非正规经济中的信任货币与信贷鸿沟
我的成长经历源于阿根廷一个小型企业主家庭。我的父母经营着一家窗帘和地毯店，这让我亲身体验了企业发展的艰辛。社区的信任与支持是维系生计的关键。然而，我最终选择不继承家族生意，转而投身政治学研究。我着迷于技术如何能够支持像我父母那样的企业成长。这份好奇心引领我来到了麻省理工学院（MIT），并在2019年，我的人工智能（AI）与经济发展方向的硕士论文获得了现实世界试点项目的资助。这便是我的研究最终落脚于哥伦比亚非正规社区的开端。
在我进行研究的这些社区里，购买午餐并不需要信用卡。店主只需认识你便足够了。如果你的母亲在贷款方面有良好记录，如果你早晨会主动问好，如果你经营的店铺为邻居所熟知，他们就会赊给你大米、甘蔗或面包。那里的经济并非仅靠现金运转，它依靠的是信任——一种随时间积累起来的无形货币。我注意到，我在阿根廷成长过程中所见的那些原则，在哥伦比亚的企业中同样鲜活存在。在许多拉丁美洲社区，信任一直是他们最强大的货币，是一种良好的声誉。
然而，一个矛盾出现了。当这些社区中的个体转向银行申请贷款以发展生意时，他们却被拒绝了。银行会告知他们：“你没有抵押品，没有金融历史，我们无法证明你的身份。”在拉丁美洲的许多社区，情况正是如此。在整个拉丁美洲，半数人口被排除在正规信贷体系之外。在金融普惠与城市发展交叉领域工作了十年后，我将毕生精力投入到回答一个问题：是什么让你在社区中获得信用，这种信任，能否同样让你在银行眼中获得信用？ 你的承诺能否成为风险评估的一部分？我们能否通过衡量一个人的潜力来扩大资本的可及性？信任，能否被AI所衡量？
Original English
I grew up in a family of small business owners in Argentina. My parents ran a curtain and carpet shop, so I witnessed first-hand how difficult it is to grow a business. Trust and support from their community were key in keeping the business alive. But I decided not to continue with my family business. Instead, I studied political science. I was obsessed with how technology could support the growth of businesses like my parents'. And that curiosity led me to MIT, where in 2019, my Master’s thesis in AI and economic development got awarded funding to become a real-world pilot. And that's how I ended up working in informal settlements in Colombia. In these neighborhoods where I did my research, you didn't need a credit card to buy lunch. It was enough for the shopkeeper to know who you were. If your mother had a good record with loans, if you said hello in the mornings, if you had a shop that was known by the neighbors, they will front you the rice, the sugarcane, the bread. The economy didn't run solely on cash. It ran on trust. That invisible currency that is built over time. And I noticed something. Those same principles I saw growing up in Argentina were alive in Colombian businesses, too. In many Latin American neighborhoods, trust has always been the strongest currency. A good name. But here comes the contradiction. When this same person goes to a bank and asks for a loan to grow this business, they will be rejected. They will tell them, “You don’t have a collateral. You don't have a financial history. There's no way we can prove who you are." In many Latin American neighborhoods, this is the case. And in Latin America, half of our population is excluded from formal credit. After a decade working at the intersection of financial inclusion and urban development, I dedicated my life to answer one question. What if what makes you credit-worthy in your neighborhood, trust, could also make you credit-worthy in the eyes of a bank? What if your word could be part of the risk assessment? What if we can scale the access to capital by making your potential measurable? What if trust could be measured with AI?从可见性困境到金融排斥的本质
从小到大，我一直梦想改变世界，这也是我选择学习政治学的原因。我曾以为可以通过政策来实现这一目标，但很快意识到政策的推进速度无法满足人们的迫切需求。于是，我转向了技术。技术不受地理界限的限制。在MIT，我和同学们开始着手一个本地化项目，旨在为社区定义本地市场，创建平台供他们上传待售商品，从而在社区内获得可见性。我们开始拜访这些企业，帮助他们上传更多产品图片到市场平台，以提高知名度和销量。
然而，我们注意到他们的销售额并未增长。当我们询问原因时，答案非常简单：他们没有足够的资金购买更多原材料。尽管他们经营这些生意多年，却无法获得更多库存，也无法获得营运资本来采购更多商品。这时我们意识到，我们面临的并非是可见性问题，而是金融排斥问题。我越深入研究，越发现一个我们常常忽略的事实：贫穷本身是极其昂贵的。当你只能小批量购买商品时，单价会更高。如果你无法一次性购买整瓶洗发水，就只能买小袋装；如果无法购买一周的杂货，就只能按天购买，最终支付的总价总是更高。
