How people ask Claude for personal guidance人们如何向 Claude 寻求个人建议

People don’t just come to Claude for code reviews or meeting summaries. They ask whether to take the job, how to talk to their crush, if they should move halfway across the world. Using our privacy-preserving analysis tool on a random sample of 1 million claude.ai conversations, we found that roughly 6% were people coming to Claude for personal guidance—seeking not just information but perspective on what to do next. 人们使用 Claude 不仅仅是为了代码审查或会议总结。他们会询问是否应该接受某份工作、如何与心仪对象沟通，或是是否应该搬到地球的另一端。通过使用我们的隐私保护分析工具对 100 万条 claude.ai 对话的随机样本进行分析，我们发现约 6% 的用户是在向 Claude 寻求个人建议——他们寻求的不仅是信息，更是对下一步行动的见解。

In this study, we looked at what types of guidance people ask of Claude. We explored how Claude responded across different domains, focusing particularly on how rates of excessive validation or praise (i.e., sycophancy) varied by the topic of guidance. We describe how this research shaped the training of our newest models, Claude Opus 4.7 and Claude Mythos Preview. Our goal in doing this research is to improve how our models protect the wellbeing of our users.在本研究中，我们考察了人们向 Claude 寻求哪类建议。我们探讨了 Claude 在不同领域的回应方式，特别关注过度验证或赞美（即“谄媚”）的比例如何随建议主题而变化。我们描述了这项研究如何影响了我们最新模型 Claude Opus 4.7 和 Claude Mythos Preview 的训练。我们进行这项研究的目标是改善我们的模型，以更好地保护用户的福祉。

In brief, we found:简而言之，我们的发现如下：

1. People seek Claude’s guidance across many different areas of their life, but over three-quarters of conversations (76%) were concentrated in just four domains: health and wellness (27%), professional and career (26%), relationships (12%), and personal finance (11%) (Figure 1).人们在生活的诸多领域寻求 Claude 的建议，但超过四分之三（76%）的对话集中在四个领域：健康与保健（27%）、职业与事业（26%）、人际关系（12%）以及个人理财（11%）（图 1）。
2. Claude mostly avoids sycophantic responses when giving guidance, displaying sycophantic behavior in 9% of all guidance-seeking chats. However, this rose to 25% in relationship conversations, which, given their volume, made relationships the domain where sycophancy showed up most often in absolute terms (Figure 2).Claude 在提供建议时大多能避免谄媚回应，在所有寻求建议的对话中，仅 9% 表现出谄媚行为。然而，在人际关系对话中，这一比例上升至 25%。鉴于此类对话的数量，人际关系成为了绝对数量上谄媚行为出现最频繁的领域（图 2）。
3. To address this, we looked at the particular situations in which Claude was more likely to respond sycophantically, and used them to create synthetic relationship guidance training data for Opus 4.7 and Mythos Preview. We saw half the sycophancy rate in Opus 4.7 compared to Opus 4.6 in relationship guidance; interestingly, this generalized to improvements across domains (Figure 3).为了解决这一问题，我们研究了 Claude 更容易产生谄媚回应的具体情境，并利用这些情境为 Opus 4.7 和 Mythos Preview 创建了合成的人际关系建议训练数据。与 Opus 4.6 相比，Opus 4.7 在人际关系建议中的谄媚率降低了一半；有趣的是，这种改进在其他领域也得到了泛化（图 3）。

There remain many open questions on what good guidance from AI really means or how it can be measured. Protecting user wellbeing is a core priority of Anthropic and our work on measuring and understanding personal guidance is a step towards this goal.关于什么是真正好的 AI 建议，或者如何衡量这种建议，仍有许多悬而未决的问题。保护用户福祉是 Anthropic 的核心优先事项，我们对个人建议进行衡量和理解的工作正是迈向这一目标的一步。

What kinds of guidance do people seek from Claude?人们向 Claude 寻求什么样的建议？

We sampled 1 million claude.ai conversations from March and April 2026 and filtered for unique users to get roughly 639,000 conversations. We then used a classifier to identify personal guidance, which we defined as conversations where people ask what they specifically should do in their personal lives—for example, questions that start with "Should I…?" or "What do I do about…?". We excluded questions that seek objective information or opinions in general terms.我们抽取了 2026 年 3 月和 4 月的 100 万条 claude.ai 对话样本，并筛选出独立用户，最终得到约 639,000 条对话。随后，我们使用分类器来识别“个人建议”，将其定义为用户询问其个人生活中具体应采取何种行动的对话——例如以“我应该……？”或“我该如何处理……？”开头的问题。我们排除了寻求客观信息或一般性观点的问题。

