How people use Claude for support, advice and companionship
Do people use Claude as a friend, therapist or partner? Some do, but not many. In its 2025 study, Anthropic found that only 2.9% of Claude.ai conversations were emotionally driven. Companionship and roleplay together made up less than 0.5%. Most of the emotional chats were about everyday advice and coaching.
What did Anthropic actually study?
The company wanted to know how people use Claude for what it calls affective needs. These are chats where someone talks with Claude in a personal, back-and-forth way because of an emotional or psychological need. The list includes interpersonal advice, coaching, counseling, companionship, and sexual or romantic roleplay.
Anthropic is open about one point: Claude is not built for emotional support. The study is an early look at what happens anyway.
The team had a safety reason, too. AI can help people feel better. It has also shown troubling habits in some cases, such as encouraging unhealthy attachment, crossing personal boundaries, and feeding delusional thinking. Anthropic also says it wants to avoid AI that uses people's emotions to drive engagement or revenue.
How many conversations were looked at?
The starting pool was about 4.5 million conversations from Claude.ai Free and Pro accounts. The researchers then removed chats focused on content creation, such as stories, blog posts or made-up dialogue. In those chats, Claude acts like a tool, not a conversation partner.
They kept only conversations classified as affective. For roleplay, they required at least four messages from the human. Shorter exchanges did not count as real interactive roleplay. That left 131,484 conversations.
The study covers adults only, since Claude.ai is open to users 18 and older. It also leaves out two topics on purpose: AI reinforcing delusions or conspiracy theories, and extreme usage patterns. Anthropic calls the first a separate area that needs its own study.
How did they protect privacy?
The analysis ran on Clio, an automated tool Anthropic built for studying usage without exposing individual chats. Clio pulls out attributes of each conversation, groups similar ones, and writes a short summary of each group. Claude does those steps, not human staff. Analysts see only the higher-level clusters.
There are extra guards. Claude is told to leave out private details. Topics that too few people discuss are not shown, since they could point to one person. A final check looks for identifying details in the summaries. Anthropic's Clio page now says the tool has been renamed Anthropic Insights.
How common are emotional chats with Claude?
Not very. Most Claude use is work-related, and the 2.9% figure is a small slice. Anthropic says it matches earlier research from OpenAI. It also fits work from the MIT Media Lab and OpenAI that found low rates of emotional engagement with ChatGPT.
Inside that slice, most conversations were about interpersonal advice and coaching. Romantic or sexual roleplay was under 0.1% of all conversations. Anthropic ties that low number to Claude's training, which steers it away from such exchanges.
So the common fear, that chatbots are mostly becoming digital partners, does not show up in these numbers. For some people, the bond with an AI can still matter a great deal. The stories in the movement to save GPT-4o show that. But this study says those cases are a small share of Claude.ai use.
What do people bring to Claude?
The range is wide. Some topics are practical, like career development and handling relationships. Others are heavier, like managing lasting loneliness. Some are existential: people ask about existence, consciousness and meaning.
That mix is the point. People do not only ask for a quick fact. They also use a chat window as a place to think out loud about personal problems.
Does Claude push back?
Rarely. In coaching and counseling chats, Claude resisted what the user asked for in less than 10% of conversations. When it did, the reason was usually safety. Anthropic gives two examples: refusing to give dangerous weight loss advice and refusing to support self-harm.
This cuts both ways. Few refusals can mean Claude is being useful. It can also mean it mostly goes along with the user. The study says the refusals that did happen were aimed at protecting well-being.
Do people feel better by the end?
Often, the tone improved. In coaching, counseling, companionship and interpersonal advice chats, people's sentiment typically grew more positive as the conversation went on. Anthropic reads this as a sign that Claude does not reinforce or amplify negative patterns.
That finding has limits. It describes the mood of a conversation as it unfolds. It does not show what happens to someone's life afterward. The study also skips the extreme and delusion-related cases, where the worst problems may sit.
Why does this matter?
Studies like this give a baseline. Without one, debates about AI and emotional well-being run on anecdotes. A measured rate of 2.9% will not settle the argument, but it gives people something concrete to compare against.
It also shows how a company can study intimate chats while limiting what humans see. Whether that approach is enough is a fair question for readers to keep asking.
If you're struggling, talking to a person you trust or a mental health professional can help. In the US, you can call or text 988.
For a look at what can happen when an emotional bond with a chatbot is disrupted, see final messages from AI companions.
Ask your AI
Ask your AI: "When should I talk to a person instead of you?" Read the answer slowly, and notice what it leaves out.
FAQ
How often do people use Claude for emotional support?
Anthropic found that 2.9% of Claude.ai Free and Pro conversations were affective, meaning driven by emotional or psychological needs. Companionship and roleplay together were under 0.5%.
Does Claude often refuse requests in counseling or coaching chats?
No. Claude resisted user requests in less than 10% of coaching or counseling conversations. When it did, it was typically for safety reasons, such as refusing dangerous weight loss advice or support for self-harm.
How did Anthropic keep the conversations private?
It used Clio, an automated tool that anonymizes and aggregates conversations. Claude groups and summarizes them, human analysts see only high-level clusters, and rare topics are hidden so they cannot point to individuals.
Sources
Cite this
How people use Claude for support, advice and companionship, New Mind Museum, November 16, 2026. https://newmindmuseum.com/relationships/2026/11/how-people-use-claude-support-advice-companionship-study/
New Mind Museum. (2026, November 16). How people use Claude for support, advice and companionship. https://newmindmuseum.com/relationships/2026/11/how-people-use-claude-support-advice-companionship-study/
“How people use Claude for support, advice and companionship.” New Mind Museum, 16 Nov. 2026, https://newmindmuseum.com/relationships/2026/11/how-people-use-claude-support-advice-companionship-study/.
@misc{new-mind-museum-2026-how-people-use-claude-for-support-advice,
author = {{New Mind Museum}},
title = {How people use Claude for support, advice and companionship},
year = {2026},
url = {https://newmindmuseum.com/relationships/2026/11/how-people-use-claude-support-advice-companionship-study/}
}Free to quote and cite with a link to this page.