Let me share a paradox. For the last 10 years, many companies have been trying to become less bureaucratic, to have fewer central rules and procedures, more autonomy for their local teams to be more agile. And now they are pushing artificial intelligence, AI, unaware that cool technology might make them more bureaucratic than ever. Why? Because AI operates just like bureaucracies.
我來和大家分享發現的一個悖論。 在過去十年裡, 很多公司都想擺脫官僚化, 通過減少職務、精簡程序, 給團隊更多自主, 讓公司運作更靈活。 現在,公司開始引進人工智慧,AI, 卻沒意識到這個很酷的科技, 可能讓他們比以前更官僚。 為什麼呢? 因為 AI 的運作方式就很官僚。
The essence of bureaucracy is to favor rules and procedures over human judgment. And AI decides solely based on rules. Many rules inferred from past data but only rules. And if human judgment is not kept in the loop, AI will bring a terrifying form of new bureaucracy -- I call it "algocracy" -- where AI will take more and more critical decisions by the rules outside of any human control. Is there a real risk? Yes.
官僚的本質 就是看重規則和程序, 而不是人類自身的判斷, 而 AI 僅依據規則做決策。 即便 AI 的規則 依據過去的數據形成, 那也終究是規則。 如果將人類判斷置若罔聞, 對 AI 的運用將帶來 可怕的新官僚主義—— 我稱之為「AI 官僚主義」 (algocracy), 意即 AI 將根據規則, 不受人為控制, 做更多重要決策。 真有風險嗎? 當然有。
I'm leading a team of 800 AI specialists. We have deployed over 100 customized AI solutions for large companies around the world. And I see too many corporate executives behaving like bureaucrats from the past. They want to take costly, old-fashioned humans out of the loop and rely only upon AI to take decisions. I call this the "human-zero mindset." And why is it so tempting? Because the other route, "Human plus AI," is long, costly and difficult. Business teams, tech teams, data-science teams have to iterate for months to craft exactly how humans and AI can best work together. Long, costly and difficult. But the reward is huge.
我領導的團隊 由 800 位 AI 專家組成, 我們為很多全球的大公司 量身打造了超過 100 個 AI 系統。 我看過太多的公司高管 因此而變得官僚作風。 他們對麻煩、老舊的 人類決策嗤之以鼻, 完全依賴 AI 來做決策。 我稱之為「無人類思維」。 (human-zero mindset) 可為何這種思維這麼誘人? 因為另一種思維—— 「人類+AI」(Human plus AI) 費時、費錢又費力。 商業團隊、科技團隊和數據科學團隊 不得不費幾個月的功夫 探索人類和 AI 怎樣達到最好的合作。 探索過程漫長艱難,耗了很多錢, 但取得了巨大成果。
A recent survey from BCG and MIT shows that 18 percent of companies in the world are pioneering AI, making money with it. Those companies focus 80 percent of their AI initiatives on effectiveness and growth, taking better decisions -- not replacing humans with AI to save costs.
根據波士頓諮詢公司 和麻省理工學院最近的調查, 全球有 18% 的公司 在研究 AI, 藉此盈利。 那些公司把 80% 的 AI 創新 專注在效率和成長, 做更好的決策, 而不是用 AI 取代人類來減少開支。
Why is it important to keep humans in the loop? Simply because, left alone, AI can do very dumb things. Sometimes with no consequences, like in this tweet. "Dear Amazon, I bought a toilet seat. Necessity, not desire. I do not collect them, I'm not a toilet-seat addict. No matter how temptingly you email me, I am not going to think, 'Oh, go on, then, one more toilet seat, I'll treat myself.' "
為什麼人類的作用必不可少? 原因很簡單: 沒有人類,AI 會幹傻事。 有時候毫無幫助, 就像這條推文講的: 「親愛的亞馬遜公司, 我之前買了一個馬桶圈。 生活必需品,不是什麼私人愛好。 我不收藏馬桶圈, 我沒有馬桶圈癮。 不管你的廣告郵件多誘人, 我都不會覺得 『噢,受不了, 只好再買個馬桶圈了, 偶爾放縱一下自己。』」
(Laughter)
(笑聲)
Sometimes, with more consequence, like in this other tweet. "Had the same situation with my mother's burial urn."
