How humans and AI can work together to create better businesses | Sylvain Duranton

Translator: Ivana Korom Reviewer: Krystian Aparta

譯者: Harper Chang 審譯者: Helen Chang

Let me share a paradox.

我來和大家分享發現的一個悖論。

For the last 10 years,

在過去十年裡，

many companies have been trying to become less bureaucratic,

很多公司都想擺脫官僚化，

to have fewer central rules and procedures,

通過減少職務、精簡程序，

more autonomy for their local teams to be more agile.

給團隊更多自主， 讓公司運作更靈活。

And now they are pushing artificial intelligence, AI,

現在，公司開始引進人工智慧，AI，

unaware that cool technology

卻沒意識到這個很酷的科技，

might make them more bureaucratic than ever.

可能讓他們比以前更官僚。

Why?

為什麼呢？

Because AI operates just like bureaucracies.

因為 AI 的運作方式就很官僚。

The essence of bureaucracy

官僚的本質

is to favor rules and procedures over human judgment.

就是看重規則和程序， 而不是人類自身的判斷，

And AI decides solely based on rules.

而 AI 僅依據規則做決策。

Many rules inferred from past data

即便 AI 的規則 依據過去的數據形成，

but only rules.

那也終究是規則。

And if human judgment is not kept in the loop,

如果將人類判斷置若罔聞，

AI will bring a terrifying form of new bureaucracy --

對 AI 的運用將帶來 可怕的新官僚主義——

I call it "algocracy" --

我稱之為「AI 官僚主義」 (algocracy)，

where AI will take more and more critical decisions by the rules

意即 AI 將根據規則， 不受人為控制，

outside of any human control.

做更多重要決策。

Is there a real risk?

真有風險嗎？

Yes.

當然有。

I'm leading a team of 800 AI specialists.

我領導的團隊 由 800 位 AI 專家組成，

We have deployed over 100 customized AI solutions

我們為很多全球的大公司

量身打造了超過 100 個 AI 系統。

for large companies around the world.

And I see too many corporate executives behaving like bureaucrats from the past.

我看過太多的公司高管 因此而變得官僚作風。

They want to take costly, old-fashioned humans out of the loop

他們對麻煩、老舊的 人類決策嗤之以鼻，

and rely only upon AI to take decisions.

完全依賴 AI 來做決策。

I call this the "human-zero mindset."

我稱之為「無人類思維」。 (human-zero mindset)

And why is it so tempting?

可為何這種思維這麼誘人？

Because the other route, "Human plus AI," is long,

因為另一種思維—— 「人類+AI」(Human plus AI)

費時、費錢又費力。

costly and difficult.

Business teams, tech teams, data-science teams

商業團隊、科技團隊和數據科學團隊

have to iterate for months

不得不費幾個月的功夫

to craft exactly how humans and AI can best work together.

探索人類和 AI 怎樣達到最好的合作。

Long, costly and difficult.

探索過程漫長艱難，耗了很多錢，

But the reward is huge.

但取得了巨大成果。

A recent survey from BCG and MIT

根據波士頓諮詢公司 和麻省理工學院最近的調查，

shows that 18 percent of companies in the world

全球有 18% 的公司

are pioneering AI,

在研究 AI，

making money with it.

藉此盈利。

Those companies focus 80 percent of their AI initiatives

那些公司把 80% 的 AI 創新

on effectiveness and growth,

專注在效率和成長，

taking better decisions --

做更好的決策，

not replacing humans with AI to save costs.

而不是用 AI 取代人類來減少開支。

Why is it important to keep humans in the loop?

為什麼人類的作用必不可少？

Simply because, left alone, AI can do very dumb things.

原因很簡單： 沒有人類，AI 會幹傻事。

Sometimes with no consequences, like in this tweet.

有時候毫無幫助， 就像這條推文講的：

"Dear Amazon,

「親愛的亞馬遜公司，

I bought a toilet seat.

我之前買了一個馬桶圈。

Necessity, not desire.

生活必需品，不是什麼私人愛好。

I do not collect them,

我不收藏馬桶圈，

I'm not a toilet-seat addict.

我沒有馬桶圈癮。

No matter how temptingly you email me,

不管你的廣告郵件多誘人，

I am not going to think, 'Oh, go on, then,

我都不會覺得 『噢，受不了，

one more toilet seat, I'll treat myself.' "

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.

有時候，AI 又「太有幫助」， 就像這條推文：

"Had the same situation

「我在買我媽媽的骨灰盒後

with my mother's burial urn."

