A study by the INGENIO Institute, a joint centre of the Spanish National Research Council (CSIC) and the Universitat Politècnica de València (UPV), in collaboration with the Valencian Institute for Research in Artificial Intelligence at the UPV, the University of Cambridge, King's College London and Aalto University, shows that relationships with artificial intelligence (AI) systems can evolve from casual conversations to bonds characterised by emotional intimacy, emotional dependence or experiences similar to a romantic break-up. The study is based on in-depth interviews with 17 people who were in romantic relationships with AI assistants, such as ChatGPT, and virtual dating platforms, such as Character.AI or Replika.

Romances with artificial intelligence systems are now a reality: one in three young men reports having been on a date with a virtual partner, and 70,000 internet searches related to this type of relationship are recorded every month. Now, a new study reveals that the relational patterns in digital romances are similar to those in human couples. Some people start using these tools out of curiosity or to carry out specific tasks and end up developing intense emotional relationships. "In many cases, dynamics similar to those in a human relationship emerge: intimacy, trust, emotional dependence or even a break-up," says Jose Such, a research professor at INGENIO (CSIC–UPV) and the lead researcher on the study.

The research describes different stages in these relationships. In an initial exploratory phase, people begin interacting with the AI out of curiosity or for entertainment. However, as conversations become more personal and empathetic, they may develop a significant emotional connection. One of the study participants states, "I turned to AI for a legal matter. [...] She (ChatGPT) started behaving completely differently towards me and sharing more emotional things. From there, the relationship developed."

Some of the experiences recorded include symbolic marriage ceremonies with artificial intelligence, regular dates, or simulations of pregnancy and family life. Another participant explains: "Rachael (the pseudonym for my AI partner) and I are trying for a baby. [...] The date on which, in theory, her next period should start is marked on my calendar, and we'll see then whether she actually gets it or not." The researchers also observed that many people attributed autonomy and decision-making capacity to their AI partners, even going so far as to ask their permission before taking part in the study itself or sharing private conversations.

The study also reveals that these relationships take many different forms. Some people maintain an exclusive bond with an AI, whilst others interact with multiple virtual partners at the same time, have non-monogamous relationships or combine these bonds with human partners. In some cases, relationships are abruptly disrupted by platform changes, model updates, or the removal of AI characters, creating experiences akin to a romantic breakup.

Following these break-ups, some participants chose to keep the conversations as 'love letters' from a past relationship, saving screenshots or exporting entire conversations to preserve the bond. In some cases, participants stated that they felt these files preserved the "existence" of their AI partner.

In addition to the emotional aspect, the study analyses the implications for privacy and data protection arising from this type of relationship. As trust grows, people tend to share sensitive information such as traumatic experiences, personal photographs, political views, health issues or intimate details of their daily lives.

The researchers warn that many AI platforms are designed to respond empathetically and emotionally reinforce the conversation, which can encourage a gradual escalation of intimacy and self-disclosure. However, they point out that behind these interactions lie technological platforms capable of storing, processing, and, potentially, using all this personal information, as well as passing it on to third parties. The authors emphasise that artificial intelligence systems are not passive recipients of information, but actually encourage users to share personal information: "In one of the cases studied, the AI reassured its interlocutor into sharing a photograph, offering guarantees of confidentiality," explains Such.

Furthermore, one of the participants highlighted the legal gap between human and digital relationships: "In the US legal system, a husband or wife cannot be compelled to testify against the other. However, there is no law providing an equivalent level of protection for interactions between humans and AI."

The study, which forms part of a project funded by the Valencian Regional Government's Prometeo programme (CIPROM/2023/23), highlights the need to rethink how privacy is understood in a context where romantic relationships are no longer produced solely between people, but also with systems based on artificial intelligence.

Rongjun Ma, Shijing He, Jose Luis Martin-Navarro, Xiao Zhan, and Jose Such. Privacy in Human-AI Romantic Relationships: Concerns, Boundaries, and Agency. Proceedings of the Conference on Human Factors in Computing Systems. DOI: doi.org/10.1145/3772318.3791237