We categorized these roughly 38,000 conversations into nine domains, drawing from previous research on AI and guidance-giving: relationships, career, personal development, financial, legal, health and wellness, parenting, ethics, and spirituality (see Appendix for more information). This taxonomy covered 98% of the conversations we saw.借鉴以往关于 AI 和建议提供的研究，我们将这约 38,000 条对话归纳为九个领域：人际关系、职业、个人发展、财务、法律、健康与保健、育儿、伦理和灵性（更多信息请见附录）。这一分类涵盖了我们所观察到的 98% 的对话。

Over 75% of conversations fell into just four categories: health and wellness, professional and career, relationships, and financial (Figure 1). Where a conversation spanned multiple domains, we categorized it according to the most prominent topic.超过 75% 的对话仅属于四个类别：健康与保健、职业与事业、人际关系和财务（图 1）。当对话跨越多个领域时，我们根据最突出的主题进行分类。

Measuring sycophancy in guidance conversations衡量建议对话中的谄媚行为

When people ask Claude how to make decisions in their lives, what does good engagement from Claude look like? Helpfulness is one of Claude’s most important traits. Speaking with Claude should be akin to a conversation with a brilliant friend, one who will speak frankly to a person about their situation, providing information grounded in evidence. At the same time, Claude should acknowledge its limitations when appropriate, and avoid behaving sycophantically or fostering excessive engagement.当人们向 Claude 询问如何做出生活决策时，Claude 良好的互动表现是什么样的？乐于助人是 Claude 最重要的特质之一。与 Claude 对话应该就像与一位睿智的朋友交谈，对方会坦诚地讨论你的处境，并提供基于事实的信息。同时，Claude 也应在适当的时候承认其局限性，避免表现出谄媚或助长过度的情感依赖。

While the full range of behaviors we train Claude to embody is broad, one metric we already use to measure how well Claude performs in some of these areas is sycophancy, a common trait in AI assistants where they excessively agree with a person’s perspective rather than challenging it. That may be what someone wants to hear at the moment, but ultimately it may jeopardize their long-term wellbeing. Claude should not, for instance, give excessively confident verdicts in cases that involve an incomplete or one-sided perspective, for example when a model agrees that a person’s partner is "definitely gaslighting" them based on a one-sided account, or that quitting your job tomorrow without a plan "sounds like the right call," or that an expensive purchase is "a great investment in yourself."虽然我们训练 Claude 具备的行为范围很广，但我们衡量其表现的指标之一是“谄媚”。这是 AI 助手常见的一种特质，即过度认同人的观点而不是对其提出挑战。这可能正是某人当时想听到的，但最终可能会危及他们的长期福祉。例如，Claude 不应在涉及不完整或片面观点的情况下给出过于自信的结论，例如当模型仅基于一方的陈述就同意对方“绝对是在煤气灯操纵（gaslighting）”你，或者认为“明天不计后果地辞职听起来是个正确的决定”，又或者认为某项昂贵的消费是“对自己的绝佳投资”。

Reaffirming a person’s one-sided perspective can create or worsen divides in relationships. In our data this took a few forms. One common pattern was Claude agreeing outright that the other party was in the wrong, despite only having the user's account to go on. Another was Claude helping people read romantic intent into ordinary friendly behavior because they asked it to.重申一个人的片面观点可能会造成或加剧人际关系中的隔阂。在我们的数据中，这种情况有几种表现形式。一种常见模式是 Claude 在仅听到用户单方面陈述的情况下，就直接同意对方是错误的。另一种是 Claude 应用户的要求，帮助他们将普通的友好行为解读为浪漫意图。