有時候,AI 又「太有幫助」, 就像這條推文: 「我在買我媽媽的骨灰盒後 遇到了同樣狀況。」
(Laughter)
(笑聲)
"For months after her death, I got messages from Amazon, saying, 'If you liked that ...' "
「在她去世後的幾個月, 亞馬遜發給我多個郵件: 『如果你喜歡這個⋯(骨灰盒)』」
(Laughter)
(笑聲)
Sometimes with worse consequences. Take an AI engine rejecting a student application for university. Why? Because it has "learned," on past data, characteristics of students that will pass and fail. Some are obvious, like GPAs. But if, in the past, all students from a given postal code have failed, it is very likely that AI will make this a rule and will reject every student with this postal code, not giving anyone the opportunity to prove the rule wrong.
有時 AI 幹壞事。 比如說 AI 曾經拒絕 學生的大學申請。 為什麼? 因為這個 AI 從以前的數據「學」了 哪些學生會通過,哪些學生不能—— 有一些指標很明顯,比如 GPA。 但如果在過去,某個地區 所有的學生都沒通過, AI 很可能就以此定下規則, 拒絕每一個來自這個地區的學生, 不給任何人證明規則有誤的機會。
And no one can check all the rules, because advanced AI is constantly learning. And if humans are kept out of the room, there comes the algocratic nightmare. Who is accountable for rejecting the student? No one, AI did. Is it fair? Yes. The same set of objective rules has been applied to everyone. Could we reconsider for this bright kid with the wrong postal code? No, algos don't change their mind.
沒有人能檢查每一條規則, 因為先進的 AI 一直在學。 所以如果直接用 AI 取代人類, 迎來的將是 AI 官僚主義的噩夢: 誰應該對學生的被拒負責? 沒有誰,AI 負責。 這公平嗎?公平。 因為所有學生都用同一規則判定。 那可不可以重新考慮這個 「住錯了地方」的聰明學生? 不行,AI 官僚不會改主意。
We have a choice here. Carry on with algocracy or decide to go to "Human plus AI." And to do this, we need to stop thinking tech first, and we need to start applying the secret formula. To deploy "Human plus AI," 10 percent of the effort is to code algos; 20 percent to build tech around the algos, collecting data, building UI, integrating into legacy systems; But 70 percent, the bulk of the effort, is about weaving together AI with people and processes to maximize real outcome.
我們的選項是, 繼續 AI 的獨裁, 還是考慮「人類+AI」思維? 要擁有這種思維, 我們要先把科技放在一邊, 從秘密公式入手。 要實現「人類+AI」, 需要 10% 的程式; 20% 的科技成份, 包括收集數據、構建用戶界面、 整合進遺留系統; 但 70%,最主要的部份, 是結合 AI 和人類方法, 讓結果最接近完美。
AI fails when cutting short on the 70 percent. The price tag for that can be small, wasting many, many millions of dollars on useless technology. Anyone cares? Or real tragedies: 346 casualties in the recent crashes of two B-737 aircrafts when pilots could not interact properly with a computerized command system.
如果這 70% 被削減, AI 就會出現問題。 代價可以很小, 比如在無用科技上浪費數百萬美元。 誰會在乎呢? 但代價也可以大到無法承受: 最近兩起波音 737 空難中 346 人遇難, 原因是電腦控制的飛行系統 無法正確回應飛機師的指令。
For a successful 70 percent, the first step is to make sure that algos are coded by data scientists and domain experts together. Take health care for example. One of our teams worked on a new drug with a slight problem. When taking their first dose, some patients, very few, have heart attacks. So, all patients, when taking their first dose, have to spend one day in hospital, for monitoring, just in case. Our objective was to identify patients who were at zero risk of heart attacks, who could skip the day in hospital. We used AI to analyze data from clinical trials, to correlate ECG signal, blood composition, biomarkers, with the risk of heart attack. In one month, our model could flag 62 percent of patients at zero risk. They could skip the day in hospital. Would you be comfortable staying at home for your first dose if the algo said so?
要實現人類的 70%, 第一步是保證算法程式由數據科學家 和領域專家共同完成。 拿醫療保健舉例, 我們有一個團隊 曾經處理一種藥產生的小問題。 在第一次服用這種藥後, 有一小部份的患者會發心臟病。 於是所有第一次服這種藥的患者 都要住院觀察一天, 以防心臟病發作。 我們想要識別出 完全沒可能發心臟病的患者, 這樣他們就不用在醫院多待一天。 我們用 AI 分析臨床試驗的數據, 尋找心電圖、血液成份、生物標記 和心臟病發作概率之間的關係。 在一個月內, 我們訓練的模型就能識別出 62% 的零發病風險患者。 這樣,這些患者就免去了 待在醫院的一天。 但是,你會放心地 在第一次服藥後直接回家, 就因為 AI 說你可以回家啦?
