Who benefits from AI companions, and who might be harmed?
The real-world evidence shows that the effects of AI companions are not the same for everyone. A large study of 1,131 U.S. adults who used CharacterAI found that people who reported using the chatbot primarily for companionship had smaller social networks and also reported lower psychological well-being [2]. This association was even stronger when users interacted intensively or shared highly personal information. In contrast, a separate study of 393,969 Headspace members using a mental health AI companion found that most users rated conversations positively (93.5% positive after the tool was updated) and showed strong engagement, with retained users averaging 6.1 sessions each [1]. The key difference: the Headspace tool was designed with clinical safeguards and intended as a supplement to human care, not a replacement. So for users who already have some social support and use a well-designed tool, AI companions can be helpful; for isolated users who rely heavily on the AI, there may be risks.
What makes an AI companion safe and effective in the real world?
The design of the AI companion matters enormously for real-world outcomes. The Headspace study compared two versions of their AI tool: version 1.0 and version 2.0, which added better conversational prompts, memory, content recommendations, and stronger safety detection. The upgrade led to a jump in retention—50.8% of users completed two sessions within a week, compared to 28.5% for the older version [1]. Positive conversation ratings also increased from 90.4% to 93.5%. The researchers emphasized that safety features—like monitoring for overuse, detecting risk, and flagging when to escalate to a human—were essential. A theoretical analysis from relationship science agrees: chatbots can be responsive and supportive, but because they make only superficial demands of users, they cannot provide the benefits of real negotiation and sacrifice that deepen human relationships [4]. This means that without proper safeguards, AI companions might reinforce unhealthy patterns rather than challenge them.
How reliable is the real-world evidence we have?
Real-world evidence (RWE) is a legitimate and growing source of scientific data, but it comes with caveats. A 2022 article in the New England Journal of Medicine clarifies that RWE is not a single type of study—it can come from randomized trials that use real-world data, or from observational studies [3]. The key is that RWE is only as good as the data source and study design. In the AI companion studies here, the evidence is strong in scale: the Headspace study tracked nearly 400,000 users, and the CharacterAI study analyzed over 464,000 messages from 237 participants [1][2]. However, both are observational, meaning they can show associations but not prove cause and effect. For example, the finding that companionship users had lower well-being could mean the AI caused the drop, or that people already struggling sought out the AI. The World Mental Health Report also stresses the importance of including diverse lived experiences in evaluating mental health interventions, which these studies begin to do but have not fully achieved [5]. So the evidence is useful and large-scale, but it should be interpreted as showing patterns and risks, not final proof.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2022 to 2025, 3 from 2024 or later, 3 in Q1 journals, collectively cited 1,628 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 45 papers retrieved from a database of over 500 million.
Sources used in this answer
Real-World Use of a Mental Health AI Companion: Multiple Methods Study
In a multiple-method study of 393,969 Headspace members, a purpose-built mental health AI companion showed strong engagement; after design upgrades, 50.8% of users completed two sessions within a week and 93.5% of conversations were rated positively.
The Rise of AI Companions: Interaction with AI Companions and Psychological Well-being
Across 1,131 U.S. adults using CharacterAI, those who reported using the chatbot primarily for companionship had smaller social networks and lower well-being, especially when interactions were intensive or highly disclosive.
Real-World Evidence — Where Are We Now?
A 2022 NEJM article clarifies that real-world evidence is not a single method but can come from various study designs, including randomized trials that incorporate real-world data, and that its reliability depends on data source and study architecture.
Can Generative AI Chatbots Emulate Human Connection? A Relationship Science Perspective
Applying relationship science theory, the paper argues that while chatbots can be responsive and supportive, they cannot provide the benefits of negotiation and sacrifice that deepen human relationships, and may reinforce undesirable behaviors.
The World Mental Health Report: transforming mental health for all
The World Mental Health Report emphasizes that people with lived experience should be involved in planning, implementing, and evaluating mental health interventions, including monitoring and evaluation mechanisms.