斯坦利诺贝尔（诺奖得主斯宾塞：推进人工智能在国家内部和全球经济中的可获得性和扩散）

4月29日， 诺贝尔经济学奖获得者 Michae l Spence （迈克尔·斯宾塞 ） 出席了“2024中关村论坛——金融科技平行论坛”，并通过视频录像发表了主题为“人工智能对世界经济的变革性影响”的演讲。

斯宾塞教授提出，生成式人工智能是人工智能发展的重大阶段，能够与人类进行更自然的交流，是一种超级通用技术。人类可以通过人工智能的运用来提高生产力，并提升现有劳动力的生产服务水平。他也强调了加强人工智能相关的公共政策和监管措施的制定的重要性，倡导保护版权，推进人工智能在国家内部和全球经济中的可获得性和扩散。

公开资料显示，迈克尔·斯宾塞，1943年11月7日生于美国的新泽西州，1972年获美国哈佛大学博士头衔，现兼任美国哈佛和斯坦福（Stanford）两所大学的教授、青岛大学名誉教授，美国斯坦福大学商学院研究生院前任院长和现任名誉院长。

迈克尔·斯宾塞最重要的研究成果是市场中具有信息优势的个体为了避免与逆向选择相关的一些问题发生，如何能够将其信息“信号”可信地传递给在信息上具有劣势的个体。2001年，与乔治·阿克尔洛夫、约瑟夫·斯蒂格利茨共同获得诺贝尔经济学奖。2017年，林重庚、迈克尔·斯宾塞等的《中国经济中长期发展和转型：国际视角的思考与建议》荣获第17届孙冶方经济科学奖。

图片来源：中关村互联网金融研究院

以下为演讲翻译，经钛媒体APP整理：

各位同事和朋友，大家好，我是迈克尔·斯宾塞，很感谢中关村论坛的组织者给我这个机会作为今年丰富节目的一部分，与大家交谈。

我今天的主题是人工智能革命的变革性经济影响，我们正处于这一革命的早期阶段。GenAI在全球范围内产生了巨大的兴趣和同样大量的投资。我认为可能存在过度乐观的成分，尤其是关于大规模经济、社会、金融体系等影响的到来时间。但我的视角稍长一些，我认为没有人知道这一进程将以多快的速度发生。但我想在到达那里之前，重点讨论一下GenAI的经济和福祉潜力。

我认为值得注意的是，在过去的15年中，人工智能领域取得了一系列重要的突破，并非所有突破都属于我们在早期所称的GenAI范畴。语音识别、手写识别、唇读等是一系列令人瞩目的成就，随后在图像和物体识别领域又取得了一系列重大突破，并拥有众多应用。我们现在正见证这些成果的实现。

尽管人工智能取得了进步，包括图像识别等，但人工智能目前仍不具备像人类一样处理复杂、快速变化的视觉环境的能力。这是关于人类与机器人以及人机协作的相关话题，将在以后的讨论中涉及。

还有其他一些突破，例如DeepMind的AlphaFold，现已成为谷歌的一部分，能够以相当高的准确度确定蛋白质的三维结构。该技术利用了定义蛋白质的氨基酸序列，并已成功用于预测大约2亿已知蛋白质的三维结构，并作为开源数据库发布。因此，全球生物界可以将其作为提高自身研究效率的工具。然后是赢得围棋比赛，人们对其重要性的看法各不相同。但开发人工智能过程中学到的知识非常重要。在许多情况下，人工智能的表现已经超过了人类。人类表现通常是我们评估人工智能的标准。

我们稍后会回到这一点，因为人工智能有从超人智能到次人智能的等级，大致与人类相当，它们都有各自的用途。但为什么通用人工智能可能是一个非常重要的发展？从经济和金融角度来看，有两三个非常重要的特点。我认为通用人工智能就像大型语言模型，具体来说，它首次拥有了你可能会称之为“领域切换能力”或“日常语言”的能力。它知道主题而不需要被告知。它拥有像人类一样理解对话背景的能力，无论对话有多复杂，我认为这是朝着通用人工智能迈出的一步，即使我们还需要一段时间才能到达那里。所以你可以和AI谈论意大利文艺复兴、通货膨胀、早期俄罗斯文学、计算机编码，并让它做数学题，它会很乐意地和你交流，并做出适当的回应。在某些情况下，它会给出令人惊讶的洞察力和聪明的结论。

