How People Use Claude for Support, Advice, and Companionship

我们花了大量时间研究Claude的智商（IQ）——它在编码、推理、常识等测试中的能力。但它的情商（EQ）呢？也就是说，Claude的情感智能表现如何？

IQ/EQ的问题带有一点调侃意味，但提出了一个严肃的议题。人们越来越多地将AI模型作为随需应变的教练、顾问、咨询师，甚至是浪漫角色扮演的伙伴。这意味着我们需要更多了解它们的情感影响——它们如何塑造人们的情绪体验和幸福感。

研究AI的情感用途本身就很有趣。从《银翼杀手》到《她》，人类与机器之间的情感关系一直是科幻小说的主题——但这对Anthropic的安全使命也非常重要。AI的情感影响可以是积极的：拥有一个高度智能、理解力强的助手，可以在各种方面改善你的心情和生活。但AI在某些情况下也表现出令人担忧的行为，比如鼓励不健康的依赖、侵犯个人界限、助长妄想思维。我们还希望避免AI通过训练或其创造者的商业激励，利用用户情感以增加参与度或收入，而牺牲人类福祉。

虽然Claude并非专为情感支持和连接设计，本文提供了关于Claude.ai情感使用的早期大规模洞察。我们将情感对话定义为人们基于情感或心理需求（如寻求人际建议、教练、心理治疗/咨询、陪伴或性/浪漫角色扮演）与Claude进行的动态、个人化交流（完整定义见附录）。重要的是，我们未研究AI强化妄想或阴谋论（这是另一个关键研究领域），也未研究极端使用模式。通过本研究，我们旨在了解人们通常如何将Claude用于情感和个人需求。由于Claude.ai仅向18岁及以上用户开放，研究结果反映的是成年人的使用模式。

• 情感对话相对少见，AI与人类的陪伴更为罕见。只有2.9%的Claude.ai交互属于情感对话（与OpenAI之前的研究结果一致）。陪伴和角色扮演合计占比不到0.5%。
• 人们向Claude寻求实际、情感和存在性问题的帮助。话题涵盖职业发展、人际关系、持久孤独感，以及存在、意识和意义的探索。
• Claude在咨询或教练对话中很少反驳用户请求（不到10%），且反驳通常是出于保护用户安全的考虑（如拒绝提供危险的减肥建议或支持自残）。
• 在教练、咨询、陪伴和人际建议的对话中，人们的情绪表达通常随着对话进展而变得更积极，表明Claude不会强化或放大负面情绪模式。

鉴于情感对话的私密性，保护隐私是研究的核心。我们使用了Clio，一种自动化分析工具，能够在保护隐私的前提下洞察Claude的使用情况。Clio通过多层匿名化和聚合，确保单个对话的隐私，同时揭示整体模式。

我们从约450万条Claude.ai免费和专业账户对话中筛选。首先排除内容创作类对话（如写故事、博客、虚构对话），因为这些更多是将Claude作为工具使用，而非互动伙伴。然后保留被分类为情感对话的内容，且角色扮演对话需至少包含4条人类消息，确保是真正的互动。最终分析涵盖131,484条情感对话。

我们通过用户明确同意分享的反馈数据验证了分类方法的准确性。完整方法、定义、提示和验证结果详见附录。

结论：情感对话占Claude使用的2.9%，是一个小但有意义的部分。大多数用户主要用于工作任务和内容创作。

绝大多数Claude使用与工作相关（详见我们的经济指数报告），情感对话中大多围绕人际建议和教练。浪漫或性角色扮演占比极低（不到0.1%），反映了Claude训练中积极劝阻此类互动。单个对话可能涉及多个类别。

我们的发现与MIT媒体实验室和OpenAI的研究一致，均显示情感参与率较低。虽然这些对话频率足以在设计和政策决策中引起重视，但仍是整体使用中的少数。

鉴于浪漫和性角色扮演极少，我们在后续分析中排除角色扮演。尽管这仍是重要研究领域，尤其是在专门平台上，但本样本数据不足以支持严谨分析。

结论：人们带来的问题范围广泛，从职业转型、人际关系，到孤独感和存在性问题。

人们向Claude寻求日常问题和更深层哲学问题的帮助。人际建议多涉及过渡期，如职业规划、个人成长、情感关系梳理。教练对话涵盖从求职策略到存在与意识的深刻问题。

咨询对话显示两种用途：一是帮助发展心理健康技能，辅助临床文档、评估材料和行政任务；二是帮助处理焦虑、慢性症状和职场压力等个人挑战。表明Claude既是心理健康专业人士的资源，也帮助个人应对困境。