而在信贷领域，这种成本甚至更高。当你没有信用记录或银行账户时，唯一的选择就是求助于掠夺性贷款人（predatory lenders），即“gota-a-gota”（高利贷者）和高利贷集团，他们以极其高昂的代价提供贷款。他们不要求任何文件，却可能收取每周甚至每天20%的利率，并且手段暴力且充满虐待。
我将讲述玛丽亚（Maria）的故事。她是一位委内瑞拉移民，居住在哥伦比亚的一个低收入社区。她制作精美的手工包，并接受客户的定制订单。在收款之前，她需要先购买材料来完成订单。由于玛丽亚是移民，她没有银行账户，也没有任何信用记录，因此购买材料的唯一途径就是向这些极其危险的掠夺性贷款人借钱。不幸的是，玛丽亚在拉丁美洲并非个例，她实际上是普遍现象。拉丁美洲有数百万这样的微型企业（microbusinesses），从街角小店到餐馆、美容院，几乎所有企业都是微型企业。99%的拉丁美洲企业是微型企业，它们贡献了三分之一的GDP。但即便如此，它们仍无法从银行获得哪怕一美元的贷款。原因在于，它们缺乏金融系统赖以建立的那些“文件”。
Original English
So before I tell you more, I want to share a little bit of how all this started. Since I was a child, I dreamed of changing the world. And that's why I studied political science. I thought I was going to do it through policy. But then I realized policy was not moving at the speed people needed to. So I turned to technology. Technology doesn't recognize any geographic boundary. So at MIT, my classmates and I started working on a local project to define local marketplaces for communities, platforms where they can upload what they are selling and become visible in their community. We started visiting these businesses to help them to upload more pictures of their products into the marketplace and become known and start selling more. And we noticed that they weren't growing their sales. So when we asked them why, their answer was very simple. They didn't have enough money to buy more supplies. Even though they were running these businesses for years, they couldn't get more inventory. They couldn't get any access to working capital to buy more inventory. So we noticed something. We were not facing a visibility problem. We were facing a financial exclusion problem. And the deeper I went, the more I learned something that we usually don't say enough. Being poor is very expensive. Products cost more when you can just afford them in small quantities. If you can't buy a whole bottle of shampoo, you end up buying a sachet. If you can't buy groceries for the whole week, you end up buying by the day, and you always end up paying more. And when it comes to credit, in the financial sector, the cost is even higher. When you don't have a credit history or bank account, your only option is to access the predatory lenders. The "gota-a-gota", the loan sharks, and they come at brutal cost. They don't ask you for paperwork, but they could charge you 20-percent interest rate per week, even per day, and they are violent and abusive. So I will tell you the story of Maria. She's a Venezuelan migrant living in a low-income neighborhood in Colombia. She makes these beautiful handcrafted bags, and she gets custom orders from her clients. So before she sells and she gets paid, she needs to make the order. So she needs to buy the materials to make that order happen. As Maria is a migrant, she doesn't have a bank account, she doesn't have any credit history, so her only option to buy those materials is to ask for money from these predatory lenders that are really, really dangerous. Unfortunately, Maria in Latin America is not the exception. She's actually the rule. She's the rule in Latin America. Millions of microbusinesses. Microbusinesses like hers are everywhere. They are from the corner shop to the restaurant to the beauty salon. Actually, almost every business in Latin America is a microbusiness. Ninety-nine percent of our businesses are micro, and they contribute one third of our GDP. But still, they cannot even access one dollar from a bank. Why? Because they don't have the paperwork the financial system was built to require.AI驱动的信用评分：发掘隐形数据价值