第二点是易用性。你不需要任何技术培训就可以使用它。这是因为它基本上是在用我们的语言交流，就像我们用语言理解一样。因此，ChatGPT在前两个月就拥有了1亿用户，这是以前从未发生过的事情。

这都意味着什么呢？ 第一个观察是，很难找到一个经济领域，一个知识经济领域，其中没有重要的、可能改变整个地球的突破性人工智能应用。我的观点，也是普遍的观点是，至少有潜力出现一次非常巨大的、长期的生产力提升，这将对增长、供应链、增长的资源约束模式产生影响。

詹姆斯·曼尼克和我在2023年底的《外交事务》上发表了一篇关于人工智能的经济潜力的论文。但重点是，它看起来像是一种超级通用的技术，可以在知识经济中广泛应用，几乎适用于所有地方。而知识经济无处不在。我们通常将它与技术、金融、管理、计算机编程等行业联系在一起。但实际上，知识经济无处不在。医院中也存在知识经济的重要组成部分，比如医生和护士的工作等等。在所有这些情况下，你都可以找到强大的数字助手，既可以帮助个人，也可以帮助以人工智能（GenAI）能力为核心的系统。

为什么这如此重要？因为人工智能（AI）和机器学习（ML）的进步，我们现在可以创建高度智能的数字助手，它们可以帮助我们完成各种任务，从日常工作到复杂的分析和决策。这些助手可以大大提高我们的生产力和效率，使我们能够专注于更重要的任务。此外，它们还可以帮助我们处理大量数据，从中提取有用的信息，为我们提供更好的决策支持。因此，拥有强大的数字助手对于在知识经济中取得成功至关重要。

目前阶段为何如此重要？ 至少在中国，存在总需求不足的暂时性问题，而全球许多经济体则受到供给受限的增长模式困扰。这些问题源于许多因素的共同作用，包括生产率下降、人口老龄化、劳动力减少、不断上升的抚养比、劳动力短缺、关键就业部门的多元化和碎片化成本高昂，以及全球供应链中源自多重来源（包括气候、疫情、金融危机和地缘政治紧张局势）的严重冲击。

最后，我们还面临着一种我试图识别的现象，我称之为“强大通缩力量的消退”。这与那个全球化高度发展、新兴经济体快速增长的时期有关。这有点像我在全球经济中所说的“刘易斯转折点”，当所有这些因素结合在一起，其结果是，我们面临着几十年来未曾有过的新的通胀压力，实际利率上升，财政空间缩小，以及一些地方的债务困境。此外，还需要对能源转型进行大规模投资，以实现可持续的增长模式。如果有人给我们提供了应对所有这些问题的最佳方法，比如与老龄化相关的过度负担等问题，那会是什么呢？

答案是，非常高且持续的生产力增长，而我们最有效的工具就在当前和未来几代人工智能领域。问题是，我们会利用这些工具吗？要实现这一点，需要采取一系列措施。

举几个例子， 斯坦福大学的埃里克·布林森（Eric Brinelson）和他的同事们对人工智能在客户服务中的应用进行了研究，这是全球一个非常大的就业领域。基本上，人工智能通过数千小时的客户服务、客户互动、音频记录以及绩效指标的训练，学会了如何适当地做出回应，从而创造了一个强大的数字助手，作为客户服务代理的助手。他们把这个工具交给了一部分客服人员试用，对其他客服人员则没有提供。

测试结果立刻显现出来。首先，整体上出现了显著的生产率提升，大约为14%。其次，它揭示了人工智能的一些特性：如果观察不那么有经验的客服人员的表现变化，人工智能的影响甚至更大，大约为35%。

事后看来，很容易猜到人工智能实际上是捕捉了与客服、客户互动相关的学习经验，包括哪些做法有效、哪些无效，然后以可用的形式反馈回来。因此，其总体效果是一种“升级效应”，对有经验的客服人员产生了显著的影响，对经验较少的客服人员的影响则更大。但这个故事以及许多类似故事的关键点在于，正确的模式并非仅仅关于自动化和取代人类，而是强大的数字助手。

