38.网球与AI监督

Passage Thirty-Eight

Tennis and AI oversight

网球与AI监督

【1】Wimbledon’s centre court has seen its share of rivalries; think of McEnroe vs Borg, or Williams vs Williams. But for David Almog, a behavioural economist at Northwestern University, the match worth tuning in for is umpire(judge) vs machine.

温网中央球场上的对决往往是火力全开，想想麦肯罗对战博格，或者大小威廉姆斯奋战的场景。但对美国西北大学的行为经济学家大卫·阿尔莫格来说，值得关注的则是裁判与机器的较量。

【2】How AI oversight affects human decision-making is an important question in a world where algorithms play an ever-larger role in everyday life. Car drivers, financial traders and air-traffic controllers already routinely(habitually) see their decisions overruled by AI systems put in place to rapidly(quickly) correct poor judgment. Doctors, judges and even soldiers could be next.

随着算法在当下生活中发挥着日益重要的作用，AI监督如何AI影响人类决策成为一个重要问题。系统用于快速纠正错误判断，汽车司机、金融交易员和空中交通管制员已经习惯了AI系统推翻他们决策的日常。医生、法官甚至军人可能会是下一批受影响的对象。

【3】Much of this correction happens out of the public eye, thwarting would-be analysts. But, says Mr Almog, “tennis is one of the most visible settings where final decision rights are granted to AI.”

大部分AI修正都发生在公众视线之外，蒙蔽了潜在的分析人士。但是，阿尔莫格表示：“在AI获得最终决策权后，网球成为了最透明的监管场景之一。”

【4】That is why, together with colleagues in America and Australia, he has looked at whether tennis umpires and line judges correctly called balls in or out during nearly 100,000 points played in some 700 matches across the world, both before and after the introduction of the Hawk-Eye ball-tracking system in 2006.

这就是为什么他与美国和澳大利亚的同事一起研究了全球范围内大约700场比赛中近10万个球的情况。他们着重研究了裁判员和线判员对于球是否出界的判断是否正确，其中包括了2006年鹰眼追踪系统推出前后的数据。

【5】The Hawk-Eye system, now used at most elite tournaments, uses between six and ten cameras positioned(placed) around the court to create a three-dimensional representation(establishment) of the ball’s trajectory.

鹰眼系统现在用于大多数精英锦标赛中，它使用放置在球场周围的6到10台摄像机来创建球的三维轨迹。

【6】This can then be presented on a screen visible to players, spectators and officials—as well as TV viewers. Players can use it to appeal human decisions, with the AI’s verdict considered final. Bad calls from line judges and umpires are now often overturned.

随后在屏幕上呈现给球员、观众和裁判，以及电视观众。由于AI拥有最终裁决权，球员可以利用鹰眼系统质疑裁判的判决。现在裁判和线判员的错误判罚经常被推翻。

【7】The latest analysis from Mr Almog and his colleagues, published as a preprint last month, showed that Hawk-Eye oversight has prompted human officials to up their game and make 8% less mistakes than before it was introduced. (That comparison can be made thanks to a 2005 trial period in which Hawk-Eye was used without the ability to influence calls.) Such an improvement in performance is to be expected, the researchers say, given the heightened watchfulness that accompanies the threat of public shaming.

阿尔莫格和他的同事最新的分析上个月以预印本形式发表。分析表明鹰眼监督促使人类裁判提高水平，失误率比引入之前减少了8%。（此番比较得益于2005年鹰眼系统的试用，当时使用了鹰眼，但无法影响裁判的判决。）研究人员表示，这样的表现改善是意料之中的，因为公开丢面子的风险增加，由此裁判不得不打起精神。

【8】Most of the improvement came during the multi-shot rallies that follow a successful serve and return. But when the researchers looked at serves in particular(extraordinary), and especially in cases where the served ball landed within 20mm either side of a line, they were surprised to see the error rate soar(enhance).

对于成功发球和接发球后的连续对打，裁判准确度的进步最为明显。但是，当研究人员特别关注发球时，尤其是发球落在球线两侧20毫米范围内的情况时，他们惊讶地发现失误率飙升。

【9】The umpires and line judges, it turned out, had switched strategy(tactic). Before Hawk-Eye, they were more likely to call a serve out when it was in. But afterwards, they were even more likely to wave through balls that were actually out. For every 100 mis-hit serves, post-Hawk-Eye umpires left 39 unchallenged. The comparable figure in the earlier era was 26.

事实证明，裁判员和线判员改变了策略。在没有鹰眼系统之前，他们更倾向于在球实际上入界的情况下判为出界。但是在鹰眼系统引入后，他们更倾向于放宽实际出界的球。鹰眼引进后，每100个失误发球中，裁判员有39次未对错误发球作出裁决，早期这一数字是26。

【10】Such a shift is easily understood. Overlooked faults are less disruptive in tennis than incorrect cries of “out” because these end the point prematurely. They can also trigger(cause) dissent(objection) from both the player and crowd when the error is identified on the big screen. It seems that human officials take the less reputationally risky option, even if it leads to more incorrect calls.

这种转变很好理解。在网球比赛中，相比被忽视的错误，错误地喊“出界”更具破坏性，因为这会过早地结束比赛。当错误在大屏幕上被揭示出来时，还可能引发球员和观众的异议。即使可能会出现更多的错误判断，但是裁判员似乎选择了声誉风险较小的选项。

【11】Tennis, with its binary outcomes and clear evidence of whether a decision was right or wrong, offers a highly simplified model for AI oversight. But many of the same tendencies will be at play in fields like medicine and law, says Mr Almog, and should be considered before algorithms are allowed to trump human decisions.

网球具有明确的二元结果以及判断正确性的证据，为AI监督提供了一个高度简化的模型。但是，阿尔莫格表示，在医学和法律等领域可能会出现相同的AI驱动趋势，在算法胜过人类决策之前，我们应该对其有所思量。

【12】Most important, perhaps, is the social cost of getting an important call wrong, which will vary between disciplines, and could distort decision-making in different ways. judges, for example, may prefer to under-convict. Doctors, on the other hand, might over-diagnose. Stay tuned.

重要决策失误所带来的社会成本在不同领域可能存在差异，也可能以不同方式导致决策失真，或许这才是最重要的。例如，法官可能会倾向于轻判。然而，医生可能会过度诊断。AI监督，是福是祸，让我们敬请期待。