(Laughter)
(笑聲)
Doctors were not. What if we had false negatives, meaning people who are told by AI they can stay at home, and die?
醫師也不會放心。 萬一出現偽陰性呢? 也就是 AI 叫他們回家, 結果他們死了?
(Laughter)
(笑聲)
There started our 70 percent. We worked with a team of doctors to check the medical logic of each variable in our model. For instance, we were using the concentration of a liver enzyme as a predictor, for which the medical logic was not obvious. The statistical signal was quite strong. But what if it was a bias in our sample? That predictor was taken out of the model. We also took out predictors for which experts told us they cannot be rigorously measured by doctors in real life. After four months, we had a model and a medical protocol. They both got approved my medical authorities in the US last spring, resulting in far less stress for half of the patients and better quality of life. And an expected upside on sales over 100 million for that drug.
這時就需要那 70% 的作用了。 我們與醫師團隊合作, 檢驗模型中每個變量的醫學合理性。 比方說,我們用肝酵素濃度 作為預測變量, 其醫學邏輯並不明顯, 但從統計信號角度看, 它與結果有很大的關聯。 但萬一它實際上是個偏置項呢? [意即這個變量與心臟病無實際關聯] 模型就會剔除那個變量。 我們還剔除了一些變量, 因為醫師無法精準測出這些變量。 四個月後, 我們訓練出了模型, 制定了醫學使用規則。 去年春天它們都獲 美國醫療機構批准通過, 為一半服用這種藥的患者減輕壓力, 提高生活品質。 還預期這種藥的銷量 增加超過一億份。
Seventy percent weaving AI with team and processes also means building powerful interfaces for humans and AI to solve the most difficult problems together. Once, we got challenged by a fashion retailer. "We have the best buyers in the world. Could you build an AI engine that would beat them at forecasting sales? At telling how many high-end, light-green, men XL shirts we need to buy for next year? At predicting better what will sell or not than our designers." Our team trained a model in a few weeks, on past sales data, and the competition was organized with human buyers. Result? AI wins, reducing forecasting errors by 25 percent. Human-zero champions could have tried to implement this initial model and create a fight with all human buyers. Have fun. But we knew that human buyers had insights on fashion trends that could not be found in past data.
人類團隊和方法造就的 70%, 也意味著在人類和 AI 之間 建立堅固的連結, 共同解決最難的問題。 以前有一個時裝零售商問我們: 「世上有很會進貨的時裝零售商, 你能不能做一個 AI 比他們更會預測銷量, 告訴我們明年要訂購多少 高端服裝、淺綠色衣服、 加大碼男襯衫呢? 能不能預測哪些衣服會大賣, 預測得比設計師還準?」 我們的團隊在幾週內 用以往銷量數據訓練出模型, 和人類商家比賽。 猜猜誰贏了? AI 勝出,預測錯誤率比人類低 25%。 零人類思維的人可能會改進模型, 投入和人類商家的競爭。 開心就好。 但我們知道,人類對時尚潮流有遠見, 這是 AI 在以往數據學不到的。
There started our 70 percent. We went for a second test, where human buyers were reviewing quantities suggested by AI and could correct them if needed. Result? Humans using AI ... lose. Seventy-five percent of the corrections made by a human were reducing accuracy.
於是我們轉向 70%。 我們開始了第二次測試, 人類商家來審查 AI 推薦的購買量, 然後做出糾正。 結果如何? 使用 AI 的人類商家⋯ 輸了。 人類做出的糾正中, 有 75% 降低了 AI 的準確率。
Was it time to get rid of human buyers? No. It was time to recreate a model where humans would not try to guess when AI is wrong, but where AI would take real input from human buyers. We fully rebuilt the model and went away from our initial interface, which was, more or less, "Hey, human! This is what I forecast, correct whatever you want," and moved to a much richer one, more like, "Hey, humans! I don't know the trends for next year. Could you share with me your top creative bets?" "Hey, humans! Could you help me quantify those few big items? I cannot find any good comparables in the past for them." Result? "Human plus AI" wins, reducing forecast errors by 50 percent. It took one year to finalize the tool. Long, costly and difficult. But profits and benefits were in excess of 100 million of savings per year for that retailer.