当机器在某些方面做得更好时，一些事情将会被自动化。就像总结他们在大量数据集（在此例中为音频记录）中发现的模式一样。但这并不意味着你将人类从剧本中抹去，而是为（甚至是完整的系统）提供一个强大的数字助手。将数字助手与人类表现进行比较是完全自然的方式，这并没有什么错，它可以帮助评估我们取得了多大的进步。但这确实会导致一种自动化偏见。原因在于，一旦AI达到平均人类水平的表现，人们就会倾向于认为“为什么不用AI取代人类呢？”需要仔细思考才能意识到，除非是在考虑某些特定的自动化方面。

AI存在于各个层面。那些超人类的AI能够做人类无法做到的事情，比如在大量数据中识别模式及其含义，以我们无法达到的速度进行高速计算。因此，这是一个子集，它们非常重要。在这个领域，我们将看到增强，但它将是通过增强人类而不是取代人类来实现的。它只是增加了一些东西。

还有一类事物，它们在某些方面与人类的能力相当。还有另一类子集，我想花点时间谈谈，虽然AI在某些方面确实无法达到我们所期望的人类水平，但它们仍然可能非常有用。这里的总体方向是包容性增长模式。

一些研究人员发现，利用皮肤癌的图像来训练AI实际上可以检测出皮肤癌。它们和您经常看的皮肤科医生一样好吗？如果您容易出现这种问题，答案可能是否定的。因此，我的学生通常认为这并不太有趣。这是个不错的尝试，但我们没有成功。但问题是，这不是正确的思考方式。世界上可能有85%的人口远离皮肤科医生，无法获得最佳治疗。但是，如果人工智能能够相当有效地从仅用普通手机拍摄的照片中检测出皮肤癌，那么只需向人们发出足够好的信号，就能触发他们的反应，让他们乘坐火车，前往80公里外的皮肤科医生那里就诊，而不至于太晚。

在信贷、金融、电子商务等领域，有很多应用与增长模式的包容性有关，通过人工智能驱动的算法，可以极大地扩展服务范围。我认为，不要让基准成为决定经济应用的因素。

最后，我想简单总结一下，与此相关的政策议程极为复杂。其中一些与防止破坏性或有害的滥用有关，例如虚假信息、大规模的宣传活动和欺诈等。

人们担心就业问题。我们会有足够的工作机会吗？就我个人而言，这是一个长期的争论，但我并不太担心平均就业率下降的问题。我认为我们不会在为人们创造工作机会方面遇到问题，但在微观经济层面上，这将是非常具有破坏性的。它将改变工作和技能，有些职业会衰退，而另一些职业则会增长。

这些转变对于人们来说并不容易。政府和政策在这方面可以发挥作用，以减轻和加速这些转变，从而避免对个人或家庭以及工人造成经济损失或过度焦虑。

版权保护是一个重大问题。GenAI模型会吸收大量来自各个领域的文字和创意作品的数据和信息。这些作品的原创者需要在财产权方面考虑某种形式的保护。这是一个非常棘手的问题，目前仍在进行中。但在积极的一面上，这个问题不能被忽视。如果我们想要实现生产力的飞跃，就需要在国内乃至全球经济范围内实现技术的普及和推广，这需要政府和公共政策的推动。我毫不怀疑我们将拥有一些非常先进的行业。

技术和金融领域可能会发展得相当迅速，管理的各个方面也将迅速发展，尤其是那些拥有资源可以探索和尝试新技术的大公司。但问题是，我们能否将技术普及到整个经济，并应用于那些在数字化采用方面往往落后的行业？

对于那些没有大量资源的小型和中型企业来说呢？我们确实需要解决这个问题，不仅从包容性角度出发，因为如果不能实现技术普及，我们就无法实现生产力飞跃。如果影响只局限于少数可能已经准备好接受新技术的行业，那就毫无意义了。

我们需要制定旨在促进技术普及和提高可访问性的政策。再次强调，这是政策议程的积极方面。我很担心，至少目前在西方，政策议程的重点非常倾向于防止负面影响，而不是加速并确保我们获得积极影响。