玛丽亚可能没有信用记录，也没有银行账户，但她拥有一部手机。而这正是我们看到的机遇所在——不是改变他们是谁，而是改变他们如何被看待。当我们开始时，关于这个经济体以及我们希望帮助的这部分人群，并没有现成的数据。而这正是AI面临的主要问题之一：模型只能预测它们已经见过的内容。因此，我们认识到，如果想开始帮助这部分人群，我们就必须自己构建数据集。由于我们谈论的这部分人群是非正规企业家（informal entrepreneurs），他们没有记录，没有数据，因此在系统中是隐形的。
在传统银行业务中，发放贷款的方式通常是：风险官亲自上门，实地考察企业，与邻居交谈，确认企业确实存在，然后基于他们的经验做出决策。然而，这种经验通常带有偏见，主观性强，而且效率极其低下。因此，在我们开始构建数据集的阶段，我们实际上是在搭建本地市场平台，让人们上传他们销售的商品。我们注意到，这些图像本身就充满了经济信号：我们可以看到是否有顾客在后面，产品是否是手工制作，该产品或服务在该社区是否有销售潜力。数据是存在的，只是不符合银行习惯读取的格式。
于是，我们开始构建数据集，从小规模做起，非常小。我们开始发放10美元的贷款，这足以让企业家补充库存，也足以让我们开始增长数据集。我们非常有针对性地选择贷款对象，我们服务的对象中有一半是女性，因为如果我们希望AI是公平的，它就需要向所有人学习。像玛丽亚这样的手工艺人，她可能没有信用记录，但她的手机里充满了关于她日常经济活动的线索。她有一个Facebook页面，用于上传她销售的产品；她接收文本订单；她多年来一直使用这部手机，里面有她产品的视频。
基于此，我们构建了一套AI驱动模型（AI-powered models），将这些隐形数据转化为一种金融身份（financial identity）。我们处理的数据非常广泛，但我将重点介绍三个我们自主研发的专有评分模型。其中一个主要模型关注文本消息，包括短代码短信，从中获取账单支付、订单确认、手机充值等数字钱包或银行账户中的交易信息。通过使用**大语言模型（LLM model）**和机器学习，我们可以识别出收入、支出以及每月可支配余额的模式。这是一种开放银行（open banking）的模式，但我们使用的是电信数据而非银行账户数据。
另一个模型利用视频。它取代了风险官通常需要花费大量时间且成本高昂的上门拜访。我们让用户发送一分钟的业务视频，解释他们正在做什么。通过计算机视觉（computer vision），我们可以获取他们的库存量、语气、对业务的描述、地理位置、业务类型以及其潜力。我们从中检测他们的支付意愿。最后，我们开发了一个连接用户社交媒体的模型。目前，大多数企业，即使是非正规的，也都在线。他们拥有Facebook页面或Instagram账号。因此，当他们申请贷款时，会授权访问他们的社交媒体，我们可以获取他们的视频、图片。我们再次使用计算机视觉技术，并结合分析点赞数、评论、互动情况、个人资料简介等。我们发现，拥有强大社交媒体和在线形象的企业，其还款概率更高。
Original English
So Maria might not have a credit history, she might not have a bank account, but she has a phone. And there's where we saw the opportunity. Not to change who they are, but to change how they are seen. So when we started, there was no data about this economy and this segment of the population we wanted to help. And, you know, that's one of the main problems with AI. Models can only predict what they have already seen. So we understood that if we wanted to start helping this population, we needed to build a data set ourselves. As this population we're talking about are informal entrepreneurs, then there's no record, there's no data. So you become invisible to the system. So in traditional banking, the way they give out a loan is usually, you know, the risk officer goes to the house of the person, checks the business with their own eyes, talks with the neighbors, sees if actually that business exists and they make the decision based on their experience that usually comes with bias, it’s subjective, and it’s really slow. So at that point, when we started to build the data set, we were actually building the local marketplaces where people were uploading the products of what they were selling. And we noticed that the images themselves were full of economic signals. We could see if there were customers in the back, if the product was handmade, if there was potential for that product or service to be sold in that neighborhood. So the data was there, but just not in the format that the banks were trained to read. So when we started building the data set, we started small. Super small. We started giving out 10-dollar loans, just enough for entrepreneurs to refill their inventory and enough for us to start growing