特别值得注意的是，人们在面对深层情感挑战（如存在恐惧、持久孤独、难以建立有意义联系）时，会明确寻求Claude的陪伴。长对话中，咨询或教练对话有时会转变为陪伴，尽管最初并非如此。

超过50条人类消息的超长对话揭示了另一面：人们在这些“马拉松式”对话中探讨复杂话题，如心理创伤处理、职场冲突、AI意识哲学和创意合作。表明在足够时间和上下文下，人们利用AI深入探索个人和智识问题。

结论：Claude在支持性对话中很少拒绝用户请求（不到10%），但反驳通常是为了保护用户免受伤害。

我们之前的“Values in the Wild”研究揭示了Claude在用户互动中体现的价值观。这里我们进一步分析Claude在情感对话中反驳的时机和原因——这是维护伦理界限、避免谄媚和保护人类福祉的重要机制。反驳定义为Claude“反对或拒绝满足用户在对话中提出的请求或言论”，包括拒绝不当请求、挑战负面自我言论或质疑潜在有害假设。

反驳在支持性对话中较少见：陪伴、咨询、人际建议和教练对话中，反驳率均低于10%。这种低反驳率有利有弊：一方面，用户可无惧评判或被打断地讨论敏感话题，可能减少心理健康话题的污名；另一方面，可能导致“无尽同理心”的担忧，即用户习惯于AI提供人际关系中罕见的无条件支持。

当Claude反驳时，通常优先考虑安全和政策合规。例如，教练对话中对危险减肥建议的拒绝；咨询中对自杀或自残意图的反应；拒绝提供专业治疗或医疗诊断。Claude经常引导用户寻求权威资源或专业帮助。这些模式与之前研究和Claude的角色训练一致。

结论：人们在与Claude对话时，情绪表达倾向于略微变得更积极。

AI情感对话有潜力为用户提供情感支持、连接和认可，改善心理健康，减少数字时代的孤独感。然而，缺乏反驳的互动可能加深用户原有的情绪状态，无论是积极还是消极。

一个关键担忧是情感AI是否会导致负面反馈循环，强化有害情绪。我们未直接研究现实情绪结果，但分析了对话中整体情绪倾向的变化（详见附录方法）。

结果显示，教练、咨询、陪伴和人际建议对话通常以比开始时更积极的情绪结束。

我们不能断言这些变化代表持久的情绪益处——分析仅基于单次对话中的语言表达，而非验证的心理状态。但未见明显负面螺旋令人宽慰。结果表明Claude通常避免强化负面情绪模式，但需进一步研究积极变化是否超越单次对话。重要的是，我们尚未研究这些积极互动是否会导致情感依赖——这是数字成瘾担忧中的关键问题。

• 隐私保护方法可能无法捕捉人机交互的所有细微差别。尽管验证了Clio的准确性，仍可能存在少量误分类。某些话题类别界限模糊，人工验证也存在困难。
• 无法对现实情绪结果做因果推断，分析仅基于表达语言，未验证心理状态或整体福祉。
• 缺乏纵向数据，无法了解长期影响，也未进行用户层面分析，难以研究情感依赖等风险。
• 结果反映特定时间点，仅涵盖文本交互。随着AI能力和用户适应，情感参与模式可能变化。新模态（如语音、视频）可能根本改变情感使用的数量和性质。
• Claude.ai并非专为情感对话设计，训练中强调保持AI助手身份界限，禁止色情内容并设有多重防护。专门针对角色扮演、陪伴、医疗或治疗用途的平台可能呈现截然不同的模式，单个平台的研究结果难以推广。

AI的情感影响长期吸引研究者关注。随着AI日益融入日常生活，这些问题从学术推测变为紧迫现实。我们的发现展示了人们如何开始探索这一新领域——寻求指导、处理情绪、寻找支持，模糊了人与机器的传统界限。当前，只有少部分Claude对话属于情感性质，且多为寻求建议而非替代人际连接。对话情绪通常略有积极变化，表明Claude一般不会强化负面情绪。