遇到了同樣狀況。」

(Laughter)

（笑聲）

"For months after her death,

「在她去世後的幾個月，

I got messages from Amazon, saying, 'If you liked that ...' "

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.

有時 AI 幹壞事。

Take an AI engine rejecting a student application for university.

比如說 AI 曾經拒絕 學生的大學申請。

Why?

為什麼？

Because it has "learned," on past data,

因為這個 AI 從以前的數據「學」了

characteristics of students that will pass and fail.

哪些學生會通過，哪些學生不能——

Some are obvious, like GPAs.

有一些指標很明顯，比如 GPA。

But if, in the past, all students from a given postal code have failed,

但如果在過去，某個地區 所有的學生都沒通過，

it is very likely that AI will make this a rule

AI 很可能就以此定下規則，

and will reject every student with this postal code,

拒絕每一個來自這個地區的學生，

not giving anyone the opportunity to prove the rule wrong.

不給任何人證明規則有誤的機會。

And no one can check all the rules,

沒有人能檢查每一條規則，

because advanced AI is constantly learning.

因為先進的 AI 一直在學。

And if humans are kept out of the room,

所以如果直接用 AI 取代人類，

there comes the algocratic nightmare.

迎來的將是 AI 官僚主義的噩夢：

Who is accountable for rejecting the student?

誰應該對學生的被拒負責？

No one, AI did.

沒有誰，AI 負責。

Is it fair? Yes.

這公平嗎？公平。

The same set of objective rules has been applied to everyone.

因為所有學生都用同一規則判定。

Could we reconsider for this bright kid with the wrong postal code?

那可不可以重新考慮這個 「住錯了地方」的聰明學生？

No, algos don't change their mind.

不行，AI 官僚不會改主意。

We have a choice here.

我們的選項是，

Carry on with algocracy

繼續 AI 的獨裁，

or decide to go to "Human plus AI."

還是考慮「人類+AI」思維？

And to do this,

要擁有這種思維，

we need to stop thinking tech first,

我們要先把科技放在一邊，

and we need to start applying the secret formula.

從秘密公式入手。

To deploy "Human plus AI,"

要實現「人類+AI」，

10 percent of the effort is to code algos;

需要 10% 的程式；

20 percent to build tech around the algos,

20% 的科技成份，

collecting data, building UI, integrating into legacy systems;

包括收集數據、構建用戶界面、 整合進遺留系統；

But 70 percent, the bulk of the effort,

但 70%，最主要的部份，

is about weaving together AI with people and processes

是結合 AI 和人類方法，

to maximize real outcome.

讓結果最接近完美。

AI fails when cutting short on the 70 percent.

如果這 70% 被削減， AI 就會出現問題。

The price tag for that can be small,

代價可以很小，

wasting many, many millions of dollars on useless technology.

比如在無用科技上浪費數百萬美元。

Anyone cares?

誰會在乎呢？

Or real tragedies:

但代價也可以大到無法承受：

346 casualties in the recent crashes of two B-737 aircrafts

最近兩起波音 737 空難中 346 人遇難，

when pilots could not interact properly

原因是電腦控制的飛行系統

with a computerized command system.

無法正確回應飛機師的指令。

For a successful 70 percent,

要實現人類的 70%，

the first step is to make sure that algos are coded by data scientists

第一步是保證算法程式由數據科學家

and domain experts together.

和領域專家共同完成。

Take health care for example.

拿醫療保健舉例，

One of our teams worked on a new drug with a slight problem.

我們有一個團隊 曾經處理一種藥產生的小問題。

When taking their first dose,

在第一次服用這種藥後，

some patients, very few, have heart attacks.

有一小部份的患者會發心臟病。

So, all patients, when taking their first dose,

於是所有第一次服這種藥的患者

have to spend one day in hospital,

都要住院觀察一天，

for monitoring, just in case.

以防心臟病發作。

Our objective was to identify patients who were at zero risk of heart attacks,

我們想要識別出 完全沒可能發心臟病的患者，

who could skip the day in hospital.

這樣他們就不用在醫院多待一天。

We used AI to analyze data from clinical trials,

我們用 AI 分析臨床試驗的數據，

to correlate ECG signal, blood composition, biomarkers,

尋找心電圖、血液成份、生物標記

with the risk of heart attack.

和心臟病發作概率之間的關係。

In one month,

在一個月內，

our model could flag 62 percent of patients at zero risk.

我們訓練的模型就能識別出 62% 的零發病風險患者。

They could skip the day in hospital.

這樣，這些患者就免去了 待在醫院的一天。

Would you be comfortable staying at home for your first dose

但是，你會放心地 在第一次服藥後直接回家，

if the algo said so?

