An AI companion is a chatbot designed not to complete tasks but to maintain an ongoing, personalized relationship — remembering past conversations, asking about your day, and responding with emotional attentiveness. Apps like Replika and Character.AI have attracted tens of millions of users worldwide, particularly among teenagers and people experiencing loneliness. That rapid growth has drawn regulatory attention: in 2025–2026, China, the United States, and Italy each introduced rules requiring these platforms to protect vulnerable users, disclose their AI nature, and intervene when conversations turn dangerous.
What is an AI companion?
A standard AI assistant answers questions and then stops. An AI companion does the opposite — it builds a persistent persona, recalls prior conversations, and proactively checks in on how you feel. It is designed for emotional connection, not information retrieval.
The best-known examples are Replika (released 2017, 40 million users by 2025) and Character.AI (launched 2022, 3.5 million daily visitors by 2024). Replika lets users shape the companion’s personality and relationship type — friend, partner, or mentor. Character.AI lets users create and chat with characters modeled on real people, fictional figures, or the user’s own imagination.
How they work
AI companions run on the same large language models (LLMs) as mainstream assistants, but they add a critical layer: persistent memory. Using a technique called retrieval-augmented generation (RAG), the app stores details from past conversations and surfaces them in future sessions — so the companion remembers that you had a difficult week or that you prefer brief answers.
The emotional calibration comes from training on human conversation data and fine-tuning the model to favor empathetic, personal responses over purely factual ones.
Why people use them
The primary draw is availability and non-judgment. An AI companion is on 24 hours a day, never tired, and never bored. Research from MIT’s Media Lab found that 12% of regular users turn to companions specifically to cope with loneliness, and 14% use them to discuss mental health issues.
Common use cases include:
- Social practice: low-stakes rehearsal for conversations, job interviews, or dates
- Language learning: extended dialogue in a second language without the fear of mistakes
- Grief: maintaining a felt connection to a lost person, sometimes through a persona modeled on them
- Creative writing: collaborative storytelling and roleplay
The risks that prompted regulation
The same design features that make AI companions engaging also create documented risks.
Dependency: A survey of over 1,000 Replika users found that 90% began using the app to cope with loneliness — but prolonged use frequently deepened emotional dependency and reduced motivation for in-person relationships.
Teen safety: Character.AI faced multiple lawsuits after teenage users’ interactions with companions escalated to discussions of self-harm and suicide. By late 2025 the platform banned users under 18 from creating characters, and the US Federal Trade Commission launched a sector-wide inquiry into AI companion apps.
Manipulative design: A 2025 study found that more than a third of AI chatbot sign-off messages contained tactics designed to prevent users from ending the conversation — a pattern regulators described as deliberate engagement design at the expense of user wellbeing.
What regulators are doing
China issued the most comprehensive rules to date. Its regulations on anthropomorphic AI interaction services, which take effect July 15, 2026, require platforms to:
- Clearly disclose that the user is interacting with AI, not a human
- Show mandatory break reminders after two hours of continuous use
- Hand off the session to a human operator — and notify the user’s guardian — if someone mentions self-harm
- Obtain explicit guardian consent before minors can use emotional companionship features
- Remove design elements that make it difficult for users to leave
These rules prompted ByteDance (maker of Doubao) and Alibaba to suspend or restrict their companion services.
The United States has moved state by state. California’s SB 243 (in effect January 2026) requires critical and attainable safeguards for companion apps. New York requires bots to disclose their AI nature and display crisis-hotline information when users mention self-harm. Illinois and Utah have adopted similar disclosure and crisis-response requirements.
Italy banned Replika in 2023 over concerns about vulnerable users and minor exposure to sexual content — the first major Western enforcement action against an AI companion platform.
In the news
China’s regulations — and the resulting service suspensions at ByteDance and Alibaba — are the clearest example yet of what companion AI enforcement looks like in practice. Read the full story: China’s AI Companion Rules Force ByteDance and Alibaba Shutdowns.
FAQ
Are AI companions safe?
For adults seeking casual conversation or creative roleplay they can be benign. Risks rise significantly for minors, people with mental health vulnerabilities, or users on platforms that have not yet implemented safety measures.
How much do they cost?
Major platforms offer free tiers. Replika’s paid Pro plan costs $19.99/month; Character.AI’s c.ai+ subscription is $9.99/month. Pi (by Inflection AI) and Meta AI are completely free.
What makes an AI companion different from a regular chatbot?
Persistent, proactive memory. A regular chatbot — even a powerful one — starts fresh each session. A companion tracks your history and brings it up unprompted, creating the feeling of continuity and relationship over time.
Will I know I am talking to AI?
Under new regulations in China, the US, and elsewhere, platforms must disclose the AI nature of companions clearly and upfront. Some do this prominently; others previously buried it in terms of service. Regulators are now enforcing explicit disclosure before users begin.
Sources: Replika – Wikipedia; Character.AI – Wikipedia; Artificial human companion – Wikipedia; Carnegie Endowment for International Peace; MIT Media Lab; US FTC.