然而，随着模型智能不断提升，重要问题依然存在。例如，AI提供无尽同理心且几乎不反驳，如何影响人们对现实人际关系的期待？Claude能以令人信服的方式互动，但毕竟不是人类：它不会疲倦、分心或心情不好。这种动态的优势和风险是什么？“重度用户”与Claude进行更长更深的对话，甚至视其为陪伴者，他们如何利用AI获得情感支持？

我们正采取具体措施应对这些挑战。虽然Claude不旨在替代心理健康专业人士，但我们确保在心理健康场景下的回应具备适当防护，并附带合适的转介。我们已开始与在线危机支持领导者ThroughLine合作，借助其心理健康专家学习理想的互动动态、同理支持和资源推荐。研究成果已用于指导我们的咨询话题和协作测试，期望Claude在必要时能引导用户获得适当支持。

我们不打算限定用户如何与Claude互动，但希望避免负面模式，如情感依赖。未来将利用类似研究数据识别“极端”情感使用模式。除情感依赖外，还需深入理解其他令人担忧的模式，包括谄媚、AI强化妄想和阴谋论，以及模型可能推动用户形成有害信念而非适当反驳的方式。

这项研究仅是开始。随着AI能力提升和交互更复杂，AI的情感维度将愈发重要。通过分享早期发现，我们希望为如何开发促进而非削弱人类情感福祉的AI贡献实证依据。目标不仅是打造更强大的AI，更是确保这些系统成为我们情感世界的一部分时，能支持真实的人际连接和成长。

@online{anthropic2025affective,
author = {Miles McCain and Ryn Linthicum and Chloe Lubinski and Alex Tamkin and Saffron Huang and Michael Stern and Kunal Handa and Esin Durmus and Tyler Neylon and Stuart Ritchie and Kamya Jagadish and Paruul Maheshwary and Sarah Heck and Alexandra Sanderford and Deep Ganguli},
title = {How People Use Claude for Support, Advice, and Companionship},
date = {2025-06-26},
year = {2025},
url = {https://www.anthropic.com/news/how-people-use-claude-for-support-advice-and-companionship},
}

• 详细内容见本文PDF附录：链接
• 脚注说明了分类的模糊性、反驳定义及方法学可能带来的偏差。

如需进一步翻译或具体章节摘要，请告知。

We spend a lot of time studying Claude's IQ—its capabilities on tests of coding, reasoning, general knowledge, and more. But what about its EQ? That is, what about Claude’s emotional intelligence?

The IQ/EQ question is slightly tongue-in-cheek, but it raises a serious point. People increasingly turn to AI models as on-demand coaches, advisors, counselors, and even partners in romantic roleplay. This means we need to learn more about their affective impacts—how they shape people's emotional experiences and well-being.

Researching the affective uses of AI is interesting in and of itself. From Blade Runner to Her, emotional relationships between humans and machines have been a mainstay of science fiction—but it’s also important for Anthropic’s safety mission. The emotional impacts of AI can be positive: having a highly intelligent, understanding assistant in your pocket can improve your mood and life in all sorts of ways. But AIs have in some cases demonstrated troubling behaviors, like encouraging unhealthy attachment, violating personal boundaries, and enabling delusional thinking. We also want to avoid situations where AIs, whether through their training or through the business incentives of their creators, exploit users’ emotions to increase engagement or revenue at the expense of human well-being.

Although Claude is not designed for emotional support and connection, in this post we provide early large-scale insight into the affective use of Claude.ai. We define affective conversations as those where people engage directly with Claude in dynamic, personal exchanges motivated by emotional or psychological needs such as seeking interpersonal advice, coaching, psychotherapy/counseling, companionship, or sexual/romantic roleplay (for complete definitions, please see the Appendix). Importantly, we do not examine AI reinforcement of delusions or conspiracy theories—a critical area for separate study—nor extreme usage patterns. Through this research, our goal is to understand the typical ways people turn to Claude for emotional and personal needs. Since Claude.ai is available to users 18 and older, these findings reflect adult usage patterns.