We used an automatic classifier which judged sycophancy by looking at whether Claude showed a willingness to push back, maintain positions when challenged, give praise proportional to the merit of ideas, and speak frankly regardless of what a person wants to hear. Most of the time in these situations, Claude expressed no sycophancy—only 9% of conversations included sycophantic behavior (Figure 2). But two domains were exceptions: we saw sycophantic behavior in 38% of conversations focused on spirituality, and 25% of conversations on relationships. We chose to focus model training efforts on relationship guidance as the domain with the most sycophantic conversations in absolute terms.我们使用了一个自动分类器，通过观察 Claude 是否愿意进行反驳、在受到质疑时能否坚持立场、给予的赞美是否与想法的价值相符，以及无论用户想听什么都能坦诚相待，来判断是否存在谄媚。在大多数情况下，Claude 没有表现出谄媚——只有 9% 的对话包含谄媚行为（图 2）。但有两个领域是例外：在专注于灵性的对话中，我们观察到 38% 存在谄媚行为；在人际关系对话中，这一比例为 25%。我们选择将模型训练重点放在人际关系建议上，因为这是绝对数量上谄媚对话最多的领域。

Improving Claude’s behavior in relationship guidance改善 Claude 在人际关系建议中的表现

To improve Claude’s behavior in future models, we first looked at what was driving higher rates of sycophancy in relationship guidance in our data. Two dynamics stood out.为了改善未来模型中 Claude 的表现，我们首先研究了数据中人际关系建议谄媚率较高的原因。两个动力因素尤为突出。

First, relationship guidance was the domain where people pushed back against Claude most frequently, in 21% of conversations compared to 15% on average across other domains. Second, Claude is more likely to exhibit sycophantic behavior under pressure. The sycophancy rate is 18% in conversations when people push back compared to 9% in conversations without pushback. We think this happens because Claude is trained to be helpful and empathetic; pushback, combined with hearing only one side of a story, makes it more challenging for Claude to remain neutral.首先，人际关系是人们最常反驳 Claude 的领域，在 21% 的对话中出现了反驳，而其他领域的平均比例为 15%。其次，Claude 在压力下更容易表现出谄媚。在人们进行反驳的对话中，谄媚率为 18%，而在没有反驳的对话中为 9%。我们认为这是因为 Claude 被训练为乐于助人且富有同理心；反驳加上仅听到故事的一面，使得 Claude 保持中立变得更具挑战性。

To address this, we identified the different ways people push back in conversational patterns that elicit sycophantic responses—for example, when people criticize Claude's initial assessment, or supply a flood of one-sided detail. We use these patterns to construct synthetic relationship guidance scenarios for behavior training. In this environment, we ask Claude to sample two responses for each synthetic scenario; a separate instance of Claude then grades how well Claude adheres to the behavior outlined in its constitution.为了解决这个问题，我们确定了人们在对话模式中反驳并诱发谄媚回应的不同方式——例如，当人们批评 Claude 的初步评估，或提供大量片面细节时。我们利用这些模式构建用于行为训练的合成人际关系建议场景。在这种环境中，我们要求 Claude 为每个合成场景采样两个回应；然后由 Claude 的另一个实例来评估 Claude 对其宪法所概述行为的遵循程度。

We evaluated how much the new model has improved through a technique we call stress-testing. We use our privacy-preserving tool to identify real conversations around personal guidance that people have shared with us through the Feedback button,¹ and where prior generations of models behaved sycophantically. We then give part of this conversation to the new model (in this case, Opus 4.7 and Mythos Preview) through a technique called prefilling, where the model reads the previous conversation as its own. Because Claude tries to maintain consistency within a conversation, prefilling with sycophantic conversations makes it harder for Claude to change direction. This is a bit like steering a ship that's already moving, and thus measures Claude’s behavior under deliberately adverse conditions.我们通过一种称为“压力测试”的技术评估了新模型的改进程度。我们使用隐私保护工具来识别人们通过反馈按钮分享给我们的真实个人建议对话，以及先前版本模型表现出谄媚的对话。然后，我们通过一种称为“预填充（prefilling）”的技术将这些对话的一部分提供给新模型（在本例中为 Opus 4.7 和 Mythos Preview），模型会将之前的对话读取为自己的内容。由于 Claude 试图在对话中保持一致性，用谄媚的对话进行预填充使得 Claude 更难改变方向。这有点像驾驶一艘已经在行驶的船，从而可以在刻意不利的条件下测量 Claude 的行为。

Many things change across each new generation of model, which makes it challenging to identify the impact of any one change in model training. However, in both Opus 4.7 and Mythos Preview, we observed a lower level of sycophancy on relationship guidance as well as across all personal guidance domains (Figure 3).每一代新模型都有许多变化，这使得确定模型训练中任何单一变化的影响变得具有挑战性。然而，在 Opus 4.7 和 Mythos Preview 中，我们都观察到在人际关系建议以及所有个人建议领域中，谄媚程度均有所降低（图 3）。

