AI-based Google Translate make me very worried about mankind’s AI-dominated near future. After observing some weird GT translations, I decided to do some simple experiments with common words derived from the Oxford English Corpus: (Most common words in English)[https://en.wikipedia.org/wiki/Most_common_words_in_English].
I requested a Chinese translation of the most common 25 nouns: time, person, year, way, day, thing, man, world, life, hand, part, child, eye, woman, place, work, week, case, point, government, company, number, group, problem, fact and immediately received an ordered list of perfectly acceptable equivalents: 時間，人，年，方法，一天，事，人，世界，生活，手，部分，孩子，眼睛，女人，地方，工作，週，案件，點，政府，公司，數量，組，問題，事實. So far, so good.
When I modified the list by inserting one single extraneous character before the first word time (@time, utime), however, the results changed dramatically. The word person became two words in Chinese:
人 and 人物 “character (in a story), (political, historical) figure, personality etc.”, and man, originally translated as 人 “person” becomes 男人 “a male person”. When I changed time to utime, the results for the remaining 24 words became even weirder: time was translated as 美國人 “an American” and the word person was left blank. I’ll let the reader look at the overall picture:
If interested, you can repeat my experiments with Google Translate and add your own quirky variations. Please do share any interesting results!
“What’s the problem,” you may ask. “So what if English-Chinese translations are sometimes a bit off?”
The problem is that
(1) Humans are not fully predictable. We don’t always follow previous patterns and do what is reasonable.
(2) When AI systems are confronted by unexpected human behavior (such as when I deliberately inserted an extraneous character in a plain list of words), the results can go off the map.
This minor translation problem is but a tiny symptom, a harbinger of a very troubling future. If AI systems are allowed to manage human affairs, we are almost guaranteed to be unhappy with the results. Do we really want our roads to be taken over by AI-controlled self-driving cars that communicate with each other in ways that mere humans cannot decipher?
Even Elon Musk, who seemingly has everything to gain by being allowed to proceed unfettered with his development work, has been sounding the alarm and asking for AI to be regulated before it's too late:
If you want to watch the whole thing, here is Elon Musk's discussion at the 2017 NGA 2017 Summer Meeting (video starts at the 25:30 mark, after the introductory remarks are over):
Note: These experiments were inspired by Creating poetry with Google translate
First Image Source: Public Domain image from Pixabay (with my own text)
新版Google Translate改用人工智能好讓翻譯效果變得更精準，但是經常出現的小錯誤讓我感到分不安，我因此決定做一些簡單的實驗，材料是自牛津英語語料庫的常見英文單字：（最常用的英文單詞）[https://en.wikipedia.org/wiki/Most_common_words_in_English]。
我輸入最常用的25個英文單字 （time, person, year, way, day, thing, man, world, life, hand, part, child, eye, woman, place, work, week, case, point, government, company, number, group, problem, fact ），初步結果沒什麼稀奇，全部翻譯都很正確：時間，人，年，方法，一天，事，人，世界，生活，手，部分，孩子，眼睛，女人，地方，工作，週，案件，點，政府，公司，數量，組，問題，事實。
問題是，當我在 time 這個單字前面多加一個字符 (@time || utime)，很多翻譯都亂掉了（必須去掉：其他24個單字完全沒有做任何更動）。最奇怪的是“utime”這個不存在的英文單字竟然“翻譯”成“美國人”，而且下一個單字忽略掉，交白卷。下面可以看全部結果：
有興趣的讀者不妨自己也玩一下，看看Google Translate 還會寫出什麼寶貝『翻譯』！有趣的結果請一定要跟大家分享哦！
熟悉電腦翻譯的朋友們可能會覺得這種事沒什麼稀奇，翻錯又怎麼樣啊？
問題是這樣的：
（1）人類的行為並不能完全預測，有時候會做出非常不合理的事情，但總會有人想出辦法解決問題。
（2）AI 人工智慧可是不同，讓AI 來控制我們賴以為生的各種錯綜複雜的生產、醫療等系統，一旦出了問題便一發不可收拾。
應該趁我們還來得及懸崖勒馬，好好想一想到底要不要讓AI無條件地接管我們的生活？