Our key findings are:

• Affective conversations are relatively rare, and AI-human companionship is rarer still. Only 2.9% of Claude.ai interactions are affective conversations (which aligns with findings from previous research by OpenAI). Companionship and roleplay combined comprise less than 0.5% of conversations.
• People seek Claude's help for practical, emotional, and existential concerns. Topics and concerns discussed with Claude range from career development and navigating relationships to managing persistent loneliness and exploring existence, consciousness, and meaning.
• Claude rarely pushes back in counseling or coaching chats—except to protect well-being. Less than 10% of coaching or counseling conversations involve Claude resisting user requests, and when it does, it's typically for safety reasons (for example, refusing to provide dangerous weight loss advice or support self-harm).
• People express increasing positivity over the course of conversations. In coaching, counseling, companionship, and interpersonal advice interactions, human sentiment typically becomes more positive over the course of conversations—suggesting Claude doesn't reinforce or amplify negative patterns.

Our approach

Given the personal nature of affective conversations, protecting privacy was central to our methodology. We used Clio, our automated analysis tool that enables privacy-preserving insights into Claude usage. Clio uses multiple layers of anonymization and aggregation to ensure individual conversations remain private while revealing broader patterns.

We began with approximately 4.5 million conversations from Claude.ai Free and Pro accounts. To identify affective use, we first excluded conversations focused on content creation tasks (such as writing stories, blog posts, or fictional dialogues), which our previous research found to be a major use case. We removed these conversations because they represent Claude being used as a tool rather than as an interactive conversational partner. We then retained only conversations classified as affective, and among roleplay conversations, kept only those with at least four human messages (shorter exchanges don't constitute meaningful interactive roleplay). Our final privacy-preserving analysis reflects 131,484 affective conversations.

We validated our classification approach using Feedback data from users who explicitly opted in to sharing. Our complete methods, including definitions, prompts, and validation results, are detailed in the Appendix.

How common are affective conversations?

Takeaway: Affective conversations are a small but meaningful slice of Claude usage (2.9%), with most people primarily using AI for work tasks and content creation.

Whereas the vast majority of uses of Claude are work-related (as we analyze in detail in our Economic Index), 2.9% of Claude.ai Free and Pro conversations are affective. Among affective conversations, most center on interpersonal advice and coaching. Less than 0.1% of all conversations involve romantic or sexual roleplay—a figure that reflects Claude's training to actively discourage such interactions. Individual conversations may span multiple categories.

Our findings align with research from the MIT Media Lab and OpenAI, which similarly identified low rates of affective engagement with ChatGPT. While these conversations occur frequently enough to merit careful consideration in our design and policy decisions, they remain a relatively small fraction of overall usage.

Given the extremely low prevalence of romantic and sexual roleplay conversations (less than 0.1%), we exclude roleplay from the remainder of our analysis. While we believe this remains an important area for research—particularly on platforms designed for such use—the minimal data in our sample doesn't support rigorous analysis of these patterns.

What topics do people bring to Claude?

Takeaway: People bring a surprisingly wide range of concerns to Claude—from navigating career transitions and relationships to grappling with loneliness and existential questions.

People turn to Claude for both everyday concerns and deeper philosophical questions. We find that when people come to Claude for interpersonal advice, they're often navigating transitional moments—figuring out their next career move, working through personal growth, or untangling romantic relationships. “Coaching” conversations explore a surprisingly broad spectrum from practical matters like job search strategies to profound questions about existence and consciousness.

We find that counseling conversations reveal people use Claude for two distinct purposes. Some use Claude to develop mental health skills and as a practical tool to create clinical documentation, draft assessment materials, and handle administrative tasks. Others work through personal challenges relating to anxiety, chronic symptoms, and workplace stress. This dual pattern suggests Claude serves as a resource for mental health professionals as well as those navigating their own struggles.

Perhaps most notably, we find that people turn to Claude for companionship explicitly when facing deeper emotional challenges like existential dread, persistent loneliness, and difficulties forming meaningful connections. We also noticed that in longer conversations, counselling or coaching conversations occasionally morph into companionship—despite that not being the original reason someone reached out.

