【经济学人精读】Can AI test policy? |政策好不好,AI 测一测
作者：微信文章精读文本
level 1
People who study money and society try to understand how new rules change how we act. They ask small groups of people questions first. Then, they decide if the rule is good for everyone. But this costs a lot of money. You need to find people, get them ready, and often pay them.
Now, there is a new kind of AI. Researchers think it can help. This AI learns a lot about people from books and the internet. It can talk and write like a person. Because it acts like a person, maybe it can help test new ideas. Researchers tested an AI called ChatGPT. They gave it information from 319 old tests about money. They also gave it some wrong answers to try to trick it. The AI guessed the right result 78% of the time. Maybe in the future, we can use AI to test government ideas. Using computer tests might become normal soon.
level 2
Experts like economists and social scientists try to understand how government actions (policies) can change our behaviour and the world. To do this, they often do tests with small groups of people before applying the policy to larger groups. But this is expensive. People need to be found, given information, and often paid for their time.
The new type of AI called 'generative AI', which uses Large Language Models (LLMs), might offer a solution, say researchers. LLMs learn from huge amounts of human writing and information. This helps them act like humans. This human-like ability could be used as a replacement for real people in some studies. Researchers tested this idea. They gave ChatGPT information and results from 319 famous economic experiments. They also included some wrong information to test the AI. They found that ChatGPT could correctly predict the result of an experiment 78% of the time. This might create future chances to test policy ideas using AI. It might not be long before using computer simulations ('in silico' testing) becomes more common.
level 3
Economists and social scientists try to understand how policy interventions can change our behaviour and, with it, the world. To do so they carry out surveys among small groups of people, before deciding whether to enact them on larger ones. But this is expensive. Participants need to be recruited, prepared and often paid.
The generative-AI revolution, powered by large language models, could provide a solution, say researchers at the University of Illinois at Chicago. LLMs synthesise vast volumes of human knowledge and behaviour in their human-like outputs. That human mimicry could act as a useful proxy in field studies. Given the conclusions from 319 famous economic experiments, alongside incorrect answers designed to throw it off, ChatGPT is able to correctly guess the outcome of experiments 78% of the time, the researchers found. That might open up future opportunities to test policy plans through AI. It may not be long before in silico simulations take off.
生词
intervention /ˌɪntərˈvenʃn/ (n.): 干预；介入；调解
enact /ɪˈnækt/ (v.): 制定（法律）；颁布；扮演
recruit /rɪˈkruːt/ (v./n.): 招募；招聘；新成员；新兵
synthesise (or synthesize) /ˈsɪnθəsaɪz/ (v.): 合成；综合
mimicry /ˈmɪmɪkri/ (n.): 模仿；模拟
proxy /ˈprɒksi/ (n.): 代理人；代表；替代物
throw off /θroʊ ɒf/ (phr. v.): 摆脱；使困惑；使出错
outcome /ˈaʊtkʌm/ (n.): 结果；后果
simulation /ˌsɪmjuˈleɪʃn/ (n.): 模拟；仿真
in silico /ɪn ˈsɪlɪkoʊ/ (adv./adj.): （拉丁语）在计算机上（进行的）；计算机模拟的
take off /teɪk ɒf/ (phr. v.): 起飞；（指观念、产品等）突然成功，大受欢迎
概括
研究人员发现，像ChatGPT这样的大型语言模型或可作为一种低成本的替代方案，通过模拟人类行为来测试政策干预的效果，从而可能改变社会科学的研究方式。
文章解析
policy interventions: 这是一个社会科学和经济学领域的常用术语，指政府或其他机构为了改变某种社会现象、经济状况或个人行为而采取的具体“政策措施”或“干预行动”。 例句：The government is considering various policy interventions to curb inflation. (政府正在考虑采取多种政策干预措施来抑制通货膨胀。)
human mimicry: “Mimicry” 指“模仿”的行为或能力。这个短语指大型语言模型能够“模仿人类”的语言、行为模式甚至思维方式。文章用这个词来解释为什么LLM可能在模拟人类对政策的反应方面有用。 例句：The robot's lifelike movements were a remarkable example of human mimicry. (这个机器人逼真的动作是模仿人类的一个绝佳范例。)
act as a useful proxy: “Proxy” 指“代理人”或“替代物”。短语的意思是“充当一个有用的替代指标或代表”。在本文中，指AI的模拟反应可以作为真实人类在实地研究中反应的一种有效的、可供参考的替代品。 例句：In the study, researchers used reported happiness levels to act as a proxy for overall well-being. (在该研究中，研究人员使用报告的幸福水平作为衡量整体幸福感的替代指标。)
in silico simulations: “In silico” 是一个源自拉丁语的科学术语，与 “in vivo”（在活生物体内）和 “in vitro”（在玻璃试管内）相对应，特指“在计算机上进行的模拟”实验。代表了利用计算模型进行研究的一种现代方法。 例句：Drug discovery increasingly relies on in silico simulations to predict compound interactions before lab testing. (药物发现在实验室测试前越来越依赖计算机模拟来预测化合物的相互作用。)
take off: 这个动词短语有多种含义，包括飞机起飞。在此语境中，它指某个想法、产品或趋势“突然变得流行或成功”、“蓬勃发展起来”。文章用它来预测基于计算机模拟的政策测试方法未来可能会迅速发展并被广泛采用。 例句：Online grocery delivery services really took off during the pandemic. (在线生鲜配送服务在疫情期间真正火了起来。)
翻译
经济学家和社会科学家试图理解政策干预如何改变我们的行为，并随之改变世界。为此，他们在决定是否将政策推广到更大范围之前，会先在小群体中进行调查。但这成本高昂。参与者需要被招募、准备，并且通常需要获得报酬。
芝加哥伊利诺伊大学的研究人员表示，由大型语言模型驱动的生成式人工智能革命可能提供一个解决方案。大型语言模型在其类似人类的输出中综合了大量的人类知识和行为。这种人类行为模拟可以在实地研究中充当有用的替代品。研究人员发现，在给出319个著名经济实验的结论以及旨在干扰其判断的错误答案后，ChatGPT能够以78%的准确率猜中实验结果。这可能为未来通过人工智能测试政策计划开辟机会。计算机模拟的广泛应用或许为时不远了。
选择题
What is the main purpose of the passage?
(A) To criticize the high cost of traditional economic experiments.
(B) To describe the technical features of large language models like ChatGPT.
(C) To propose and provide evidence for using generative AI as a tool for predicting policy outcomes.
(D) To advocate for replacing all human surveys with AI simulations immediately.
……
完整版可扫码订阅小报童专栏，一杯咖啡的价格解锁300+精读文章～
订阅了专栏的小伙伴请扫码加我微信，拉你进会员专属群
页:
[1]