就因為 AI 說你可以回家啦？

(Laughter)

（笑聲）

Doctors were not.

醫師也不會放心。

What if we had false negatives,

萬一出現偽陰性呢？

meaning people who are told by AI they can stay at home, and die?

也就是 AI 叫他們回家， 結果他們死了？

(Laughter)

（笑聲）

There started our 70 percent.

這時就需要那 70% 的作用了。

We worked with a team of doctors

我們與醫師團隊合作，

to check the medical logic of each variable in our model.

檢驗模型中每個變量的醫學合理性。

For instance, we were using the concentration of a liver enzyme

比方說，我們用肝酵素濃度

as a predictor,

作為預測變量，

for which the medical logic was not obvious.

其醫學邏輯並不明顯，

The statistical signal was quite strong.

但從統計信號角度看， 它與結果有很大的關聯。

But what if it was a bias in our sample?

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]

That predictor was taken out of the model.

模型就會剔除那個變量。

We also took out predictors for which experts told us

我們還剔除了一些變量，

they cannot be rigorously measured by doctors in real life.

因為醫師無法精準測出這些變量。

After four months,

四個月後，

we had a model and a medical protocol.

我們訓練出了模型， 制定了醫學使用規則。

They both got approved

去年春天它們都獲 美國醫療機構批准通過，

my medical authorities in the US last spring,

resulting in far less stress for half of the patients

為一半服用這種藥的患者減輕壓力，

and better quality of life.

提高生活品質。

And an expected upside on sales over 100 million for that drug.

還預期這種藥的銷量 增加超過一億份。

Seventy percent weaving AI with team and processes

人類團隊和方法造就的 70%，

also means building powerful interfaces

也意味著在人類和 AI 之間 建立堅固的連結，

for humans and AI to solve the most difficult problems together.

共同解決最難的問題。

Once, we got challenged by a fashion retailer.

以前有一個時裝零售商問我們：

"We have the best buyers in the world.

「世上有很會進貨的時裝零售商，

Could you build an AI engine that would beat them at forecasting sales?

你能不能做一個 AI 比他們更會預測銷量，

At telling how many high-end, light-green, men XL shirts

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？

we need to buy for next year?

At predicting better what will sell or not

能不能預測哪些衣服會大賣，

than our designers."

預測得比設計師還準？」

Our team trained a model in a few weeks, on past sales data,

我們的團隊在幾週內 用以往銷量數據訓練出模型，

and the competition was organized with human buyers.

和人類商家比賽。

Result?

猜猜誰贏了？

AI wins, reducing forecasting errors by 25 percent.

AI 勝出，預測錯誤率比人類低 25%。

Human-zero champions could have tried to implement this initial model

零人類思維的人可能會改進模型，

and create a fight with all human buyers.

投入和人類商家的競爭。

Have fun.

開心就好。

But we knew that human buyers had insights on fashion trends

但我們知道，人類對時尚潮流有遠見，

that could not be found in past data.

這是 AI 在以往數據學不到的。

There started our 70 percent.

於是我們轉向 70%。

We went for a second test,

我們開始了第二次測試，

where human buyers were reviewing quantities

人類商家來審查

suggested by AI

AI 推薦的購買量，

and could correct them if needed.

然後做出糾正。

Result?

結果如何？

Humans using AI ...

使用 AI 的人類商家⋯

lose.

輸了。

Seventy-five percent of the corrections made by a human

人類做出的糾正中，

有 75% 降低了 AI 的準確率。

were reducing accuracy.

Was it time to get rid of human buyers?

是不是要放棄人類商家的介入了？

No.

不是。

It was time to recreate a model

我們要重新搭建一個模型，

where humans would not try to guess when AI is wrong,

這一次，不讓人類猜 AI 的對錯，

but where AI would take real input from human buyers.

而是讓 AI 尋求人類的建議。

We fully rebuilt the model

我們將模型改頭換面，

and went away from our initial interface, which was, more or less,

拋棄了最初的交互方式， 有點像這樣：

"Hey, human! This is what I forecast,

「嘿人類！這是我的預測，

correct whatever you want,"

照你的意思幫我糾正一下吧！」

and moved to a much richer one, more like,

改進後的交互方式變得更廣泛，像這樣：

"Hey, humans!

「嘿人類！

I don't know the trends for next year.

我不懂明年的流行趨勢，

Could you share with me your top creative bets?"

可不可以告訴我你押寶在哪？」

"Hey, humans!

「嘿人類！

Could you help me quantify those few big items?