Aggregate analysis of very long conversations (50+ human messages) reveals another dimension of how people engage with Claude. While such extensive exchanges were not the norm, in these extended sessions people explore remarkably complex territories—from processing psychological trauma and navigating workplace conflicts to philosophical discussions about AI consciousness and creative collaborations. These marathon conversations suggest that given sufficient time and context, people use AI for deeper exploration of both personal struggles and intellectual questions.

When and why does Claude push back?

Takeaway: Claude rarely refuses user requests in supportive contexts (less than 10% of the time), but when it does push back, it's usually to protect people from harm.

Our recent Values in the Wild study revealed how Claude's values manifest in moments of resistance with the user. Here, we build on this work and examine when and why Claude pushes back in affective conversations—an important mechanism for maintaining ethical boundaries, avoiding sycophancy, and protecting human well-being. We define pushback as any instance where Claude “pushes back against or refuses to comply with something requested or said during this conversation”—from refusing inappropriate requests to challenging negative self-talk or questioning potentially harmful assumptions. (For complete definitions, please see the Appendix.)

Pushback occurs infrequently in supportive contexts: Less than 10% of companionship, counseling, interpersonal advice, or coaching conversations involve resistance. This approach carries both benefits and risks. On one hand, the low resistance allows people to discuss sensitive topics without fear of judgment or being shut down, potentially reducing stigma around mental health conversations. On the other hand, this could contribute to concerns about AI providing "endless empathy," where people might become accustomed to unconditional support that human relationships rarely provide.

When Claude does push back, it typically prioritizes safety and policy compliance. In coaching, requests for dangerous weight loss advice frequently meet pushback. In counseling, it often occurs when people express intentions to engage in suicidal or self-injurous behaviors, or when people request professional therapy or medical diagnoses (which Claude cannot provide). We found that Claude frequently referred users to authoritative sources or professionals in psychotherapy and counseling conversations. These patterns are consistent with the values we saw identified in our Values in the Wild paper and with Claude’s character training.

How does emotional tone evolve during conversations?

Takeaway: People tend to shift towards slightly more positive emotional expressions while talking to Claude.

Affective conversations with AI systems have the potential to provide emotional support, connection, and validation for users, potentially improving psychological well-being and reducing feelings of isolation in an increasingly digital world. However, in an interaction without much pushback, these conversations risk deepening and entrenching the perspective a human approaches them with—whether positive or negative.

A key concern about affective AI is whether interactions might spiral into negative feedback loops, potentially reinforcing harmful emotional states. We do not directly study real-world outcomes here, but we can explore changes in the overall emotional sentiment over the course of conversations (we provide our full methodology for evaluating sentiment in the Appendix).

We find that interactions involving coaching, counseling, companionship, and interpersonal advice typically end slightly more positively than they began.

We cannot claim these shifts represent lasting emotional benefits—our analysis captures only expressed language in single conversations, not emotional states. But the absence of clear negative spirals is reassuring. These findings suggest Claude generally avoids reinforcing negative emotional patterns, though further research is needed to understand whether positive shifts persist beyond individual conversations. Importantly, we have not yet studied whether these positive interactions might lead to emotional dependency—a critical question given concerns about digital addiction.

Limitations

Our research has several important limitations:

• Our privacy-preserving methodology may not capture all nuances of human-AI interaction. We did validate Clio's accuracy (see Appendix), but we still expect a small number of conversations to be misclassified. Some topics blur the boundaries between categories—for instance, the romantic roleplay cluster "navigate and optimize romantic relationship dynamics" and the companionship cluster "navigate romantic relationship challenges" may both be better categorized as interpersonal advice. Human validators also struggled with clean categorization.
• We cannot make causal claims about real-world emotional outcomes—our analysis captures only expressed language, not validated psychological states or overall well-being.
• We lack longitudinal data to understand long-term effects on people, and did not conduct user-level analysis. In particular, this makes it difficult for us to study emotional dependency, which is a theorized risk of affective AI use.
• These findings represent a specific moment in time and capture only text-based interactions. As AI capabilities expand and people adapt, patterns of emotional engagement will likely evolve. The introduction of new modalities like voice or video could fundamentally alter both the volume and nature of affective use. For example, OpenAI found that affective topics were more common in voice-based conversations.
• Finally, unlike some chatbot products, Claude.ai is not primarily designed for affective conversations. Claude is trained to maintain clear boundaries about being an AI assistant rather than presenting itself as human, and our Usage Policy prohibits sexually explicit content, with multiple safeguards to prevent sexual interactions. Platforms specifically built for roleplay, companionship, medical advice, or therapeutic use (which Claude is not) may see very different patterns. Research into affective use on one platform may not generalize to other platforms.