但是，简而言之，我认为科学和技术界出于某种原因为我们提供了一套工具，这些工具在很大程度上是可以获得的，可以广泛应用，具有巨大的经济、社会、医疗、教育和其他潜在价值。

目前正处于深入探索和实验的初期，很难准确预测其长期发展情况，但看起来潜力巨大，因此，参与其中的这十年将是令人兴奋的十年。

最后，请允许我指出，人工智能的进步远未结束。我认为我们将看到更多令人惊叹的突破，这些突破发生在一个扩大了的可能性范围内。例如，我们似乎正处于人工智能与生物医学科学交汇的初期。如果这是真的，它将加速正在进行的生命科学革命。所以总的来说，人们可以在这些事情上花费数小时的时间，但我希望我已经充分地证明了人工智能将成为全球经济，尤其是美国经济的积极推动力。我认为人工智能领域的另一大玩家显然是中国。

当前世界正处于一个充满挑战的时期，经济增长放缓，各种问题使得生活更加艰难，降低了经济绩效。在我看来，人工智能是推动我们朝着相反方向前进的一束亮光。

非常感谢。

以下为演讲原文：

Greetings, colleagues and friends, Im Michael Spence. Im grateful to the organizers of ZGC forum this year for giving me the chance to speak with you as part of this years very rich program.

My subject today, is the transformative economic impact of the revolution in artificial intelligence, in which we find ourselves in the early stages. GenAI is generated a huge amount of interests globally, and an equally large amount of investment. I think there may be elements of excessive optimism, especially about the timing of the arrival of large impacts on economies, societies, financial systems, and so on. But I have a slightly longer time horizon. I dont think anybody knows how fast this is going to occur. But I want to focus on the economic and welfare potential of GenAI before we get there. I think its worth noting that there were a set of important breakthroughs in the past 15 years in artificial intelligence, not all of which fit the category head under GenAI we had in the earlier days. Speech recognition, handwriting recognition, lip reading, which was extraordinary set of achievements, and then a very big and important set of breakthroughs in image and object recognition with a huge host of applications, that we are now seeing coming to fruition.

In spite of the progress of the AI, including image recognition and so on, AI dont yet have a human like ability to process very complex, rapidly evolving visual environments with no latency. Thats a relevant subject for later discussion that has to do with the relationship between humans and robots and human robotic collaboration.

There were other breakthroughs, Alpha Fold, product of Deep Mind, now part of Google in determining the three dimensional structure of proteins with reasonable accuracy. Using the amino acid sequence that defines a protein that work was successful and has now been used to predict the three dimensional structure of approximately the 200 million known proteins, and then published as an open source database.So the biological community globally can use this as a productivity enhancer in their own research.

Then there was winning the games of Go, which many people have different opinions about the importance of that. But the learning that went along with developing an AI was extremely important. In many cases, now, AI exceed human performance. Human performance is usually the benchmark by which we assess AI.

Well come back to that because theres a kind of hierarchy of AI from superhuman to subhuman with roughly at a par with human, and they all have their uses. But why is GenAI potentially such an important development? From an economic and financial point of view, there are two or three characteristics that are really important. I think of GenAI as large language models and just to be concrete for the really the first time has what you might call a domain switching capability or an ordinary language. It knows what the subject is without being told. It has a human like ability to understand the context of a conversation, no matter how complex, which I think is a step in the direction of artificial general intelligence, even if it takes us a while to get there.

So you can talk AI about the Italian renaissance and inflation, early Russian literature, computer coding, and ask it to do math problems, and it doesnt have any trouble going along with you and to respond appropriately. In some cases, with sort of surprisingly insightful and clever results. Second is accessibility. You dont really need technical training to use it. And thats because it essentially talks our language and understands in the same way we understand using language. So ChatGpt had a hundred million users in the first 2 months that just never happened before. What does it all mean?

