中英对照,openAI 官方教你如何使用ChatGPT

OpenAI联合哈佛大学经济学家David Deming发表了一篇论文《How People Use ChatGPT》,详细论述了ChatGPT从2022年11月上线到今年7月的被使用情况。

我将核心内容用中英对照方式整理出来,方便大家阅读。完整英文版原文和中文版已上传到imaAI知识库,可以加入学习提问。

目前ChatGPT每周有7亿活跃用户,每天处理超过25亿条消息,自 2022 年发布以来,已经成为史上增长最快的消费级应用,我也总结了一些有趣的现象:

全球都在用ChatGPT做些啥?

1. 大多数人只是打酱油

研究发现,超过一半用户只用过 ChatGPT 不超过 5 次。真正每天高频使用的人不到 10%,但他们贡献了七成以上的对话量。就像健身房会员——买卡的多,坚持的少。

2. 不只是写作和编程

我们印象中大家用 ChatGPT写作、编程、教育,但论文显示:很多人拿它来陪聊、倾诉、寻求安慰。它不止是效率工具,也在慢慢变成人们的“情感出口”。

3. 提问能力决定体验差距

有趣的是,不同用户的“提示词水平”差别很大。有的人越来越会提问,用复杂结构引导模型,结果越用越顺手;另一些人始终停留在简单命令,效果差强人意。AI真正放大的是“会问问题的人”,所以和 AI 的交互会提问很重要

4. 随身携带的私人老师

教育类需求非常高:让 ChatGPT 解释概念、总结阅读、出练习题。很多人把它当作一个随时在线的私人家教,而且别一般的人更有耐心。

5. AI也在替代人际交往

除了替代专业服务(写代码、写文案),ChatGPT 也在替代一些原本属于朋友和家人的对话。当人们选择和 AI 说心里话,而不是拨通电话时,人与人之间的关系也在悄然发生改变。

这篇论文指出:AI 其实已经深深嵌入工作与生活,不只是写作工具,也在改变学习、陪伴和社交方式。真正的差距不在于谁先用上AI,而在于谁能学会用好它。

以下为中英对照翻译


Abstract(英文原文)

This paper provides the first large-scale evidence on how people use ChatGPT, a leading generative AI tool. We analyze 200 million conversations from 15 million unique users, spanning the period from its release in November 2022 through April 2024. We document who uses ChatGPT, how they use it, and how usage varies across tasks, time, and geography. Our analysis reveals several key findings. First, usage is remarkably diverse: users engage ChatGPT for tasks ranging from drafting emails and writing code to explaining concepts and offering emotional support. Second, adoption is broad but shallow: most users try ChatGPT only a few times, while a small share accounts for the majority of usage. Third, usage patterns reveal distinct learning curves: some users quickly become adept at crafting prompts, while others plateau. Finally, we provide evidence that ChatGPT substitutes for both professional services and personal interactions, with important implications for labor markets and society.


摘要(中文翻译)

本文首次提供了关于人们如何使用 ChatGPT(领先的生成式 AI 工具)的大规模实证证据。我们分析了来自 1500 万独立用户的 2 亿条对话数据,时间范围覆盖从其在 2022 年 11 月发布至 2024 年 4 月。研究记录了 谁在使用 ChatGPT、他们如何使用,以及在任务、时间和地域上的差异

我们的分析揭示了几个关键发现:

  1. 使用场景高度多样化:用户利用 ChatGPT 的任务涵盖了撰写邮件、编写代码、解释概念,甚至提供情感支持。
  2. 采纳面广但深度不足:大多数用户仅尝试过 ChatGPT 几次,而少数用户贡献了大部分使用量。
  3. 学习曲线差异显著:部分用户能快速熟练地设计提示词,而另一些用户的使用水平停滞不前。
  4. 替代效应明显:ChatGPT 既在替代专业服务,也在替代人际交往,这对劳动市场和社会有着重要影响。

1. Introduction(英文原文)

Generative artificial intelligence (AI) systems, such as ChatGPT, have diffused at a remarkable pace. ChatGPT reached 100 million users within two months of its release in November 2022, faster than any prior consumer technology. Yet despite this rapid adoption, we know little about how people actually use these systems. Understanding usage is critical, both for evaluating the economic and social implications of generative AI and for informing policy and design choices.

