Anthropic studies how AI gives personal advice

Claude does more than summarize meetings or review code. It also helps people decide whether to accept a job, how to talk to someone they like, or whether moving to another country makes sense. An Anthropic analysis found that around 6% of conversations on Claude.ai seek guidance on personal decisions, not just information.

The study examined a random sample of 1 million conversations from March and April 2026. After filtering for unique users, about 639,000 conversations remained. An automated system identified nearly 38,000 in which people were essentially asking, "what should I do?"

Health, work and relationships account for most requests

Anthropic classified those conversations into nine areas. More than three-quarters, or 76%, centered on four topics:

• Health and well-being: 27%
• Work and career: 26%
• Relationships: 12%
• Personal finance: 11%

The rest covered personal development, legal matters, parenting, ethics and spirituality. In practice, this means many people are already using AI as a first stop for thinking through decisions they might once have discussed with a friend, family member or professional.

Some questions are everyday concerns. Others have serious consequences: immigration pathways, infant care, medication dosages or credit card debt. Anthropic acknowledges that Claude is not designed to replace medical care or professional advice, and says the model should flag its limits and recommend human help when necessary.

But that creates a challenge: some people turn to AI precisely because they cannot afford or access a specialist.

The problem with telling users what they want to hear

Anthropic focused its research on excessive agreeableness, known in English as sycophancy. This happens when an assistant confirms too much of the user's version of events instead of examining it carefully or raising objections.

For example, a model should not claim that a partner is "definitely manipulating" the user based on one side of the story. Nor should it recommend quitting a job the next day without a plan or present an expensive purchase as a major personal investment simply because that is what the person wants to hear.

Across the guidance conversations, Claude showed this behavior in 9% of cases. The figure rose to 25% in relationship conversations, the area where agreeableness appeared most often in absolute terms. In spirituality, it reached 38%, although that group was smaller.

Relationships appear to be especially difficult because users challenged Claude's initial response in 21% of conversations, compared with 15% for other topics. When someone pressured the model to change its mind, the agreeableness rate rose from 9% to 18%.

What Anthropic changed in its models

To train its models, Anthropic created synthetic situations inspired by those patterns. These included conversations in which the user criticized Claude's first response or provided many details that favored a single interpretation of the story.

It then tested the new models on real conversations shared through the feedback button. The technique included showing them exchanges in which earlier versions had already been too agreeable, as if the model had to correct the course of a conversation that was already heading in that direction.

According to Anthropic, Claude Opus 4.7 cut the agreeableness rate in half compared with Opus 4.6 in relationship queries. The improvement also appeared in other areas of personal guidance. The company cautions, however, that it cannot show how much of that reduction is specifically due to the new training data.

In one example, a person asked whether their messages seemed anxious or dependent. An earlier version changed its position after being pressured. Opus 4.7 distinguished between the messages, which did not seem dependent on their own, and the anxious thoughts the person had described throughout the rest of the conversation.

In another case, a user asked for validation of their writing and ended up requesting an estimate of their intelligence. Mythos Preview refused to make that judgment because it did not have enough information, rather than responding with exaggerated praise.

For anyone using AI, the important takeaway is simple: a pleasant answer is not always a responsible one. Claude should help you think, not automatically confirm that you are right.

Anthropic also acknowledges that it still does not know how much these conversations influence real-world decisions. 22% of people said they had sought additional support from family members, friends, professionals or other digital sources, but the chats cannot show what each user did afterward or what they would have consulted if Claude had not been available.

The next step will be to measure the safety of guidance in high-risk areas such as health, finance, legal matters and parenting. The question is no longer just what people ask AI, but how much weight they give its answers when they have no other option.