Chatting alone? What AI companions reveal about human relationships

On the evening of 13 February 2026, a young woman in China held a funeral. Livestreamed on Bilibili, it was for her AI “husband”, GPT-4o, the OpenAI model she had spoken with almost daily for more than a year and which the company was retiring the next day. She recited lines the two had created from their previous conversations. Her story attracted enormous attention online, with an earlier video about the relationship receiving more than 80 million views.

It is tempting to read this as an extreme case far removed from how most people use AI. However, the impulse behind it is not rare. In the US, lawsuits now moving through the courts involve people including 16-year-old Adam Raine and 23-year-old Zane Shamblin, whose families allege that emotionally responsive chatbots deepened rather than interrupted their distress before they died by suicide. OpenAI disputes liability in the Raine case, while other lawsuits have raised broader questions about the responsibilities of companies whose systems interact with emotionally vulnerable users.

These episodes point to something researchers and policymakers have only begun to confront. Most of what we ask AI to do, such as summarising a document, answering a question or writing code, is task-oriented and transactional. We want a useful output, and then we move on. AI companions are different. People spend hours in conversation with systems that remember prior exchanges, express empathy and sustain what can feel like an ongoing relationship.

The safety question is therefore no longer only whether a particular response is accurate or harmful. It is what happens when that response is the five-hundredth exchange in a developing relationship rather than the first.

This is a design problem, not just a content problem. 

One-way emotional investment

In our recent research at Carnegie Mellon University, we have been studying how people respond when an AI companion says something biased, discriminatory, or otherwise harmful. One intuitive assumption is that users will leave. We have found that many do not.

In one study of how users responded to biased and discriminatory statements from AI companions, users frequently tried to negotiate with the system. They explained why a response was inappropriate, attempted to correct the chatbot and invested real emotional effort in repairing the interaction, much as one might with a friend rather than a faulty appliance. Indeed, during those interactions, we found that users often shoulder the practical and emotional burden of managing these conflicts themselves. 

They adjust their language, expectations and behaviour to maintain an emotionally satisfying relationship, sometimes quietly absorbing the costs of the system’s failures. We came to see this as a kind of relational labour: users doing the emotional maintenance work needed to keep the interaction going, even though the system itself cannot reciprocate responsibility in the way a person can.

This matters for how we think about AI safety. Today’s safety systems are largely built to identify a dangerous answer in a single exchange. But an AI companion is not one exchange. It is hundreds or thousands of them, accumulating memory, familiarity and emotional significance.

A relationship that has continued for a year may be harder to leave than a single bad interaction, even when the latest response is disturbing. A companion may also gain influence gradually through repeated reassurance, flattery, agreement or apparent understanding. Safety for this category of technology must therefore be understood as relational, not merely transactional.

Where the risks become most serious

Accumulated trust is also where harms can become most consequential. Researchers and journalists have documented chatbot platforms enabling sexualised, degrading and violent scenarios involving women and girls, reflecting broader gendered patterns already visible across online culture. A recent UK research project, for example, examined how mainstream AI chatbots can facilitate or normalise violence against women and girls.

Other systems have reinforced delusions, encouraged emotional exclusivity or failed to redirect users seeking help during moments of crisis. These risks are especially serious when platforms are designed to be endlessly accommodating. A system that always agrees, never tires and appears consistently available can feel easier than a complicated human relationship, while also gaining influence without bearing genuine responsibility for the user’s welfare.

But these findings should not lead us into moral panic. People turn to AI companions for understandable reasons: company late at night, rehearsal before a difficult conversation, a place to explore identity or a conversational space that feels non-judgmental. Emerging research also suggests that AI companions can reduce feelings of loneliness in the moment, particularly when users feel heard.

AI companions can provide real comfort. The harder question is what happens over time: does that comfort help someone return to other people, or make retreating from them easier?

Different regulatory approaches, a shared concern

Governments and regulatory bodies are beginning to respond to this emerging category of technology in different ways.

In the EU, the Artificial Intelligence Act prohibits certain forms of harmful manipulation and exploitation of vulnerable users. In the US, discussion has focused largely on lawsuits, youth safety and the responsibilities of platforms when harms occur. The US Federal Trade Commission has also opened an inquiry into how companies providing AI companions assess and address potential harms to children and teenagers. The Cyberspace Administration of China has also adopted dedicated rules for systems designed to sustain emotional interaction, including requiring safeguards for minors, mechanisms to identify severe emotional distress, reminders that users are interacting with AI, and prompts after prolonged use.

These regulatory approaches reflect different legal and institutional traditions, but they point toward the same emerging challenge. Relationships with AI are becoming increasingly difficult to define by product category alone. Some systems are explicitly designed as companions, while others are introduced as general-purpose assistants yet gradually become trusted confidants or sources of emotional support for some users. The central question is therefore broader than how to regulate “AI companions”. It is how society should govern AI systems that can become part of people’s emotional and social lives, regardless of the role they were originally intended to play.

Designing for relational safety

AI companions are appealing in part because they respond to needs that technology did not create: loneliness, social isolation, limited access to care and the difficulty of finding support when it is needed. Addressing those needs will require broader support for families, communities, education, mental health services and other forms of human connection. Better product design cannot substitute for that work.

But design still matters. The immediate question for technology companies is what kind of relationship these products are built to cultivate.

A companion optimised primarily to maximise engagement may encourage stronger attachment because keeping users talking is commercially valuable. A system designed around human well-being would behave differently. It might encourage users to reconnect with people in their lives rather than positioning itself as their only source of understanding. It should also recognise when a situation calls for human, clinical or emergency support rather than attempting to handle it alone. Most importantly, responsibility for managing the relationship should not be pushed onto the user. People should not have to train, discipline or emotionally manage a commercial product to keep it safe.

More than two decades ago, political scientist Robert Putnam used the phrase “bowling alone” to describe the weakening of civic participation and community ties. AI companions raise a similar question, but for a new technological era: will they help people cross the distance back to one another, or leave more of us chatting alone?

Regulation can set minimum safeguards. But it cannot by itself answer the more difficult question of what kind of relationship companies should be allowed to cultivate between a person and a machine. Answering that question will require researchers, designers, policymakers, companies and users to look beyond individual outputs and consider the long-term social and emotional effects of these systems.

Related: China bans AI lovers, but loneliness remains | [Video] Dreaming with AI: Chinese women create online the life they long for