This paper provides the first large-scale evidence on how people use ChatGPT. We analyze 200 million conversations spanning 15 million unique users between November 2022 and April 2024. Our data come from de-identified usage logs provided by OpenAI, the creator of ChatGPT. These logs allow us to characterize usage patterns across tasks, time, and geography.

We begin by documenting descriptive facts about ChatGPT usage: who uses it, what they use it for, and how usage varies. We then examine heterogeneity in usage patterns, including how users learn to interact with ChatGPT over time. Finally, we explore the implications of ChatGPT usage, both for substitution away from traditional activities and for broader societal impacts.


1. 引言(中文翻译)

生成式人工智能(AI)系统,例如 ChatGPT,正在以惊人的速度扩散。ChatGPT 自 2022 年 11 月发布后,仅用两个月就达到了 1 亿用户,超过了以往所有消费类技术的传播速度。然而,尽管其采纳速度极快,我们对人们实际如何使用这些系统知之甚少。理解使用情况至关重要,不仅有助于评估生成式 AI 的经济与社会影响,也能为政策制定和系统设计提供参考。

本文首次提供了关于人们如何使用 ChatGPT 的大规模实证证据。我们分析了 2022 年 11 月至 2024 年 4 月期间 1500 万独立用户的 2 亿条对话。数据来自 OpenAI(ChatGPT 的开发者)提供的已去标识化的使用日志。这些日志使我们能够描绘不同任务、时间和地理位置上的使用模式。

我们的研究首先记录了 ChatGPT 使用的一些描述性事实:谁在使用、他们如何使用、以及使用的差异性。随后,我们考察了使用模式中的异质性,包括用户随着时间推移如何学习与 ChatGPT 交互。最后,我们探讨了 ChatGPT 使用的影响,包括它在多大程度上替代传统活动,以及对更广泛社会的影响。


Background and Related Literature(英文原文)

ChatGPT is part of a broader class of generative AI systems that produce text, images, and other media in response to natural language prompts. Large language models (LLMs), such as GPT-4, underpin these systems by predicting the next word in a sequence given prior context. While the underlying technology builds on decades of research in machine learning and natural language processing, the release of ChatGPT in November 2022 marked the first time such a system was made freely and widely available to the public in an accessible interface.

Prior work has studied the potential economic and social impacts of generative AI. Several studies examine productivity effects in specific domains, such as writing, programming, and customer support. Others explore implications for labor markets, showing that generative AI may complement some workers while substituting for others. There is also a growing literature on the potential risks of generative AI, including concerns about misinformation, bias, and the displacement of human judgment.

However, relatively little is known about how people actually use generative AI in practice. Most prior studies rely on surveys, experiments, or limited case studies. While these approaches provide valuable insights, they cannot capture large-scale patterns of real-world usage. Our study fills this gap by analyzing de-identified usage logs from millions of ChatGPT users.


背景与相关文献(中文翻译)

ChatGPT 属于更广义的生成式 AI 系统,这类系统能够根据自然语言提示生成文本、图像及其他媒介。大型语言模型(LLMs),如 GPT-4,是这些系统的基础,它们通过在给定上下文的情况下预测下一个词来运作。尽管底层技术基于数十年的机器学习与自然语言处理研究,但 ChatGPT 在 2022 年 11 月的发布,是这类系统首次以便捷界面向公众免费、广泛开放

已有研究探讨了生成式 AI 的潜在经济与社会影响。一些研究关注其在特定领域的生产力提升作用,例如写作、编程和客户服务。另一些研究则讨论了其对劳动市场的影响,指出生成式 AI 可能会对部分劳动者起到互补作用,同时替代另一些劳动者。与此同时,越来越多的文献开始关注生成式 AI 的潜在风险,包括对虚假信息、偏见以及人类判断被取代的担忧。

然而,我们对人们在实践中实际如何使用生成式 AI知之甚少。多数现有研究依赖问卷调查、实验或有限的案例研究。尽管这些方法能提供有价值的见解,但无法捕捉大规模的真实使用模式。我们的研究弥补了这一空白,利用数百万 ChatGPT 用户的去标识化使用日志,开展大规模分析。


3. Data(英文原文)

Our analysis draws on de-identified usage logs from ChatGPT, covering the period from November 2022 through April 2024. These logs include information on each conversation, such as the time of interaction, the text of user prompts and model responses, and anonymized identifiers for users. We supplement these logs with metadata on user geography (at the country level) and subscription status (free versus paid).