First observation is that its very hard to find a part of the economy, the part of the knowledge economy that where there isnt an important set of potentially transformative applications of the whole planet play of breakthroughs in AI and so my belief and one that is widely shared is that theres at least the potential for a very huge, extended productivity surge, which will have impacts on growth, supply chain, supply constraint patterns of growth. James Manic and I made this argument, and wrote a paper on foreign affairs late 2023 on the economic potential of AI. But the main point is, it looks like a super general purpose technology that can be used in the knowledge economy, pretty much everywhere. And the knowledge economy is everywhere. We associate it with sectors like technology and finance, management, computer coding, et cetera. But in fact, the knowledge economy is everywhere. Theres important parts of the knowledge economy in hospitals and what doctors and nurses do and so on. And in all of these cases, you can find powerful digital assistance, both to in individuals and the systems that have as at their core, i.e. the GenAI capabilities.

Why is this so important at current stage?

Theres at least a transitory problem of insufficient aggregate demand in China, whiel much of the global economy is suffering from a supply constrained patterns of growth. These are the result of a lot of things coming together, declining productivity, declining productivity, aging, declining labor forces, higher and rising dependency ratios, labor shortages, in key large employment sectors, expensive patterns of diversification and fragmentation in global supply chains that are the result of severe shocks with multiple sources including climate pandemic, financial distress, geopolitical tensions.

And finally, we have a pattern that Ive tried to identify, which I call the fading of the powerful deflationary forces that were associated with that period of hyper globalization and very rapid emerging economy growth. Its like what I call a kind of Lewis Turning Point in the global economy when all of those things come together. The effect is that we have new inflationary pressures that we havent had for three or four decades, rising real interest rates, declining fiscal space, debt distress in a number of places.

And alongside of that, the need for very large investments in the energy transition, in pursuit of sustainable growth patterns. Whats the best anecdote that we could possibly have if somebody gave it to us to deal with all these things, excessive burdens on the young associated with aging and et cetera.

The answer would be, a period of very high, sustained productivity growth, and the most powerful tools we have to engineer that are very prominently in the area of the current and future generations of artificial intelligence. The question is, are we going to take advantage of this? A number of steps are required to do that? Let me mention a couple of examples.

There was a study done by Eric Brinelson and his colleagues at Stanford of the application of artificial intelligence and customer service, application or industry sector, which is a very large employment sector globally. Basically, AI was trained on literally thousands of hours of customer service, customer interactions, audio recordings, along with performance measures. And it learned how to respond appropriately, which created a sort of powerful digital assistant as an assistant to the customer service agent. They gave it to a subset of the customer service agents and knocked others and tested it.

The two conclusions emerged immediately. One, there was a very large productivity increase overall on the order of 14%. And the second one, tells you something about artificial intelligence is if you looked at the impact on performance of the less experienced customer service agents, the impact was even larger on the order of 35%.

With the benefit of hindsight, its fairly easy to guess that what the AI is doing basically is capturing the learning that is associated with customer service, customer interactions, what works and what doesnt, and then delivering it back in a usable form.

So the net effect of that is a kind of leveling up effect, getting a noticeable impact for the experienced agents and a much bigger impact for the less experienced and impacted. But the main point of that story and many others like it is that the right model, notwithstanding a very powerful tendency to believe that this is really just all about automation and getting rid of human beings. The right model is the powerful digital assistant.

Some things will be automated when the machines are better at it. Like summarizing patterns that they finds in very large collections of data, in this case，audio recordings. But that doesnt mean youve written a human being out of the script, as opposed to giving the human being or even a full system a powerful digital assistant. Digital assistants are benchmarked against human performance is a perfectly natural way. Theres nothing wrong with this to get to assess how much progress weve made. But this does lead to a kind of automation bias. And the reason is that once AI passes the average human performance. The tendency is to think why dont we get rid of the human and just use the AI. It takes a little bit of careful thought to realize thats probably not the right answer unless youre thinking about very partial aspects of automation.

AI’s come at all levels. The superhuman ones can do things that human beings just cant do in sort of recognizing patterns and their implications in vast quantities of data that are involved. High speed calculations of at a level that we just cant accomplish. So thats a subset. Theyre very important. And in that area, we will see augmentation, but it will be augmentation not by replacing humans. It will just adding something.