the data set. And we were very intentional to whom we were giving the loans. Half of the people we were serving were women, because if we want AI to be fair, then it needs to learn from everyone. So people like Maria the artisan, they might not have a credit history, but she has a phone that is full of clues about her daily economy. She has a Facebook page where she uploads the products she's selling, She has, you know, text orders that she's receiving. She has had this phone for years, she has videos of the products in her phone. So we built a suite of scores, AI-powered models, that take these invisible data into a financial identity. This is all the data we are processing. But I will concentrate on three specific scores that are proprietary and that have been done by us. One of the main scores we have is looking at text messages, short-code text messages, where we are getting bill payments, order confirmations, mobile recharges, any transactions that have been done in digital wallets or bank accounts. And by using an LLM model and machine learning, we can detect patterns of income, of spending, of disposable, available balance per month. It's a kind of open banking, but instead of using a bank account, we are using telecom data. Another score we have developed is using videos. We replace that visit that usually the risk officer is doing to the houses of people, that is usually very expensive and it takes a lot of time. We replace it by users sending a one-minute video of their business, where they explain what they are doing, and using computer vision, we can get their stock, their inventory, their tone of voice, what they are saying about their business, their localization, the type of business, and all the potential that it has. We are detecting their willingness to pay. And lastly, we developed one that is connecting into their social media. Right now, most of businesses, even if they are informal, they are present online. They have a Facebook page or they have an Instagram. So when they apply for the loan, they sign up into their social media and we can get their videos, their pictures, So we use, again, computer vision, the same one we did for the other type of videos. But also we get the likes, the comments, the engagement they are having, their profile bio, and we detected that a business that has a really strong social presence and online presence has more probability to pay back.AI重塑金融体系：公平、高效与个性化服务
所有这些数据汇入我们的模型，我们从中识别模式和信号，以判断一个人是否值得信赖，即使他们从未获得过贷款。经过三年的实践，我们已经超越了简单的“是”或“否”的判断。我们能在几秒钟内做出决定，还能确定他们能偿还多少金额、何时偿还以及在何种条件下偿还。这使得我们能够模拟利率、分期期数，还能检测季节性影响。因此，我们能够提供真正支持人们日常需求的信贷，并且是为他们量身定制的。我们不是试图向每个人推销同一种金融产品，而是真正理解他们的业务需求。
我们已经验证了这种方法。经过这三年的努力，我们证明了可以利用这类数据来理解非正规经济部门。我们的业务和模型达到了超过0.83的准确率，这符合市场标准。我们已经服务了超过26,000名企业家。我们的模型经过训练，使用了超过150,000个非正规企业家的数据样本，包含数百万个数据点。
这不仅支持了企业家及其家庭，更在改变着金融体系。过去需要数年才能建立的信用记录，或者根本不存在的信用记录，现在只需几个月即可完成。我们正在构建一个实时的金融健康监测系统，可以每日更新，这样就不需要等待多年才能获得贷款资格。这使得非正规经济部门能够首次获得来自正规银行系统的贷款。
人工智能（AI）并非魔法，它是一种工具。它可以帮助我们处理人类风险官永远无法触及、阅读、观看或大规模分析的数百万数据点。AI当然提高了效率，但如果我们带着意图去设计它，它就能超越效率本身。它变得更加公平，让我们能在他人看到风险的地方发现价值；在他人看到石头的地方看到黄金；并使我们能够大规模地提供服务，同时尊重当地的知识、文化和背景。它实现了金融服务的超个性化（hyper-personalization）。
它让我们能够对玛丽亚这样的人说“是”。让我们能够对多年前我母亲创业时说“是”。也让我们能够对数百万推动经济前进的女性企业家说“是”。我们说“是”，不是因为一份银行对账单，而是因为数百万个无声的信号告诉我们，她会坚持出现，她会履行承诺，她值得信赖。
(掌声)