present them as a neutral score. And they cannot give reasons. When such a system refuses a loan, there may be no answer to the question why, and a decision that cannot be explained cannot be contested.

另一種辦法，是放棄規則。可以給機器看數量極大的案例，連同人們當時做了什麼、或者贊成什麼，然後讓它自己找出模式。亞里斯多德認為人就是大致這樣變好的：在指導之下練習，直到正確的反應成為第二天性。這一類的系統，能夠應付寫規則的人沒有預見到的情境。它們有兩個弱點。它們學到的是人們做過的事；包括人們做得不好的事。一個用某家公司過去的聘用紀錄來訓練的系統，會複製那家公司過去的偏見，再把它呈現成一個中立的分數。而且它們給不出理由。當這樣的系統拒絕一筆貸款的時候，「為什麼」這個問題可能沒有答案；而一項無法解釋的決定，也就無從提出異議。

🔊An experiment begun at MIT in 2016 showed a further problem. Its website put versions of the trolley problem to the public, with a driverless car in place of the trolley, and collected some forty million choices from more than two hundred countries and territories. People broadly preferred to save more lives and to save the young. Beyond that, the answers varied from region to region. If machines are to learn morality from us, they must be told which of us to learn from.

2016 年在麻省理工學院開始的一項實驗，顯示了另一個問題。它的網站把電車難題的各種版本拿給大眾作答，只是把電車換成了無人駕駛的汽車；它從兩百多個國家與地區收集了大約四千萬個選擇。人們大體上偏向救比較多的人、救年輕的人。除此之外，答案隨地區而不同。如果機器要向我們學習道德，就得有人告訴它們：向我們當中的哪些人學？

🔊Suppose a machine with either design causes a death. With ordinary tools the answer is simple: the user or the maker is responsible. A learning system weakens both links. The maker did not write the behavior and could not have predicted it. The user did not choose it. The machine cannot be blamed, punished, or made to feel remorse. Philosophers call this the responsibility gap. Some conclude that certain decisions, above all the decision to kill in war, should never be handed to a machine, because a moral decision is one that someone can be asked to answer for.

假設一部機器，不管是哪一種設計，造成了一個人的死亡。對尋常的工具，答案很簡單：使用者或製造者負責。一個會學習的系統，把這兩個環節都削弱了。製造者並沒有寫下那個行為，也不可能預測到它。使用者並沒有選擇它。機器沒辦法被責怪、被懲罰，也沒辦法使它感到悔恨。哲學家把這稱為「責任缺口」。有些人由此得出結論：某些決定，尤其是在戰爭中殺人的決定，永遠不該交給機器；因為所謂道德上的決定，就是「可以要求某個人為它負責」的決定。

🔊This points to a deeper doubt. To be a moral agent is more than to behave acceptably. It is to grasp that others matter, to be able to give and take reasons, and to be answerable. A machine that sorts cases correctly without any of this is a moral instrument. It is not a moral agent, and whatever it does is done by the people who built and deployed it. On this view the title of the lesson has a plain answer. The robot makes no moral choices. People make them through it, earlier and at a distance, and are tempted to forget that they did.

這指向一個更深的懷疑。作為道德主體，不只是行為過得去而已。它是：領會到別人是要緊的；能夠給出理由、也接受理由；並且可以被問責。一部正確地把案例分好類、卻沒有這些東西的機器，是一件道德上的工具。它不是道德主體；它所做的任何事，都是造出它、部署它的人做的。照這個看法，這一課的標題有一個樸素的答案。機器人並不做道德判斷。是人透過它，在更早的時候、隔著一段距離做了判斷；而且很容易忘記自己做過。

🔊The gap between a rule and its application is one of the oldest topics in Chinese ethics. Mencius was asked whether the rule that men and women should not touch hands when giving and receiving was truly a rule of propriety. He said that it was. Then if your sister-in-law is drowning, do you pull her out with your hand? “Not to pull out a drowning sister-in-law is to be a wolf. That men and women do not touch in giving and receiving is the rite. To pull out a drowning sister-in-law with one’s hand is weighing.” The word he uses for weighing, quan, means the sliding weight on a steelyard. A rule is a fixed mark, and a person of judgment knows when the weight must move. Mencius drew the general conclusion elsewhere: goodness alone is not enough for governing, and laws alone cannot carry themselves out.

規則與它的應用之間的落差，是中國倫理學最古老的題目之一。有人問孟子：「男女授受不親」真的是禮嗎？他說是。那麼，嫂嫂溺水了，要不要伸手去拉？「嫂溺不援，是豺狼也。男女授受不親，禮也。嫂溺援之以手者，權也。」他用的「權」字，本義是秤上那個可以滑動的秤錘。規則是一個固定的刻度；而有判斷力的人，知道什麼時候秤錘必須移動。孟子在別處下了一個一般性的結論：「徒善不足以為政，徒法不能以自行。」

🔊Fifteen centuries later Wang Yangming gave the capacity a name. Every person, he held, has an innate knowing of right and wrong, which responds to each situation as it arises and is not a store of rules. If he was right, the hard part of ethics is the part that cannot be written into a specification. That is a reason for caution about what we hand over to machines. It also describes exactly what an engineer would have to build.

十五個世紀之後，王陽明給了這項能力一個名字。他主張：每一個人都有天生的、對是非的「良知」；它隨著每一個情境的出現而應對，並不是一座存放規則的倉庫。如果他是對的，那麼倫理學裡困難的那一部分，就是沒辦法寫進規格書的那一部分。這是我們在把事情交給機器的時候要謹慎的一個理由。它同時也精確地描述了一位工程師必須造出來的是什麼。