05 欧盟“上游治理”式的人工智能监管提案只会阻碍创新
麻省理工学院麦克菲：欧盟拟议法规不仅会威胁到学外语应用“多邻国”最出色的创新，还将降低类似创新出现在欧盟的可能性。
全文共722个词, By Andrew McAfee
While some of us under lockdown churned through streaming services and sourdough starters, others decided to use the time for a little self-improvement — taking up Dutch or Danish, Swahili or Esperanto. Duolingo, the free app many downloaded, has become the world’s most popular way to learn a second language. The company is now hoping to ride that interest into an initial public offering: last week it said it wanted to be valued at up to$3.4bn in its IPO.
当我们中的一些人在封锁期间忙着享受流媒体服务和鼓捣酸酵头时，其他人决定利用这段时间来进行一点自我提升——学习荷兰语或丹麦语、斯瓦希里语或世界语。很多人下载的免费应用“多邻国”（Duolingo）已成为全球最受欢迎的学习第二语言的渠道。如今，该公司希望借助人们的这股兴趣进行首次公开发行（IPO）：上周该公司表示，希望IPO估值能达到34亿美元。
But EU proposals for regulating AI threaten the use of one of Duolingo’ s niftiest innovations, the English Test, in its current form. They also make it less likely that the next round of similar innovations will be developed in the bloc. That’s a problem.
但欧盟（EU）监管人工智能（AI）的提案，威胁到“多邻国”最出色的创新之一——“英语测试”（English Test）的使用。它们还将降低下一轮类似创新出现在欧盟的可能性。这就成问题了。
The English Test provides a way for people to demonstrate their proficiency to more than 3,000 educational institutions around the world. Test-takers don’t need to register in advance or travel anywhere; they just need an internet-connected device with a webcam and an hour to spare. The test guards against cheating (that’s what the webcam is for); assesses literacy, conversation and comprehension; returns results in two days; and costs less than $50.
“多邻国”英语测试为人们提供了一种向世界各地3000多所教育院校展示自己语言能力的方式。测试者无需提前报名或前往任何地方；他们只需要一个带有网络摄像头的联网设备和一个小时的空闲时间。该测试可以防止作弊（这正是网络摄像头的作用）；评估读写、会话和理解能力；两天内便可收到结果；花费只需不到50美元。
It’s also a high-risk AI system, according to the EU proposal. This label applies because the test uses AI, both for personalisation — questions appropriate to the taker’s skill level are generated on the fly — and for grading. Systems that use AI for “assessing participants in tests commonly required for admission to educational institutions” are put in the high-risk category by the EU’s proposal.
而根据欧盟的提案，它也是一个高风险的人工智能系统。之所以给它贴上这一标签，是因为该测试使用人工智能来进行个性化服务（即时生成适合测试者技能水平的问题），并进行评分。使用AI“评估教育机构入学考试参与者”的系统，被欧盟上述提案列入高风险类别。
Providers of high-risk AI systems have to take a number of steps under the proposed regulations related to a long list of factors: data and data governance, transparency, oversight, robustness, accuracy and security, among others. The requirements for high-risk AI systems and the obligations placed on their providers take up 10 pages in the proposal. And as far as I can tell, all these requirements and obligations would kick in before the systems can be offered in any form: before a first beta offering is released or before the provider has any indication of product-market fit. Or whether there’s a market at all.
根据拟议的法规，高风险AI系统的提供商必须采取一系列措施——这些法规涉及一长串因素：数据和数据治理、透明度、监督、鲁棒性、准确性和安全性等等。在提案中，针对高风险AI系统的要求及其提供商应承担的义务占了10页内容。据我所知，在这些AI系统以任何形式推出之前，必须满足所有这些要求和义务：在首个beta版本发布之前，或在提供商透露任何有关产品与市场匹配方面的信息之前。或者究竟有没有市场。
It’s a safe bet that AI-using entrepreneurs and early-stage investors — which these days means essentially all tech entrepreneurs and investors — will balk at these expensive and time-consuming requirements and direct their energies away from high-risk application areas. The EU, then, will generate less tech innovation in, among other important activities,admitting students to schools and grading their exams, making hiring and promotion decisions, establishing creditworthiness, dispatching first responders, letting prisoners out on parole and using analytics to fight crime. On the other hand, Europeans will know that all AI used in these areas has gone through an extensive vetting process.
可以肯定的是，AI使用领域的企业家和早期投资者——如今这基本上意味着所有科技企业家和投资者——将回避这些昂贵且耗时的要求，并将精力投入到高风险的应用领域之外的地方。然后，欧盟在各种重要活动方面的技术创新将减少，包括学校招生和考试打分、做出招聘和提拔决定、建立信誉、派遣急救员、允许囚犯假释和利用分析打击犯罪。另一方面，欧洲人将知道，这些领域使用的所有AI都经过了广泛的审查程序。
Your views about that trade-off indicate which of the two main schools of thought about tech regulation you belong to. One school holds that precisely because some techs are so powerful, they must be rolled out with a lot of upfront planning and continued oversight.This is the “upstream governance” approach, of which the EU’s proposed AI regulation is a clear example.
针对这里面的权衡取舍，你的看法将表明你属于两大科技监管派别的哪一派。一种观点认为，正是因为一些技术如此强大，因此它们必须经过大量前期规划才推出，而且要持续接受监督。这就是“上游治理”方法，欧盟拟议的AI监管就是一个明显的例子。
Advocates of the other approach, often called “permissionless innovation”, say the activities listed above as important need to be exposed to lots of attempts to improve them,even wacky ones that come from unknowns and outsiders on a shoestring budget (like Duolingo when it started). That’s the best way to find real breakthroughs. Restricting the field of potential innovators to those who can afford high upfront costs is a bad idea. It leads to slower progress and growth and fewer hometown success stories, which are also risks.
另一种方法——通常被称为“无需经过许可的创新”——的倡导者认为，上面所列的那些重要活动需要通过大量尝试——即便是那些由预算有限的无名之辈和外行人发起的古怪活动（就像推出之初的“多邻国”）——来改善。这是发现真正突破的最好方法。将潜在创新者的领域局限于那些能够承担高额前期成本的人不是一个好主意。这会导致进步和增长的步伐放缓，本土成功故事减少，这些同样是风险。
These risks aren’t just theoretical. The EU’s General Data Protection Regulation, which came into effect in May 2018, has been accompanied by increased market share for Google and reduced venture investment in Europe. And the benefits to the EU of all the extra governance are not obvious; the EU’s own progress report, released in June 2020, found fragmented enforcement and insufficient resources.
这些风险不仅仅存在于理论中，欧盟2018年5月实施《通用数据保护条例》（GDPR）以后，谷歌（Google）市场份额增加，欧洲的风险投资减少。所有额外增加的治理给欧盟带来的好处并不明显；欧盟2020年6月发布的进展报告发现，执行工作支离破碎，资源不足。
The EU is a rich and well-educated region with great technological strengths. Yet it is lagging as we move deeper into the “second machine age”, and by some measures lagging badly. There are many reasons why. One of them, I believe, is that more upstream governance translates to less downstream innovation.
欧盟是一个富裕、受教育程度高的地区，拥有强大的技术优势。然而，随着我们更深地步入“第二次机器时代”，它已经落后了，而且在某些方面严重落后。原因有很多。我认为其中一个原因是，上游治理增多导致下游创新变少。