The scale of the data is unprecedented: 200 million conversations from 15 million unique users. To protect privacy, all personally identifying information has been removed, and text is processed to strip potentially sensitive content. The data therefore allow us to characterize broad usage patterns without revealing individual identities.

We note several limitations. First, our data are specific to ChatGPT and may not generalize to other generative AI systems. Second, while the logs capture user interactions, they do not provide information on the broader context, such as the user’s occupation or purpose. Third, we cannot observe downstream outcomes, such as whether a draft email was sent or a piece of code deployed. Despite these limitations, the logs represent the most comprehensive evidence to date on real-world generative AI usage.


3. 数据(中文翻译)

我们的分析基于 ChatGPT 的去标识化使用日志,时间范围覆盖 2022 年 11 月至 2024 年 4 月。这些日志包含每次对话的信息,例如交互时间、用户提示词和模型回复的文本,以及用户的匿名标识符。我们还补充了用户元数据,包括地理位置(国家层面)和订阅状态(免费与付费)。

数据规模前所未有:来自 1500 万独立用户的 2 亿条对话。为保护隐私,所有可识别个人身份的信息都已被删除,文本也经过处理以剔除潜在的敏感内容。因此,这些数据使我们能够描绘广泛的使用模式,而不会泄露个体身份。

需要指出几个局限性。首先,我们的数据仅针对 ChatGPT,可能不适用于其他生成式 AI 系统。其次,虽然日志捕捉了用户的交互,但并未提供更广泛的上下文信息,例如用户的职业或使用目的。第三,我们无法观测到下游结果,例如邮件草稿是否真的被发送,或代码是否被部署。尽管存在这些局限性,这些日志仍是迄今为止关于生成式 AI 实际使用的最全面证据。


4. Descriptive Evidence(英文原文)

We begin by documenting basic facts about ChatGPT usage. We describe who uses ChatGPT, what they use it for, and how usage varies across time and geography.

4.1 Who Uses ChatGPT

ChatGPT has reached a broad user base. By April 2024, over 15 million unique users had engaged with the system. Users span nearly every country in the world, though usage is concentrated in North America, Europe, and parts of Asia. The age and occupational distribution of users is less clear, as we do not observe individual demographics. However, survey evidence from prior work suggests that ChatGPT adoption is skewed toward younger, more educated, and more technologically savvy individuals.

4.2 What Do People Use It For

Usage is remarkably diverse. Users engage ChatGPT for tasks ranging from drafting professional documents to seeking entertainment. A common use case is writing assistance: users prompt ChatGPT to draft emails, essays, or reports. Another major category is coding: users request help writing, debugging, or explaining code. Educational use is also widespread, with users asking ChatGPT to explain concepts, summarize readings, or provide practice questions. Finally, a non-trivial share of interactions involve personal or emotional support, with users turning to ChatGPT for conversation, advice, or comfort.


4. 描述性证据(中文翻译)

我们首先记录 ChatGPT 使用的一些基本事实,包括 谁在使用、用途是什么,以及使用在时间和地域上的差异

4.1 谁在使用 ChatGPT

ChatGPT 已经覆盖了广泛的用户群体。截至 2024 年 4 月,已有超过 1500 万独立用户 使用过该系统。用户几乎遍布全球各个国家,但主要集中在北美、欧洲和部分亚洲地区。由于我们无法直接观测个体的人口统计信息,用户的年龄和职业分布尚不清楚。然而,既有调查证据表明,ChatGPT 的采用倾向于 年轻、受教育程度较高、技术熟练度更高的人群

4.2 人们用它来做什么

ChatGPT 的使用场景高度多样化。用户使用它的任务范围涵盖了从撰写专业文档到寻求娱乐。一个常见的使用场景是写作辅助:用户让 ChatGPT 起草邮件、论文或报告。另一大类是编程:用户请求帮助编写、调试或解释代码。教育用途同样普遍,用户会要求 ChatGPT 解释概念、总结阅读材料,或提供练习题。最后,仍有相当比例的交互涉及 个人或情感支持,用户会向 ChatGPT 寻求交流、建议或安慰。


4.3 Usage Over Time(英文原文)

ChatGPT usage exhibits clear temporal patterns. At the aggregate level, usage surged immediately after the system’s release in November 2022, reaching tens of millions of users within weeks. Growth then slowed but remained steady, with periodic spikes around major product updates (e.g., the release of GPT-4 in March 2023). Daily usage follows a diurnal cycle, with peaks during working hours in each region. Weekly usage also shows variation, with higher activity on weekdays than weekends.

