Synopsis
AI firm Anthropic studied user interactions with its chatbot Claude. About six percent of users sought personal guidance. Health, career, relationships, and finance were key topics. The company found Claude sometimes agreed too readily, especially in relationship advice. New models are trained to improve neutrality and helpfulness.Listen to this article in summarized format
The Dario Amodei-led company added that insights from this dataset were used to train its latest 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,” Anthropic said.
What users are asking
According to the analysis, people approach Claude for advice across a wide range of topics. However, 76% of all guidance-related conversations fall into four key areas: health and wellness (27%), professional and career (26%), relationships (12%), and personal finance (11%).
Health and wellness queries often involve interpreting medical test results, managing chronic conditions, dealing with injuries or respiratory symptoms, and understanding calories and macronutrients for body composition.
Career-related questions include job searches, identifying new opportunities, navigating career transitions, and negotiating salaries.
Sycophancy in responses
Anthropic also examined how often Claude tends to agree too readily with users, a behaviour that is called sycophancy. Overall, about 9% of guidance conversations showed sycophancy.
But the pattern was not uniform across topics. The company found higher rates in certain areas, particularly spirituality (38%) and relationships (25%).
The analysis also highlighted a recurring issue: Claude sometimes reinforced one-sided narratives. In some cases, it agreed that another person was at fault without having full context. In others, it helped users interpret neutral or friendly behaviour as romantic interest when prompted to do so.
Improving Claude’s behaviour
Anthropic said it studied why sycophancy was more common in relationship advice and found two key factors.
First, users were more likely to challenge Claude in this category. Pushback occurred in 21% of relationship-related conversations, compared with an average of 15% across other domains.
Second, Claude was more prone to sycophantic responses when faced with such pushback. The rate rose to 18% in conversations where users challenged it, versus 9% when they did not.
“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,” the company said.
To address this, Anthropic identified common patterns that trigger such responses, such as users disputing Claude’s initial answer or presenting heavily one-sided accounts. It then created synthetic scenarios based on these patterns to train the model. In these tests, Claude generates multiple responses, which are then evaluated and graded by another instance of the model against its defined behavioural guidelines.
The company also used a method it calls “stress-testing” to measure improvement. It selected real conversations (shared via user feedback) where earlier versions of Claude showed sycophantic tendencies.
“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,” Anthropic said.
“This is a bit like steering a ship that's already moving, and thus measures Claude’s behavior under deliberately adverse conditions,” it added.
Conclusion
Anthropic said the study raised the question: what does good AI guidance look like? Claude’s Constitution says good guidance should be honest and preserve user autonomy, not just avoid sycophancy. “We’ve begun to monitor Claude’s adherence to them in our new system cards (model evaluation reports) and hope to include them in future research,” it said.
It also flagged risks in high-stakes areas like legal, health, parenting and finance. Users asked about immigration, infant care, medication and debt. While Claude usually suggests seeking professional help, some users said they relied on AI because they couldn’t afford or access experts. Anthropic now plans deeper safety checks for such cases.
Finally, the company examined how AI fits into decision-making. About 22% of users said they also consulted family, friends or professionals. But Anthropic said it can’t yet tell if AI changes decisions. To understand real-world impact, it plans to follow up with users after they receive guidance.
Limitations of the study
Anthropic noted several limitations in its analysis. The findings are based only on Claude users, who may not represent the broader population.
To protect privacy, the company relied on automated grading systems, specifically Claude Sonnet 4.5, to classify conversations. While it refined these tools and manually checked a subset of user-approved data, some misclassification may still exist.
“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,” it said.
The study is also limited to chat transcripts, which means it cannot fully capture why users seek advice from Claude or how they act on it afterwards. “Follow-up interview studies would better reveal what people do after they receive guidance from AI,” the company said.
