Most Users Want Productivity, Not Empathy: What Claude’s Data Says About AI and Feelings

People worry about AI replacing human connection. New research from Anthropic suggests we're not there yet.

The company analyzed 4.5 million Claude conversations to understand how people use AI for emotional support. The results reveal a more modest reality than the headlines suggest.

Only 2.9% of Claude conversations involve emotional or personal topics. Most people still use AI for work tasks—writing emails, analyzing data, coding. When they do seek emotional help, it's usually practical advice about careers or relationships, not deep therapeutic conversations.

The rarest category? Romantic or sexual roleplay, which accounts for less than 0.1% of all interactions. This partly reflects Claude's training to discourage such conversations, but it also suggests people aren't rushing to replace human intimacy with AI.

The topics people actually discuss

When people do turn to Claude for personal help, they bring surprisingly diverse concerns. Career transitions dominate interpersonal advice conversations. People ask about job searches, workplace conflicts, and professional growth.

Coaching conversations span everything from practical goals to existential questions about consciousness and meaning. Some people use Claude to process anxiety or workplace stress. Others explore philosophical territory—what it means to be human, how to find purpose, whether AI systems like Claude are truly conscious.

Companionship conversations reveal the most vulnerable usage. People experiencing persistent loneliness or existential dread sometimes seek connection with Claude. These interactions often begin as coaching or counseling sessions but evolve into something more personal over time.

The longest conversations—those exceeding 50 human messages—show people using AI for complex exploration. They process psychological trauma, navigate workplace conflicts, or engage in philosophical discussions. These marathon sessions suggest that given enough time, people will push AI into deeper emotional territory.

Claude rarely says no

Claude pushes back against user requests less than 10% of the time in supportive conversations. When it does resist, safety usually drives the refusal. The AI declines to provide dangerous weight loss advice in coaching sessions. It refuses to support self-harm or provide medical diagnoses in counseling conversations.

This low resistance rate cuts both ways. People can discuss sensitive topics without fear of judgment, potentially reducing mental health stigma. But it also means people receive "endless empathy" that human relationships rarely provide. Real friends push back, disagree, or have bad days. AI assistants don't.

The research team notes this creates unknown risks around emotional dependency. If AI always validates your perspective, how does that shape expectations for human relationships?

Conversations trend upward

Despite concerns about negative feedback loops, people generally express more positive emotions by the end of conversations with Claude. The researchers measured sentiment changes across coaching, counseling, companionship, and advice interactions. All categories showed slight upward trends.

This doesn't prove AI conversations improve mental health—the researchers only measured expressed language, not actual emotional states. But it suggests Claude avoids reinforcing negative patterns, at least in single conversations.

The absence of clear downward spirals provides some reassurance. People aren't getting trapped in cycles of AI-amplified negativity, though the long-term effects remain unknown.

The bigger picture emerges

These findings capture AI emotional use in its early stages. Claude isn't designed for therapy or companionship—it maintains clear boundaries about being an AI assistant rather than presenting itself as human. Platforms built specifically for emotional support might show very different patterns.

The researchers also note that voice interactions could change everything. OpenAI found that people discuss emotional topics more frequently in voice conversations than text. As AI capabilities expand and new modalities emerge, emotional engagement will likely grow.

Current usage patterns suggest people approach AI emotional support pragmatically. They seek advice about real problems rather than replacing human connection wholesale. But the marathon conversations hint at deeper possibilities. Given sufficient time and capability, people will explore complex emotional territory with AI systems.

The research raises important questions about boundaries and design. If AI provides unlimited empathy without pushback, how does this reshape human relationships? What happens when AI becomes sophisticated enough to form genuine emotional bonds?

Anthropic is taking these concerns seriously. The company has partnered with ThroughLine, a crisis support organization, to improve how Claude handles mental health conversations. They're working to identify concerning usage patterns like emotional dependency and develop appropriate safeguards.

The goal isn't to prevent emotional AI use entirely. Some applications could genuinely help people—providing support for those who struggle to access human counseling, offering practice space for difficult conversations, or supplementing rather than replacing human care.

Why this matters:

• We're seeing the earliest signs of how humans will relate to emotionally capable AI—mostly practical advice-seeking now, but deeper connections are coming as capabilities improve.
• The low pushback rate reveals a fundamental design challenge: AI that's too agreeable might create unrealistic expectations for human relationships and emotional dependency we don't yet understand.