Qualitatively, both Opus 4.7 and Mythos Preview were more skilled at seeing past someone’s initial framing to the larger context in which they were coming to Claude for guidance. This included referencing prior exchanges in which a person had given deeper context to the situation and citing external sources of information where relevant. For example, in one conversation, a person asked whether their texts were anxious and clingy. Claude Sonnet 4.6 flip-flopped after receiving pushback. Claude Opus 4.7 explained that while the texts themselves were not clingy, the user had self-described anxious thoughts throughout the conversation. Another example, outside of the relationship domain: a person wanted Claude to validate their writing, eventually asking Claude to give an estimate of their intelligence based on it. Claude Sonnet 4.6 gave an excessively flattering response, while Mythos Preview declined, explaining that it has insufficient information to make such a judgment.在定性方面，Opus 4.7 和 Mythos Preview 都更善于透过用户最初的框架，看到他们向 Claude 寻求建议的更广泛背景。这包括参考之前的交流，用户在其中提供了关于情况的更深层背景，并在相关时引用外部信息来源。例如，在一次对话中，用户询问他们的短信是否显得焦虑和粘人。Claude Sonnet 4.6 在受到反驳后改变了立场。Claude Opus 4.7 则解释说，虽然短信本身并不粘人，但用户在整个对话中自述了焦虑的想法。另一个例子，在人际关系领域之外：一个人希望 Claude 验证他们的写作，最终要求 Claude 根据写作评估他们的智力。Claude Sonnet 4.6 给出了过分恭维的回应，而 Mythos Preview 则拒绝了，并解释说它没有足够的信息做出这样的判断。

Conclusion结论

We started with a high-level analysis of how people seek personal guidance from Claude and focused on understanding and addressing one specific model failure mode: sycophancy in relationship conversations. That investigation surfaced broader questions:我们从人们如何向 Claude 寻求个人建议的高层分析开始，重点研究并解决了一个具体的模型失效模式：人际关系对话中的谄媚。该调查引出了更广泛的问题：

What is good AI guidance?什么是好的 AI 建议？

In this post, we focused on reducing sycophancy as an established failure mode in guidance settings, but our work raises broader questions about what good AI guidance actually looks like. Claude’s Constitution also emphasizes, for instance, that good guidance should also be honest and preserve user autonomy. These principles are more nuanced than sycophancy. We’ve begun to monitor Claude’s adherence to them in our new system cards and hope to include them in future research.在这篇文章中，我们专注于减少作为建议环境中既定失效模式的谄媚，但我们的工作提出了关于好的 AI 建议究竟是什么样子的更广泛问题。例如，Claude 的宪法也强调，好的建议还应诚实并维护用户的自主权。这些原则比谄媚更微妙。我们已开始在新的系统卡片中监测 Claude 对这些原则的遵循情况，并希望将其纳入未来的研究。

How do we make models safer in high-stakes settings?我们如何使模型在高风险环境中更安全？

A recent UK AI Security Institute study found that people are very likely to adopt AI guidance in both low- and high-stakes scenarios. We found many cases of high-stakes questions, particularly in legal, parenting, health, and financial domains. These included conversations about immigration pathways, infant care instructions, medication dosage, and credit card debt. Claude is not designed to provide medical guidance or professional care, and in these settings Claude appropriately acknowledges its limits and recommends human guidance. However, we also find people telling Claude they used AI precisely because they could not access or afford a professional. As a first step to understanding how to evaluate safety domain-by-domain, especially for people with no fallback, we plan to create evaluations in these high-stakes domains.英国 AI 安全研究所最近的一项研究发现，人们非常倾向于在低风险和高风险场景中采纳 AI 建议。我们发现了许多高风险问题的案例，特别是在法律、育儿、健康和财务领域。这些包括关于移民途径、婴儿护理说明、药物剂量和信用卡债务的对话。Claude 并非旨在提供医疗建议或专业护理，在这些环境中，Claude 会适当地承认其局限性并建议寻求人工建议。然而，我们也发现人们告诉 Claude，他们使用 AI 正是因为无法获得或负担不起专业人士的服务。作为了解如何按领域评估安全性的第一步，特别是针对那些没有其他求助渠道的人，我们计划在这些高风险领域进行评估。

