自动驾驶汽车难以克服的局限性

We humans are fallible, imperfect beings prone to making mistakes and basic errors of judgment. So we create technologies to help ourselves out.

我们人类容易犯错，不完美，容易犯错和出现基本的判断失误。因此，我们发明技术来帮助自己。

Take plans to launch fleets of self-driving cars. The belief here is that the technology will improve safety. But this is a very bold assumption to make at this stage in its development.

以推出自动驾驶汽车的计划为例。人们相信，这项技术将加强安全。但在当前发展阶段，这是一个非常大胆的假设。

For one thing, the evidence to support the view simply is not there yet. The best data we have come from tests conducted over the course of 2016 in California, a state which, it must be remembered, boasts a mild climate hardly representative of global driving conditions.

首先，支持这种观点的证据并不存在。我们拥有的最好数据来自加州在2016年期间进行的测试。必须记住，这个州拥有温暖的气候，很难代表全球的驾驶状况。

Google’s Waymo scored best, with one human intervention every 5,127 miles driven. This was an improvement on the year before, but nowhere near perfect. In all, Waymo’s 60 testing vehicles drove about 10,597 miles in 2016 — 3,000 miles less than the annual US average per vehicle — and within that timeframe required at least two interventions each.

谷歌(Google)的Waymo得分最高，每驾驶5127英里需要一次人为干预。这比上一年有所改善，但远远算不上完美。2016年，Waymo的60辆测试汽车的驾驶总里程约为1.0597万英里(比美国每辆汽车的年度平均里程少3000英里)，其间每辆汽车需要至少两次干预。

Tesla fared much worse. The electric vehicle maker’s four cars tested on average 137 miles each that year, encountering 45 disengagement events per vehicle — or one roughly every three miles. Each intervention represents an accident which was potentially avoided.

特斯拉(Tesla)的表现糟糕得多。去年，这家电动汽车制造商有4辆汽车接受测试，每辆汽车的平均行驶里程为137英里，每辆汽车发生45次人为干预，大约每3英里一次。每次干预代表着一起潜在被避免的事故。

Given that most industry watchers believe the public will not tolerate any faults at all, none of this is encouraging. It is certainly true that the technology is improving. But it is also the case that self-driving cars have been around since the mid- 1990s, with one vehicle achieving a 98.2 per cent “autonomous driving percentage” even back then.

鉴于多数行业观察者认为公众不会容忍任何故障，因此这些结果都不振奋人心。相关技术确实在改进。但还有一个事实，自动驾驶汽车早在上世纪90年代中期就已出现，甚至当时就有一款汽车的“自动驾驶比例”达到了98.2%。

But even if the technical challenges can be surmounted, unexpected negative externalities probably cannot.

然而，即便能够克服技术挑战，意料之外的负面外部性也很可能难以逾越。

A case in point is Uber’s latest autonomous vehicle accident in Arizona where the fault lay with the human driver of another car not Uber’s vehicle. In this case the human failed to yield, drawing attention to one of the biggest challenges for the forthcoming autonomous transition: a world in which humans and autonomous vehicles will have to interact with each other safely.

优步(Uber)在亚利桑那州的最新自动驾驶汽车事故恰恰说明了这点，事故责任方是另一辆汽车的人类驾驶员，而非优步的汽车。在这个例子里，责任人没有让路，这令人关注即将到来的自动化过渡的最大挑战之一：人类和自动驾驶汽车必须能够安全互动。

What motivates humans to act, however, is very different to what motivates algorithms. At the basic level, most drivers — save those who are drunk, suicidal or intent on sowing fear or terror — have an interest in their own self-preservation or the preservation of others. That cannot be guaranteed of complex algorithms.

然而，推动人类行动的因素与算法的运行迥然不同。在基本层面，多数驾驶者(除了那些醉酒、有自杀倾向或者蓄意制造害怕或恐惧的人)都有意保护自己，保护其他人。复杂的算法并不能保证会如此。

There are exceptions, such as last week’s terror attack in London. But eliminating humans from the wheel will not necessarily reduce the risks. Cars that drive themselves could be easily weaponised, since they need only to be hacked, not driven by a martyr.

也有例外，例如最近伦敦发生的恐怖袭击。但是淘汰人类驾驶员不一定会减轻风险。那些自动驾驶汽车可能很容易变成武器，因为他们只需要被黑客入侵，而不需要由恐怖分子驾驶。

Alcoholism, meanwhile, killed three times the number of Americans that fatal crashes did in 2014. So if driving encourages sobriety, another unintended consequence could be the rise of alcohol and drug-abuse when humans are freed of that responsibility.

与此同时，2014年，因酒精中毒导致死亡的美国人数量是致命车祸的3倍。因此，如果驾车会鼓励人们保持清醒，另一个意外后果可能是当人类摆脱了这种责任后，滥用酒精和药物的人数会上升。

Then there is the trust we have to put in the coders. Normally in the corporate world employers devise elaborate reward and penalty programmes to ensure that workers are incentivised to do the best possible job, even when it’s in their interests to take short-cuts. They are held accountable. In sectors where human sloppiness can have a disproportionately bad effect on others — banking, say, or air traffic control — these sorts of incentives matter even more.

还有就是我们不得不相信程序员。通常，在企业界，雇主会精心设计奖惩计划，以确保员工有动力把工作做得尽可能好，即便偷工减料对他们有利。他们会被追究责任。在人类草率行为可能对他人造成特别糟糕影响的领域(例如银行业或空中交通管制)，这种激励更为重要。

Self-driving cars, however, will be programmed and maintained by coder armies benefiting from safety in numbers when it comes to accountability. Can we be sure that they will always be properly motivated?

然而，自动驾驶汽车将由程序员大军编程和维护，就问责机制而言，这些程序员受益于他们人数众多。我们能够确定他们会一直得到妥善激励吗?

Finally, while there is a good case for self-driving technology to augment human driving skills, there is a risk this could lead to a degradation of those abilities. And we surely would not want to risk discovering that they were no longer there when we really needed them.

最后，尽管自动驾驶技术将增强人类驾驶技能的说法理由充足，但风险在于这可能导致人类能力的退化。我们肯定不想冒这个险：在我们真正需要这些能力时，却发现它们不存在了。