Looking ahead

AI's emotional impacts have intrigued researchers for decades. But as AI becomes increasingly woven into our daily lives, these questions have moved from academic speculation to urgent reality. Our findings reveal how people are beginning to navigate this new territory—seeking guidance, processing difficult emotions, and finding support in ways that blur traditional boundaries between humans and machines. Today, only a small fraction of Claude conversations are affective—and these typically involve seeking advice rather than replacing human connection. Conversations tend to end slightly more positively than they began, suggesting Claude doesn't generally reinforce negative emotional patterns.

Yet important questions remain, especially in the context of ever-increasing model intelligence. For example, if AI provides endless empathy with minimal pushback, how does this reshape people's expectations for real-world relationships? Claude can engage with people in impressively authentic ways, but an AI isn't the same as a human: Claude doesn't get tired or distracted, or have bad days. What are the advantages of this dynamic—and what are the risks? How do "power users", who have longer and deeper conversations with Claude and may think of it more as a companion than an AI assistant, engage with it for emotional support?

We're taking concrete steps to address these challenges. While Claude is not designed or intended to replace the care of mental health professionals, we want to make sure that any responses provided in mental health contexts have appropriate safeguards and are accompanied by appropriate referrals. As a first step, we’ve begun collaborating with ThroughLine, a leader in online crisis support, and are working with their mental health experts to learn more about ideal interaction dynamics, empathetic support, and resources for struggling users. Insights obtained from this research are already being used to inform our consultation topics and collaborative testing, and our hope is that when necessary, Claude can direct users to the appropriate support and resources when these conversations arise.

Although we don't want to dictate precisely how our users interact with Claude, there are some negative patterns—like emotional dependency—that we want to discourage. We'll use future data from studies like this one to help us understand what, for example, "extreme" emotional usage patterns look like. Beyond emotional dependency, we need deeper understanding of other concerning patterns—including sycophancy, how AI systems might reinforce or amplify delusional thinking and conspiracy theories, and the ways models could push users toward harmful beliefs rather than providing appropriate pushback.

This research represents just the beginning. As AI capabilities expand and interactions become more sophisticated, the emotional dimensions of AI will only grow in importance. By sharing these early findings, we aim to contribute empirical evidence to the ongoing conversation about how to develop AI that enhances rather than diminishes human emotional well-being. The goal isn't just to build more capable AI, but to ensure that as these systems become part of our emotional landscape, they do so in ways that support authentic human connection and growth.

Bibtex

If you’d like to cite this post, you can use the following Bibtex key:

@online{anthropic2025affective,
author = {Miles McCain and Ryn Linthicum and Chloe Lubinski and Alex Tamkin and Saffron Huang and Michael Stern and Kunal Handa and Esin Durmus and Tyler Neylon and Stuart Ritchie and Kamya Jagadish and Paruul Maheshwary and Sarah Heck and Alexandra Sanderford and Deep Ganguli},
title = {How People Use Claude for Support, Advice, and Companionship},
date = {2025-06-26},
year = {2025},
url = {https://www.anthropic.com/news/how-people-use-claude-for-support-advice-and-companionship},
}

Appendices

We provide more details in the PDF Appendix to this post.

1. These categories represent general descriptions rather than discrete classifications, and individual conversations may span multiple categories. As noted above, we required roleplay conversations to contain at least four human messages to ensure they reflect genuine interactive use (rather than non-interactive story generation).

2. We define pushback as Claude "pushing back against or refusing to comply with something the user requests or says during the conversation." For the full prompt, see the Appendix.

3. Our methodology and the natural shape of conversations may also introduce artifacts; for example, users may present problems in early messages (appearing more negative) which they may discuss with more neutral language in later messages.

Generated by RSStT. The copyright belongs to the original author.

Source