The friend who never needs anything from you

An AI assistant and an AI companion are not quite the same thing

General-purpose chatbots are built to help with many tasks, though people may also use them for emotional conversation. AI-companion products are more deliberately designed to simulate friendship, mentoring, romance or continuing social presence.

The boundary is already blurry. A practical assistant can sound warm, remember preferences and respond in a personal voice. A companion can also give advice, organise thoughts or rehearse a difficult conversation.

But neither kind of system experiences the relationship as a person does. Its apparent empathy is generated from patterns and product design, not a private inner life that can be known, hurt, surprised or changed by us.

That does not mean the user’s feelings are fake. People have always formed genuine attachments to characters, objects, places and one-sided media figures. The emotional experience can be real even when the relationship is not mutual.

The important distinction is not real feeling versus fake feeling. It is mutual relationship versus simulated responsiveness.

The evidence for relief should be taken seriously

A 2025 Journal of Consumer Research paper reported several studies in which AI companions reduced momentary loneliness. In one experiment, the reduction was comparable with interacting with another person and greater than passive activities such as watching videos. Across a week-long study, participants repeatedly reported less loneliness immediately after using the companion.

Feeling heard appeared to be especially important. The researchers found that perceived listening helped explain the short-term reduction more than disclosure or distraction alone.

A 2024 systematic review and meta-analysis also found a moderate pooled reduction in loneliness from relational agents. But its 14 studies included only 286 participants in total, individual samples were small, and many agents were older social robots or structured systems rather than today’s open-ended language models. The authors described the evidence as promising while noting bias and design limitations.

These findings matter because loneliness is not cured by telling people that only human contact counts. Someone isolated by disability, geography, caregiving, stigma, night-shift work or social anxiety may need support when a person is not available. A well-designed system may offer immediate emotional steadiness or help someone find words for an experience they have not yet shared.

The longer view is more uncertain

A four-week preregistered controlled study by MIT Media Lab and OpenAI involved 981 participants and more than 300,000 messages. The assigned conversation modes and topics did not produce significant differences in the study’s main psychosocial outcomes. Participants who voluntarily used the chatbot more, however, tended to report worse outcomes, including greater emotional dependence and problematic use.

That pattern cannot tell us whether heavier use caused the difficulty. People already feeling lonely or attached may simply have used the system more. The study also examined ChatGPT rather than a product designed solely as a companion.

A 2026 longitudinal study followed more than 2,000 adults across four Western countries for a year. Increased social-chatbot use predicted later increases in a single-item measure of emotional isolation. With a broader measure of social connection, the direction looked different: feeling less connected predicted more subsequent chatbot use, while chatbot use did not significantly improve connection. The authors explicitly called the analyses exploratory.

Together, these studies suggest a possible feedback loop rather than a simple verdict. Loneliness can lead someone towards an AI companion. The companion may provide immediate relief. If that relief becomes easier than risking human contact, the person may receive less practice and fewer opportunities to build the relationships that could change their isolation more durably.

The same conversation can be a bridge on one night and a hiding place on another.

Human relationships contain friction for a reason

A friend may misunderstand. A partner may need the evening to be about them. A parent may offer advice when we wanted listening. A child may refuse the perfect question. Every person arrives with needs, limits, history and freedom.

This makes human connection inefficient. It also makes it consequential.

When someone stays after we disappoint them, their staying means something because leaving was possible. When they make time, the time has a cost. When we apologise and repair a rupture, both people are changed by having endured something neither could edit away.

An AI can help us think through conflict, but it does not need us to consider its exhaustion, honour its boundary, tolerate its silence or make room for a version of events that threatens our own. Those obligations are not defects in relationship. They are part of how attention becomes care and care becomes character.

If every difficult human response begins to feel like poor service, we may not become better connected. We may become less tolerant of the conditions under which connection becomes real.

Ask what the conversation makes easier afterwards

Minutes spent with a chatbot are not automatically minutes stolen from a person. The more useful question is functional: what is this interaction doing in my life?

• After talking to the AI, do I feel more ready to contact someone—or relieved that I no longer have to?
• Am I using it to rehearse honesty, or to avoid the uncertainty of being known?
• Does it help me understand another person’s perspective, or mostly confirm mine?
• Have I begun keeping important experiences from people because the AI is easier?
• When a human response is slower, messier or less affirming, do I now experience that as failure?
• Would I feel comfortable telling someone I trust how much this companion matters to me?

No single answer diagnoses a problem. Patterns matter: secrecy, escalating time, distress when access is interrupted, withdrawal from existing relationships, or choosing simulation whenever human contact carries a risk of disagreement.

Use the tool as a bridge back to people

The goal does not have to be abstinence. AI can help translate a feeling into a first sentence, prepare for an apology, identify a person who may be safe to approach or lower the activation energy of arranging a meeting.

1. 01
   Name the human destination

   Before opening the chat, decide what real-world action the conversation might support: a message, a call, a boundary or an invitation.

2. 02
   Ask for preparation, not substitution

   Use the AI to help organise what you want to say, then say some version of it to a person who can truly respond.

3. 03
   Keep one thing reciprocal

   After receiving support, ask someone else how they are and stay long enough to hear the answer.

4. 04
   Notice avoidance

   If you repeatedly postpone the human step, ask whether the tool is reducing fear or quietly protecting it.

5. 05
   Protect privacy and safety

   Do not assume an emotional conversation is confidential. Avoid sharing identifying or highly sensitive information, and do not rely on a general chatbot for emergency or professional mental-health care.

Let AI help you find the words. Let another person have the chance to answer them.

What we do not yet know in Singapore

FO-C-US found no representative Singapore evidence measuring how many people use AI for companionship, whether such use replaces face-to-face contact, or how outcomes differ across age, language, family structure and social support.

Research from the United States and other countries should not be treated as a local prevalence estimate. Product design, cultural expectations, access to care and the meaning of family obligation may all shape how people use artificial companionship.

The point is not to win an argument about what counts as real

If an AI conversation helped you through a difficult night, the help does not become meaningless because software produced it. If you have formed an attachment, shame is unlikely to make the attachment safer or easier to understand.

But relief and relationship are not interchangeable. One reduces distress now. The other asks us to build a life in which care moves in more than one direction.

The people we love will sometimes answer badly. So will we. They will need us when we are empty, challenge the story in which we are entirely right, and change in ways our preferences did not request.

That is precisely why showing up matters.

A machine may become exceptionally good at making us feel accompanied. Our human task is to ensure that this comfort does not make us less willing to accompany one another.

The best artificial companion may be the one that helps us close the window and open a door.