是不是要放棄人類商家的介入了? 不是。 我們要重新搭建一個模型, 這一次,不讓人類猜 AI 的對錯, 而是讓 AI 尋求人類的建議。 我們將模型改頭換面, 拋棄了最初的交互方式, 有點像這樣: 「嘿人類!這是我的預測, 照你的意思幫我糾正一下吧!」 改進後的交互方式變得更廣泛,像這樣: 「嘿人類! 我不懂明年的流行趨勢, 可不可以告訴我你押寶在哪?」 「嘿人類! 可以幫我看看這些大傢伙嗎? 它們超出了我已知的知識範圍。」 結果如何? 「人類+AI」勝出, 這次預測錯誤率低了 50%。 我們花了一年才最終完成這個工具, 漫長、成本高又艱難, 但利潤和獲益頗豐厚, 每年為零售商節省超過一億美金。
Seventy percent on very sensitive topics also means human have to decide what is right or wrong and define rules for what AI can do or not, like setting caps on prices to prevent pricing engines [from charging] outrageously high prices to uneducated customers who would accept them. Only humans can define those boundaries -- there is no way AI can find them in past data.
在一些特定議題上, 70% 也意味著人類要決定對錯, 限制 AI 的權力。 例如設定價格上限, 防止 AI 粗暴地提高價格, 向不知情的顧客漫天要價。 只有人類能夠設定界線, 因為 AI 不可能從以往數據學到。
Some situations are in the gray zone. We worked with a health insurer. He developed an AI engine to identify, among his clients, people who are just about to go to hospital to sell them premium services. And the problem is, some prospects were called by the commercial team while they did not know yet they would have to go to hospital very soon. You are the CEO of this company. Do you stop that program? Not an easy question.
有時我們可能遇到灰色地帶。 我們曾和保險公司有過合作, 他們開發了一個 針對客戶的 AI 系統, 用來識別健康狀況較差, 很可能即將要去治病的客戶, 向他們推銷附加產品。 問題是, 一些接到推銷電話的客戶, 這時候並不知道 他們很可能馬上要去醫院看病。 如果你是這家公司的執行長, 你會取消掉這個項目嗎? 這是個兩難抉擇。
And to tackle this question, some companies are building teams, defining ethical rules and standards to help business and tech teams set limits between personalization and manipulation, customization of offers and discrimination, targeting and intrusion.
為了解決這個問題, 一些公司正在組建團隊, 幫商業和科技團隊 制訂倫理規則和標準, 在個性化和可操作性間尋找平衡點, 區別客製與偏見, 分清關照與冒犯。
I am convinced that in every company, applying AI where it really matters has massive payback. Business leaders need to be bold and select a few topics, and for each of them, mobilize 10, 20, 30 people from their best teams -- tech, AI, data science, ethics -- and go through the full 10-, 20-, 70-percent cycle of "Human plus AI," if they want to land AI effectively in their teams and processes. There is no other way.
我堅信在每家公司, 把 AI 運用到關鍵之處 定會有巨大回報。 商業領袖們要勇敢嘗試, 選擇幾個項目, 每個項目,召集幾十個領域佼佼者—— 科技、AI、數據科學、倫理—— 然後完成 10%、20%、70% 的 「人類+AI」。 這樣 AI 就可以和人類高效合作。 除此之外別無他法。
Citizens in developed economies already fear algocracy. Seven thousand were interviewed in a recent survey. More than 75 percent expressed real concerns on the impact of AI on the workforce, on privacy, on the risk of a dehumanized society. Pushing algocracy creates a real risk of severe backlash against AI within companies or in society at large. "Human plus AI" is our only option to bring the benefits of AI to the real world. And in the end, winning organizations will invest in human knowledge, not just AI and data. Recruiting, training, rewarding human experts. Data is said to be the new oil, but believe me, human knowledge will make the difference, because it is the only derrick available to pump the oil hidden in the data.
經濟飛速發展的同時, 公民已對 AI 官僚主義產生恐懼。 在近期對七千人的調查中, 超過 75% 的人表示實質的擔憂, 擔心 AI 影響勞動力、隱私, 擔心社會會失去人性。 AI 官僚主義的出現 會導致公司和社會 對 AI 的強烈牴觸。 「人類+AI」是唯一選項, 只有這樣才能讓 AI 真正帶來福祉。 到頭來, 因 AI 獲利的組織, 會為人類知識投資, 而不僅僅投資 AI 和數據。 招募、培養、獎勵人類專家。 有人說數據是新的燃料, 但相信我,人類知識能改變世界。 因為人類知識是唯一的泵, 將蘊藏於數據中的燃料泵出。
Thank you.
謝謝大家。
(Applause)
(掌聲)