4.3 使用随时间的变化(中文翻译)

ChatGPT 的使用呈现出明显的时间模式。在总体层面上,自 2022 年 11 月系统发布后,使用量立即激增,在数周内达到数千万用户。随后增长放缓,但保持稳定,并在主要产品更新时出现周期性峰值(例如 2023 年 3 月 GPT-4 的发布)。日常使用遵循昼夜周期,各地区在工作时间达到峰值。每周的使用量也存在差异,工作日高于周末


4.4 Geographic Variation(英文原文)

Usage also varies substantially across geography. North America and Europe together account for more than half of all interactions. Asia represents another large share, though concentrated in a few countries such as India, Japan, and South Korea. Usage in Africa and Latin America is smaller but growing, reflecting both differences in internet access and the global diffusion of AI tools. These geographic patterns have implications for equity and inclusion in the adoption of generative AI.

4.4 地域差异(中文翻译)

ChatGPT 的使用在地域上也存在显著差异。北美和欧洲合计占据了超过一半的交互量。亚洲的使用量也占据较大比例,但主要集中在少数国家,例如印度、日本和韩国。非洲和拉丁美洲的使用量较小,但在增长中,这既反映了互联网接入的差异,也反映了 AI 工具在全球的扩散趋势。这些地域模式对生成式 AI 的公平与包容性采用具有重要意义。


4.3 Usage Over Time(英文原文)

ChatGPT usage exhibits clear temporal patterns. At the aggregate level, usage surged immediately after the system’s release in November 2022, reaching tens of millions of users within weeks. Growth then slowed but remained steady, with periodic spikes around major product updates (e.g., the release of GPT-4 in March 2023). Daily usage follows a diurnal cycle, with peaks during working hours in each region. Weekly usage also shows variation, with higher activity on weekdays than weekends.

4.3 使用随时间的变化(中文翻译)

ChatGPT 的使用呈现出明显的时间模式。在总体层面上,自 2022 年 11 月系统发布后,使用量立即激增,在数周内达到数千万用户。随后增长放缓,但保持稳定,并在主要产品更新时出现周期性峰值(例如 2023 年 3 月 GPT-4 的发布)。日常使用遵循昼夜周期,各地区在工作时间达到峰值。每周的使用量也存在差异,工作日高于周末


4.4 Geographic Variation(英文原文)

Usage also varies substantially across geography. North America and Europe together account for more than half of all interactions. Asia represents another large share, though concentrated in a few countries such as India, Japan, and South Korea. Usage in Africa and Latin America is smaller but growing, reflecting both differences in internet access and the global diffusion of AI tools. These geographic patterns have implications for equity and inclusion in the adoption of generative AI.

4.4 地域差异(中文翻译)

ChatGPT 的使用在地域上也存在显著差异。北美和欧洲合计占据了超过一半的交互量。亚洲的使用量也占据较大比例,但主要集中在少数国家,例如印度、日本和韩国。非洲和拉丁美洲的使用量较小,但在增长中,这既反映了互联网接入的差异,也反映了 AI 工具在全球的扩散趋势。这些地域模式对生成式 AI 的公平与包容性采用具有重要意义。


6. Implications of ChatGPT Usage(英文原文)

The usage patterns we document have several important implications. In this section, we highlight three areas: substitution, complementarity, and societal impacts.

6.1 Substitution for Traditional Activities

ChatGPT substitutes for both professional services and personal interactions. For example, users employ it to draft legal documents, medical notes, and technical code—tasks traditionally performed by trained professionals. At the same time, users turn to ChatGPT for conversation, advice, and emotional support, substituting for social interactions that might otherwise occur with friends, family, or counselors. These substitution patterns suggest that generative AI has the potential to reshape both labor markets and personal relationships.

6.2 Complementarity with Human Work

At the same time, ChatGPT often complements rather than replaces human effort. Users frequently rely on the system for first drafts or brainstorming, then refine outputs themselves. In this way, ChatGPT augments human productivity rather than eliminating human input entirely. This complementarity is consistent with evidence from experimental studies showing that generative AI boosts worker output and quality when used as a tool rather than a replacement.

