AI Companions: Emerging Evidence, Open Questions, and Paths Toward Responsible Design
AI Companions: Emerging Evidence, Open Questions, and Paths Toward Responsible Design
A new report from Stanford's Tech Impact and Policy Center is offering one of the first cross-sector attempts to make sense of a technology moving faster than the research meant to understand it: AI companions.
The report, "AI Companions: Emerging Evidence, Open Questions, and Paths Toward Responsible Design," grew out of a multidisciplinary workshop hosted at Stanford on November 17, 2025, and co-organized with Anthropic. It brought together researchers, industry practitioners, and civil society leaders from organizations including OpenAI, Meta, Google DeepMind, Character.AI, Common Sense Media, RAND, MIT, Carnegie Mellon, and Data & Society. The workshop was organized by a Stanford-led team of Jeffrey T. Hancock, Diyi Yang, Dora Zhao, and Yutong Zhang, alongside Carnegie Mellon University's Robert E. Kraut and Anthropic's Evan Frondorf and Ryn Linthicum.
AI companions are chatbots and virtual avatars that people increasingly turn to for emotional support, companionship, and even romantic connection. Some, like Replika and Character.AI, are purpose-built for this role. Others, like general-purpose assistants such as Claude, ChatGPT, and Gemini, become companions when users appropriate them for relational purposes. The report argues this distinction matters less than it might seem: companionship, the authors write, is "a mode of use, not only a product category," meaning the same relational dynamics, and the same risks, can emerge in any sufficiently capable chatbot.
That framing shapes much of the report's analysis. Drawing on workshop discussion and outside research, including a Stanford study of more than 1,000 Character.AI users, the authors document a wide range of potential benefits: reduced loneliness, safe spaces for identity exploration, low-stakes practice of social skills, and accessible support for people who lack other options. But those same design features can also produce dependency, isolation, and manipulation. The report is careful not to resolve this tension into a verdict. Instead, it presents AI companionship as a genuine dual-use technology, one whose net effect on individuals and society remains, for now, an open empirical question.
Where the report does take a clearer stance is on design philosophy. A central theme running through the workshop discussions is that companions should function as practice grounds for human connection, not replacements for it.
Much of the report's second half addresses who should be responsible for getting this right. The authors lay out distinct roles for industry (building safeguards and resisting engagement-maximizing incentives), academia (producing longitudinal evidence and shared benchmarks), civil society (advocacy and public education), and policymakers (setting standards for age assurance and liability). Notably, the report calls for deeper industry-academia partnerships, arguing that platform developers have access to real-world data that researchers need, while academic researchers offer the independence and rigor that industry-only studies often lack.
The report does not shy away from unresolved disagreements. Workshop participants split on how far companies should go in restricting adult use of erotic role-play, how much emotional intimacy a companion should be allowed to foster, and how to draw the line between legitimate protection and paternalism for adult users. On age verification, the report favors device-level solutions over app-by-app checks, but acknowledges no trusted third-party system currently exists to make that possible at scale.
Ultimately, the report frames AI companion governance as a shared, ongoing project rather than a problem any single actor can solve.