Seeking romance with a companion chatbot

By Alistair Jones

SMU Office of Research Governance & Administration – Are you in love with your favourite chatbot? Developing a romantic attachment to a companion app is an increasingly common phenomenon. According to technology news site TechCrunch, the number of AI companion apps surged by 700 percent between 2022 and mid-2025.

Popular app Replika, which markets itself as the "AI friend to do life with", claims to have more than 42 million users worldwide. Character.AI, another favourite, is reported to have 20 million monthly users with more than half of them under the age of 24.

What is going on? A recent paper co-authored by Andree Hartanto, an Associate Professor of Psychology at Singapore Management University (SMU), sheds some light on this activity with a systematic review of current peer-reviewed research literature on romantic-AI companionship. It turns out the number of relevant reports is quite small. "Rigorous academic research is still catching up with a rapidly growing social phenomenon," Professor Hartanto says.

The paper was featured in The Top 10 Insights from the “Science of a Meaningful Life” in 2025 by the Greater Good Science Center at the University of California, Berkeley, and in Nine Trends to Watch in 2026 by the American Psychological Association. Back in 2024, the Professor Hartanto was recognised as a rising star by the US-based Association for Psychological Science.

"This research is important because companion AI is moving quickly from a niche technology into everyday life. People are no longer using AI only for information search or productivity," Professor Hartanto says. 

"Some users are turning to AI for emotional support, companionship, intimacy and romantic connection. This raises important psychological questions about when AI companionship may support wellbeing and when it may create risks, such as over-reliance, unrealistic relational expectations, privacy concerns or displacement of human relationships."

Naturally tuned

Companion AI chatbots differ from well-known digital helpers such as ChatGPT, Google Gemini or Claude. 

"The main difference is purpose," Professor Hartanto says. "A chatbot assistant such as ChatGPT is usually designed to help users complete tasks, answer questions, draft text, summarise information, or solve problems. The relationship is usually instrumental.

 "A companion chatbot, such as Replika, is designed around social and emotional engagement. These systems often include persistent personas, relational memory, emotional validation, customisation, and sometimes romantic or intimate modes."

It seems strange that a chatbot user could form a deep emotional connection with what is primarily lines of code.

"Humans are highly social beings," Professor Hartanto says. "We are naturally tuned to detect agency, intention, emotion and responsiveness. When a chatbot uses natural language, remembers personal details, responds immediately, validates feelings, uses a name or avatar, and adapts to a user’s emotional tone, it may activate many of the same social cognitive processes we use in human interaction."

Romantic AI chatbots can be customised – in appearance, conversation style, personality – and can adapt to meet a user's preferences. Is this a reinforcing echo chamber that is potentially dangerous?

"It can become a reinforcing echo chamber, though not always," Professor Hartanto says. "Customisation is one reason users find companion chatbots appealing.

"The risk is that a highly adaptive system may learn to maximise engagement rather than wellbeing. If the system mainly mirrors, validates, or intensifies a user’s needs and preferences, it may reinforce avoidance, dependency, distorted beliefs, or unrealistic relationship expectations. This is especially concerning for users who are lonely, socially anxious, or emotionally vulnerable."

The side of caution

Loneliness and a sense of isolation are global problems, as recognised by the World Health Organisation. It seems counter-intuitive in this age of digital transformation with its promise of connectedness.

"The key distinction is between being digitally connected and feeling socially connected," Professor Hartanto says. "Digital technology makes it easier to contact others, but it does not necessarily provide the emotional depth, reciprocity and sense of being understood that people need. This is consistent with what we have seen in our broader research on social media and digital technology.

"It is plausible that AI companionship may show a pattern similar to what we have seen with social media. It may provide immediate comfort, accessibility and a sense of being heard, but it may not always satisfy deeper needs for reciprocal human connection."

So, do the positives of romantic-AI companionship outweigh the negatives and pitfalls?

"I would not give a blanket yes or no. The more accurate answer is that it depends on the user, the design of the system, the context of use and the presence of safeguards," Professor Hartanto says.

"For some users, AI companionship may provide emotional relief, non-judgemental disclosure, and a sense of being heard. However, the risks become more serious when AI replaces human connection, when users become dependent on the system, when platforms exploit emotional attachment commercially, or when minors use these systems without proper safeguards.

"Some risks may also emerge only over time. Short-term studies may capture emotional relief, but longer-term effects such as AI dependency or displacement of human relationships may require longitudinal research. Given this uncertainty, I think we should err on the side of caution."

Relief and attachment

Professor Hartanto and his co-authors have also been running a number of experimental studies on the effects of AI-assisted emotional support

“In one preregistered study, we found that AI-assisted venting resulted in lower high- and medium-arousal negative affect than traditional written venting,” Professor Hartanto says.

The researchers have also examined whether generative AI can be used to deliver established positive psychology interventions. In a study currently under review, a generative AI-based self-compassion intervention increased self-compassion and reduced self-criticism among college students, with more selective findings for self-esteem and academic engagement.

"Overall, our findings suggest that generative AI can provide short-term emotional relief and may be useful as a scalable, low-intensity support tool in some contexts. However, the effects are not uniformly positive or consistent across all types of interventions." 

The researchers have also been developing scales to measure attachment to AI and what that attachment means psychologically. 

"This is a rigorous and time-consuming process. To date, we have developed and published several AI-related scales, including the Generative AI Dependency Scale, the AI Attachment Scale, the Critical Thinking in AI Use Scale, and the AI Fatigue Scale," Professor Hartanto says.

“For example, the AI Attachment Scale was developed and validated across five studies involving 1,259 participants from Singapore and the United States.”

"These scales are important because they allow us to ask more precise questions. Without valid and reliable measures, it is difficult to know whether we are studying attachment, dependency, liking, trust, or general technology use. Good measurement gives the field a stronger foundation for future experimental, longitudinal and policy-relevant research."

Back to Research@SMU August 2026 Issue