来源：《自然》
原文刊登日期：2023年5月3日
Since 2015, Koko, a mobile mental-health app, has tried to provide crowdsourced support for people in need. Text the app to say that you’re feeling guilty about a work issue, and an empathetic response will come through in a few minutes — clumsy perhaps, but unmistakably human — to suggest some positive coping strategies.
自2015年以来，移动心理健康应用程序Koko一直试图为有需要的人提供众包支持。发短信说你对工作上的问题感到内疚，几分钟后你就会得到一个同情的回应——也许有些笨拙，但毫无疑问是人性的——提出一些积极的应对策略。
The app might also invite you to respond to another person’s plight while you wait. To help with this task, an assistant called Kokobot can suggest some basic starters, such as “I’ve been there”.
这款应用可能还会邀请你在等待的同时对别人的困境做出回应。为了帮助完成这项任务，一个名为Kokobot的助手可以建议一些基本的开场白，比如“我去过那里”。
But last October, some Koko app users were given the option to receive much-more-complete suggestions from Kokobot. These suggestions were preceded by a disclaimer, says Koko co-founder Rob Morris: “I’m just a robot, but here’s an idea of how I might respond.” Users were able to edit or tailor the response in any way they felt was appropriate before they sent it.
但去年10月，一些Koko应用程序用户可以选择接受Kokobot提供的更完整的建议。Koko的联合创始人罗布•莫里斯说，这些建议之前都有一个免责声明：“我只是一个机器人，但这里有一个我可能会如何回应的想法。”用户可以在发送之前以任何他们认为合适的方式编辑或修改回复。
What they didn’t know at the time was that the replies were written by GPT-3, the powerful artificial-intelligence (AI) tool that can process and produce natural text, thanks to a massive written-word training set. When Morris eventually tweeted about the experiment, he was surprised by the criticism he received. “I had no idea I would create such a fervour of discussion,” he says.
用户当时不知道的是，回复是由GPT-3生成的，GPT-3是一种强大的人工智能（AI）工具，可以处理和生成自然文本，这要归功于一个庞大的书面语训练集。当莫里斯最终在推特上发布关于这项实验的消息时，他对自己受到的批评感到惊讶。他说：“我不知道自己会引起如此热烈的讨论。”
People have been trying to automate mental-health therapy for 70 years, and chatbots in one form or another have been a part of that quest for about 60. There is a need for the greater efficiency that these tools promise. Estimates suggest that for every 100,000 people worldwide, there are about 4 psychiatrists on average; that number is much lower in most low- and middle-income countries.
70年来，人们一直在尝试将心理健康治疗自动化，而60年来，各种形式的聊天机器人一直是这一探索的一部分。人们需要这些工具所承诺的更高效率。据估计，全世界每10万人中平均约有4名精神科医生；在多数中低收入国家，这一数字要低得多。
Recognizing this gap, smartphone-app developers have built thousands of programs offering some semblance of therapy that can fit in one’s pocket. There were 10,000–20,000 mobile mental-health apps available in 2021. But for many of these apps, the evidence to support their use is quite thin, says Nicholas Jacobson, a biomedical data scientist at Dartmouth College’s Center for Technology and Behavioral Health in Lebanon, New Hampshire. And the incorporation of large language models such as GPT-3, and the related chatbot ChatGPT, represents a new step that many find concerning.
认识到这一缺口，智能手机应用程序开发商已经开发了数千个程序，提供一些可以放在口袋里的类似治疗的程序。2021年，有1万到2万个移动心理健康应用程序可用。但是，新罕布什尔州黎巴嫩市达特茅斯学院技术与行为健康中心的生物医学数据科学家尼古拉斯·雅各布森说，对于许多这样的应用程序来说，支持它们使用的证据相当薄弱。而GPT-3等大型语言模型和相关聊天机器人ChatGPT的加入，代表了许多人感到担忧的新一步。
Some are worried about increased threats to privacy and transparency. And there are concerns about safety and legal liability. Earlier this year, a Belgian man reportedly committed suicide after weeks of sharing his climate-related anxieties with an AI chatbot called Eliza. His wife contends that he would still be alive if he had not engaged with this technology.
一些人担心隐私和透明度受到的威胁会增加。人们还担心安全问题和法律责任。据报道，今年早些时候，一名比利时男子在与一个名为Eliza的人工智能聊天机器人分享他与气候有关的焦虑数周后自杀。他的妻子认为，如果他没有参与这项技术，他可能还活着。