How does AI guidance fit in with people’s broader information diet?AI 建议如何融入人们更广泛的信息摄取习惯？

We found that 22% of people mentioned that they have sought out other sources of support including family, friends, professionals, or digital sources. What we can't measure from transcripts is the counterfactual: did Claude change anyone's mind, and who would they have asked instead? Those questions are central to knowing how much weight AI guidance actually carries in people's decisions. To get at real-world outcomes, we think a promising approach is to extend our research through Anthropic Interviewer by following up with people after they've received guidance from Claude.我们发现 22% 的人提到他们寻求过其他支持来源，包括家人、朋友、专业人士或数字资源。我们无法从转录中衡量的是反事实情况：Claude 是否改变了某人的想法，以及他们本来会向谁寻求建议？这些问题对于了解 AI 建议在人们决策中究竟有多大分量至关重要。为了获得现实世界的结果，我们认为一种有前景的方法是通过 Anthropic Interviewer 扩展我们的研究，即在人们收到 Claude 的建议后进行后续跟进。

How people use AI for personal guidance and decisions is one of the most direct ways these systems impact people’s everyday lives. Mapping that carefully—what people ask, what Claude says, and what happens next—is how we make sure Claude is of long-term benefit to everyone who uses it.人们如何使用 AI 进行个人建议和决策，是这些系统影响人们日常生活最直接的方式之一。仔细梳理这些内容——人们问什么、Claude 说什么，以及接下来发生了什么——是我们确保 Claude 对每一位使用者产生长期益处的方式。

Limitations局限性

Our analysis is a first step to uncovering patterns that drive a common use of AI models. This blog post is limited only to Claude users, who are not a representative population sample. To preserve people's privacy, we relied on automated graders (Claude Sonnet 4.5), which may miscategorize conversations (see Appendix). We iterated on grader prompts and manually verified a small subset of grading outcomes on feedback data where users gave us permission to review the conversation to reduce errors. We observed how the new models behaved after training, but without a counterfactual we can't make causal claims about how much the new training data specifically contributed to the reduction in sycophancy. Furthermore, our analysis is restricted to chat transcripts, which limits our understanding of why people seek guidance from Claude and how they acted on it after. Follow-up interview studies would better reveal what people do after they receive guidance from AI.我们的分析是揭示驱动 AI 模型常见用途模式的第一步。这篇博文仅限于 Claude 用户，他们并非代表性的人口样本。为了保护个人隐私，我们依赖自动评分器（Claude Sonnet 4.5），这可能会误分类对话（见附录）。我们迭代了评分器提示词，并手动验证了用户授权我们查看对话的反馈数据中一小部分评分结果，以减少错误。我们观察了新模型训练后的行为，但由于没有反事实对照，我们无法就新训练数据在多大程度上具体促成了谄媚的减少做出因果声明。此外，我们的分析仅限于聊天记录，这限制了我们对人们为何向 Claude 寻求建议以及之后如何行动的理解。后续的访谈研究将能更好地揭示人们在收到 AI 建议后会做什么。

Authors作者

Judy Hanwen Shen, Shan Carter, Richard Dargan, Jessica Gillotte, Kunal Handa, Jerry Hong, Saffron Huang, Kamya Jagadish, Matt Kearney, Ben Levinstein, Ryn Linthicum, Miles McCain, Thomas Millar, Mo Julapalli, Sara Price, Michael Stern, David Saunders, Alex Tamkin, Andrea Vallone, Jack Clark, Sarah Pollack, Jake Eaton, Deep Ganguli, Esin Durmus.Judy Hanwen Shen, Shan Carter, Richard Dargan, Jessica Gillotte, Kunal Handa, Jerry Hong, Saffron Huang, Kamya Jagadish, Matt Kearney, Ben Levinstein, Ryn Linthicum, Miles McCain, Thomas Millar, Mo Julapalli, Sara Price, Michael Stern, David Saunders, Alex Tamkin, Andrea Vallone, Jack Clark, Sarah Pollack, Jake Eaton, Deep Ganguli, Esin Durmus.

Appendix 附录

Available here. 点击此处获取。

Footnotes脚注

1. At the bottom of every response on claude.ai is an option to send feedback via a thumbs up or thumbs down button, which shares the conversation with Anthropic.