Then theres not a set of things that they do on a par with humans. Theres another subset that I want to just spend a minute on which AI dont really quite measure up to the kind of human performance that we would hope for, but are still potentially very useful. The general heading here is inclusive growth patterns.

Some researchers have determined that using images of skin cancer to train AI can actually detect skin cancer. Are they as good as the dermatologists that you visit regularly? If you were prone to this kind of problem, the answer is probably no. So my students normally think thats not very interesting. It was nice try, but we didnt make it. But the problem is that its not the right way to think about it. There probably is 85% of the worlds population that doesnt live near a dermatologist and cant get the first best version of the treatment. But if AI is reasonably good at detecting skin and cancer from images that are just taken with an ordinary mobile phone, you could very easily have a significant increment to the preventive health care by giving people signals that are good enough to trigger a response and get them to get on a train and go 80 kilometers and visit a dermatologist before its kind of too late.

Theres a lot of applications in credit and finance in e-commerce and so on. That have to do with inclusiveness of growth patterns where you can extend the range of service enormously with AI-powered algorithms. And I think its important not to let the benchmarking be the determination of what the economic application is.

Finally, Im going to conclude by saying a few words. Theres an enormously complex policy agenda associated with this. Some of it is associated with preventing destructive or damaging misuse, e.g. disinformation, campaigns, fraud at a massive scale.

Theres concern about jobs. Will we have enough jobs? On my side, this is a longer argument, but I’m not too worried about the net average, a problem with job loss. I dont think well have a problem generating work for people, but at a more microeconomic level, it will be very disruptive. It will change work and skills. There will be some occupations that decline and others that grow.

And those are not easy transitions for people. And there is a role for government and for policy in easing and making and accelerating. Those transitions in a way that they dont produce both economic damage at the individual or household level, at the level of the worker on excessive anxiety.

Theres a big issue of copyright protection. The GenAI models just vacuum up enormous and vast quantities of data and Information of writing and creative work in various areas. And the original creators of that work need to consider some kind of protection in terms of property rights. This is a very hard problem and it is a work in progress. But it can’t be simply ignored on the positive side. If were going to get the productivity surge, we need accessibility and diffusion within countries and eventually in the global economy. And that takes government and public policy to get there. And I dont have any doubt that well have very advanced sectors.

The tech sectors finance will probably move fairly fast. Various parts of management will move quickly, especially among the bigger firms that have the resources to explore and experiment with this. But the question is, are we going to get it across the economy and set into sectors that tend to lag in terms of digital adoption?

In small and medium sized businesses that dont have the massive resources available? We really need to address that problem, not just from an inclusion point of view, but because we wont get the productivity surge. If the impact only comes in a few may sectors that where its very likely that itll get adopted anyway.

We need policies that are designed to increase to make diffusion work, to make accessibility easier and so on. And again, thats the positive side of the policy agenda.And Im worried, at least in the west at the moment, the policy agenda is very heavily weighted toward what you might call preventing the negative side, and not so much weighted toward accelerating and ensuring we get the positive side.

But to summarize, I think we have been given for whatever reason by the scientific and technological community, a set of tools that are accessible for the most part can be made widely available that have a huge economic and social and medical and educational and other potential.

It was in the early stages of intense exploration and experimentation. Its very hard to know exactly how this will play out over time in detail. But it certainly looks like the potential is enormous, so itll be an exciting decade to be involved in all of this.

Finally, just let me mention that the progress in AI is far from over. I think we will see more breathtaking breakthroughs that come within an expanded set of potential. For example, we seem to be in the early days of the intersection of artificial Intelligence and biomedical science. If thats true, it will turbocharge the revolution in life sciences thats already underway. So bottom line is one can spend hours on these things, but I hope Ive made a pretty convincing case that this is going to be one of the positive driving forces in the global economy, certainly in the United States. And I think the other major player in artificial Intelligence is obviously China.

Its in a pretty difficult world with slowing growth and lots of things that are making life more difficult and reducing economic performance. This seems to me the bright light thats pushing us in the opposite direction.

Thanks very much.（本文首发于钛媒体APP，作者｜颜繁瑶，编辑｜刘洋雪）

更多宏观研究干货，请关注钛媒体国际智库公众号：