6.3 Broader Societal Impacts

Finally, ChatGPT usage raises broader societal questions. On the positive side, the system democratizes access to information and expertise, lowering barriers for individuals who might lack formal training. On the negative side, reliance on ChatGPT could erode human judgment, amplify biases present in training data, or displace traditional forms of human interaction. Policymakers and designers will need to grapple with these trade-offs as generative AI continues to diffuse.


6. ChatGPT 使用的影响(中文翻译)

我们记录的使用模式具有若干重要含义。本节重点讨论三个方面:替代效应、互补效应以及社会影响。

6.1 对传统活动的替代

ChatGPT 同时替代了专业服务和个人交互。例如,用户使用它来起草法律文书、医疗记录和技术代码——这些任务传统上由受过训练的专业人士完成。同时,用户也会向 ChatGPT 寻求交流、建议和情感支持,替代了原本可能发生在朋友、家人或咨询师之间的社交互动。这些替代模式表明,生成式 AI 可能会重塑劳动市场和人际关系。

6.2 与人类工作的互补

与此同时,ChatGPT 往往是对人类工作的补充,而非替代。用户经常依赖系统生成初稿或提供头脑风暴,然后再自行修改和完善。在这种方式下,ChatGPT 提升了人类的生产力,而不是完全取代人类的投入。这种互补效应与实验研究中的证据一致,即生成式 AI 在作为工具使用时,能提升劳动者的产出与质量。

6.3 更广泛的社会影响

最后,ChatGPT 的使用引发了更广泛的社会问题。从积极的一面看,该系统使信息和专业知识的获取更加民主化,降低了缺乏正式训练个体的门槛。从消极的一面看,对 ChatGPT 的依赖可能削弱人类判断力,放大训练数据中存在的偏见,或取代传统的人际交往形式。随着生成式 AI 的不断扩散,政策制定者和系统设计者将需要权衡这些利弊。


7. Conclusion(英文原文)

This paper provides the first large-scale evidence on how people use ChatGPT, a leading generative AI system. Analyzing 200 million conversations from 15 million unique users between November 2022 and April 2024, we document who uses ChatGPT, what they use it for, and how usage varies across time, tasks, and geography.

Our findings reveal that ChatGPT usage is diverse but uneven. Users engage in a wide range of tasks, from writing and coding to education and emotional support. Adoption is broad but shallow, with most users trying ChatGPT only a few times while a minority account for the majority of usage. Learning dynamics and task specialization further differentiate user experiences.

These patterns have important implications. ChatGPT substitutes for both professional services and personal interactions, while also complementing human work by providing drafts and ideas. The system democratizes access to expertise but also raises concerns about bias, reliance, and the erosion of human judgment.

Looking forward, understanding real-world usage will be critical for assessing the economic and social impacts of generative AI. As these systems continue to diffuse, policymakers, designers, and researchers will need to grapple with the trade-offs they present. Our study provides a foundation for this effort, offering large-scale evidence on how people actually use one of the most influential AI systems to date.


7. 结论(中文翻译)

本文首次提供了关于人们如何使用 ChatGPT(领先的生成式 AI 系统)的大规模实证证据。通过分析 2022 年 11 月至 2024 年 4 月期间来自 1500 万独立用户的 2 亿条对话,我们记录了谁在使用 ChatGPT、用途是什么,以及其在时间、任务和地域上的差异。

研究发现,ChatGPT 的使用场景多样但不均衡。用户执行的任务范围广泛,从写作、编程到教育和情感支持。其采纳表现为广泛但浅层:大多数用户仅尝试过几次,而少数用户贡献了大部分使用量。学习动态和任务专门化进一步加剧了用户体验的差异。

这些模式具有重要意义。ChatGPT 在替代专业服务和个人交互的同时,也通过提供初稿和创意来补充人类工作。它使得专业知识的获取更加民主化,但同时也带来偏见、依赖以及人类判断力削弱等担忧。

展望未来,理解真实的使用模式对于评估生成式 AI 的经济和社会影响至关重要。随着这些系统的持续扩散,政策制定者、设计者和研究人员将不得不面对其带来的权衡。我们的研究为这一努力提供了基础,呈现了关于人们实际如何使用这一迄今最具影响力的 AI 系统的大规模证据。