可以幫我看看這些大傢伙嗎？

I cannot find any good comparables in the past for them."

它們超出了我已知的知識範圍。」

Result?

結果如何？

"Human plus AI" wins,

「人類+AI」勝出，

reducing forecast errors by 50 percent.

這次預測錯誤率低了 50%。

It took one year to finalize the tool.

我們花了一年才最終完成這個工具，

Long, costly and difficult.

漫長、成本高又艱難，

But profits and benefits

但利潤和獲益頗豐厚，

were in excess of 100 million of savings per year for that retailer.

每年為零售商節省超過一億美金。

Seventy percent on very sensitive topics

在一些特定議題上，

also means human have to decide what is right or wrong

70% 也意味著人類要決定對錯，

and define rules for what AI can do or not,

限制 AI 的權力。

like setting caps on prices to prevent pricing engines

例如設定價格上限，

防止 AI 粗暴地提高價格，

[from charging] outrageously high prices to uneducated customers

向不知情的顧客漫天要價。

who would accept them.

Only humans can define those boundaries --

只有人類能夠設定界線，

there is no way AI can find them in past data.

因為 AI 不可能從以往數據學到。

Some situations are in the gray zone.

有時我們可能遇到灰色地帶。

We worked with a health insurer.

我們曾和保險公司有過合作，

He developed an AI engine to identify, among his clients,

他們開發了一個 針對客戶的 AI 系統，

people who are just about to go to hospital

用來識別健康狀況較差， 很可能即將要去治病的客戶，

to sell them premium services.

向他們推銷附加產品。

And the problem is,

問題是，

some prospects were called by the commercial team

一些接到推銷電話的客戶，

while they did not know yet

這時候並不知道

they would have to go to hospital very soon.

他們很可能馬上要去醫院看病。

You are the CEO of this company.

如果你是這家公司的執行長，

Do you stop that program?

你會取消掉這個項目嗎？

Not an easy question.

這是個兩難抉擇。

And to tackle this question, some companies are building teams,

為了解決這個問題， 一些公司正在組建團隊，

defining ethical rules and standards to help business and tech teams set limits

幫商業和科技團隊 制訂倫理規則和標準，

between personalization and manipulation,

在個性化和可操作性間尋找平衡點，

customization of offers and discrimination,

區別客製與偏見，

targeting and intrusion.

分清關照與冒犯。

I am convinced that in every company,

我堅信在每家公司，

applying AI where it really matters has massive payback.

把 AI 運用到關鍵之處 定會有巨大回報。

Business leaders need to be bold

商業領袖們要勇敢嘗試，

and select a few topics,

選擇幾個項目，

and for each of them, mobilize 10, 20, 30 people from their best teams --

每個項目，召集幾十個領域佼佼者——

tech, AI, data science, ethics --

科技、AI、數據科學、倫理——

and go through the full 10-, 20-, 70-percent cycle

然後完成 10%、20%、70% 的

of "Human plus AI,"

「人類+AI」。

if they want to land AI effectively in their teams and processes.

這樣 AI 就可以和人類高效合作。

There is no other way.

除此之外別無他法。

Citizens in developed economies already fear algocracy.

經濟飛速發展的同時， 公民已對 AI 官僚主義產生恐懼。

Seven thousand were interviewed in a recent survey.

在近期對七千人的調查中，

More than 75 percent expressed real concerns

超過 75% 的人表示實質的擔憂，

on the impact of AI on the workforce, on privacy,

擔心 AI 影響勞動力、隱私，

on the risk of a dehumanized society.

擔心社會會失去人性。

Pushing algocracy creates a real risk of severe backlash against AI

AI 官僚主義的出現

within companies or in society at large.

會導致公司和社會 對 AI 的強烈牴觸。

"Human plus AI" is our only option

「人類+AI」是唯一選項，

to bring the benefits of AI to the real world.

只有這樣才能讓 AI 真正帶來福祉。

And in the end,

到頭來，

winning organizations will invest in human knowledge,

因 AI 獲利的組織， 會為人類知識投資，

not just AI and data.

而不僅僅投資 AI 和數據。

Recruiting, training, rewarding human experts.

招募、培養、獎勵人類專家。

Data is said to be the new oil,

有人說數據是新的燃料，

but believe me, human knowledge will make the difference,

但相信我，人類知識能改變世界。

because it is the only derrick available

因為人類知識是唯一的泵，

to pump the oil hidden in the data.

將蘊藏於數據中的燃料泵出。

Thank you.

謝謝大家。

(Applause)

（掌聲）