评论
打印
Business
商业版块
Bartleby
巴托比专栏
The dangers of data
数据的危险
Human judgment is flawed, but numbers have risks too.
人类判断有缺陷，但数据也有风险。
Managers are better equipped than ever to make good decisions.
管理者比以往任何时候都更有能力做出好的决策。
They are more aware that human judgment is fallible.
他们更清楚地意识到人类的判断是容易出错的。
They have oodles of data about their customers and products.
他们拥有关于客户和产品的海量数据。
They can use artificial intelligence (AI) to analyse, summarise and synthesise information with unprecedented speed.
他们可以使用人工智能以前所未有的速度分析、总结、综合信息。
But as the pendulum swings inexorably away from gut instinct and towards data-based decisions, firms need to be alive to a different set of dangers.
但是在时代的钟摆不可阻挡地从直觉荡向基于数据的决策之际，公司需要意识到一系列不同的危险。
In a recent paper Linda Chang of the Toyota Research Institute and her co-authors identify a cognitive bias that they call “quantification fixation”.
在最近的一篇论文中，丰田研究所的琳达·张及其合著者发现了一种被称为“量化情结”的认知偏见。
The risk of depending on data alone to make decisions is familiar: it is sometimes referred to as the McNamara fallacy, after the emphasis that an American secretary of defence put on misleading quantitative measures in assessing the Vietnam war.
仅依靠数据来做决策的风险并不令人陌生：这有时被称为麦克纳马拉谬误，得名于美国国防部长麦克纳马拉在评估越南战争时极其看重具有误导性的量化指标。
But Ms Chang and her co-authors help explain why people put disproportionate weight on numbers.
但是张女士及其合著者帮助解释了为什么人们会过度重视数字。
The reason seems to be that data are particularly suited to making comparisons.
原因似乎是数据特别适合进行比较。
In one experiment, participants were asked to imagine choosing between two software engineers for a promotion.
在一项实验中，参与者被要求想象在两名软件工程师之间选择晋升人选。
One engineer had been assessed as more likely to climb the ladder but less likely to stay at the firm; the other, by contrast, had a higher probability of retention but a lower chance of advancement.
一位工程师被评估为更有可能晋升，但不太可能留在公司；相比之下，另一位工程师留在公司的概率更高，但晋升的概率更低。
The researchers varied the way that this information was presented.
研究人员改变了信息呈现的方式。
They found that participants were more likely to choose on the basis of future promotion prospects when only that criterion was quantified, and to select on retention probability when that was the thing with a number attached.
他们发现，当只有未来晋升前景被量化时，参与者更有可能根据晋升前景进行选择，而当只有留任概率被量化时，参与者更有可能根据留任概率进行选择。
One answer to this bias is to quantify everything.
对这种偏见的一个回应办法是将一切量化。
But, as the authors point out, some things are mushier than others.
但是正如作者们指出的，有些事情比其他事情更模糊。
A firm’s culture is harder to express as a number for job-seekers than its salary levels.
企业文化比薪资水平更难以数字的形式传达给求职者。
Data can tell an early-stage investor more about a startup’s financials than a founder’s resilience.
数据可以让早期投资者知道某个初创公司的财务状况，而不是创始人的坚韧品格。
Numbers allow for easy comparisons.
数字便于比较。
The problem is that they do not always tell the whole story.
问题在于数字并不总是讲述整个故事。
There are other risks, too.
此外还有其他风险。
Humans bring the same cognitive biases to their analysis of numbers as they do to other decisions.
人类在分析数字时会像做其他决策时那样有认知偏见。
Take confirmation bias, the propensity to interpret information as support for your point of view.
以确认偏差为例，即将信息解释为支持自己观点的倾向。
In another experiment Itai Yanai of New York University and Martin Lercher of Heinrich Heine University asked computer-science undergraduates to say what general correlation they expected between wealth and happiness, before showing them a fictitious dataset of the relationship between these two variables for 1,000 individuals.
在另一项实验中，纽约大学的伊泰·雅奈和海因里希海涅大学的马丁·莱凯尔让计算机科学专业的本科生说出他们预期财富和幸福之间有什么一般相关性，然后向他们展示了一个虚构的数据集，其中包含了1000个人的财富和幸福之间的关系。
Faced with an identical graph, students who expected a positive correlation were much more likely to see one in the data.
面对相同的图表，那些期待正相关的学生更有可能在数据中看到正相关。
Beliefs influenced interpretation.
信念影响了解释。
Plenty of people struggle with basic data literacy: consumers are less likely to participate in competitions with higher numbers of contestants, even when the odds of winning a prize are exactly the same.
很多人对于理解基本的数据知识都有困难：即使获奖的几率完全相同，消费者也不太可能参加参赛者数量较多的比赛。
In a world giddy with excitement over AI models, relying on algorithms may seem like the sensible solution to this.
在一个对人工智能模型兴奋不已的世界里，依靠算法似乎是解决这个问题的明智之举。
In one more experiment, Hossein Nikpayam and Mirko Kremer of the Frankfurt School of Finance and Management and Francis de Véricourt of ESMTt Berlin found that managers were unimpressed when other decision-makers ignored machine-led recommendations and exercised their own judgment.
在另一项实验中，法兰克福金融管理学院的侯赛因·尼克帕姆和米尔科·克雷默，以及欧洲管理与技术学院的弗朗西斯·德维里科特发现，当其他决策者忽视机器主导的建议并自行判断时，管理者们对此不以为意。
They blamed them if the outcome was bad, and did not reward them if it was good.
如果结果不好，管理者会责怪他们，如果结果好，也不会奖励他们。
People used to say that nobody ever got fired for buying IBM.
人们过去常说，没有人会因为购买IBM的软件而被解雇。（注：这句话表示人们对可靠技术或品牌的依赖。）
It’s not hard to imagine “nobody gets fired for following the algorithm” becoming the modern-day equivalent.
不难想象，“没人会因为遵循算法而被解雇”将成为与这句话等同的现代名言。
But there are times when humans have an advantage.
但有时候人类是有优势的。
Datasets reflect back the world as it is, for example, not the world as it might be.
例如，数据集反映的是世界的现状，而不是世界可能的样子。
It’s harder to evaluate radically new ideas by looking at existing patterns.
通过观察现有的模式来评估全新的想法是比较困难的。
In the early days of HBO, a pioneering TV channel, executives operated on a mixture of instinct and contrarianism to commission programmes that broke the mould: profane comedy specials, a prison drama that killed off a main character in the first episode.
在早期的HBO（一个开创性的电视频道），高管们依靠直觉和逆向思维的混合来找人制作打破常规的节目：亵渎性喜剧特辑，在第一集就杀死了主角的监狱剧。
Other networks turned down the idea of a violent mobster in therapy; HBO did not.
其他电视网拒绝了一个暴力匪徒接受治疗的创意，HBO没有拒绝。
Relying on data might have led to more explicable decisions, but they would also have been safer.
依靠数据可能会带来更可解释的决策，但这些决策也是更安全保守的。
None of this is to say that instinct trumps data, or to claim that humans make better decisions than machines.
这并不是说直觉胜过数据，也不是说人类能比机器做出更好的决策。
Far from it.
情况远非如此。
But it is a warning.
但这是一个提醒。
Numbers promise rigour, certainty and objectivity.
数字意味着严谨、确定、客观。
They have flaws, too.
但数字也有缺陷。
