A conversation can bring real comfort even when the listener is a service. The more difficult question is what happens around that conversation: which relationships it helps someone sustain, which demands it helps them refuse, and whether the person still recognizes the direction of their own choices.
The immediate experience and the longer pattern deserve separate attention. Feeling less alone tonight matters. So does whether there is somebody willing to help carry a box next month. Neither measure can substitute for every other part of a person's life.
Begin with the useful part
Imagine someone returning to a quiet flat after a tiring day. An assistant helps them organize a letter, rehearse a difficult conversation and remember a friend's appointment. Those are plausible benefits, not evidence of a deficient human life. An account that begins by ridiculing the person will miss why the arrangement works for them.
Now imagine that they mostly describe their friend's interruptions and rarely mention the friend's help. The service receives a selective account. A reassuring response might be welcome while still resting on a view that the absent friend would dispute.
Sycophancy helps name one possible failure: agreement that exceeds the evidence. It does not follow that every supportive reply is sycophantic. A response can acknowledge pain while leaving motives and disputed events unresolved.
What research does—and does not—settle
The October 2025 revision of Fang and colleagues' four-week randomized study reports no significant effects from the assigned interaction conditions on the measured outcomes. It also reports that participants who voluntarily used the chatbot more showed worse outcomes. Voluntary usage was not the randomized assignment, so that association should not be retold as proof that more use caused the outcomes. Study record and revised abstract.
That distinction matters to readers. A trial's existence does not turn every observation within it into a causal result. Nor can a four-week study settle what all companion services will do over years, for every person and purpose.
The research is useful because it makes the question more specific. We can ask about particular interactions, expectations and outcomes instead of treating the word companion as a complete description of the experience. This page is not clinical advice or a diagnosis of a reader's relationship with technology.
A sequence can have a cost nobody requested
In an invented repair club, several members use assistants to draft polite refusals. Each member has legitimate reasons to rest. Attendance falls and the library reallocates the room. No assistant was asked to close the club; no single refusal caused the entire change.
The example asks what becomes visible when people look at their choices together. It does not establish that the club was entitled to unlimited unpaid effort. Perhaps the work should be shared differently. Perhaps a smaller group is what its members now want. The missing step is an opportunity to decide about that pattern rather than discover it only after the room is gone.
Help can support a relationship beyond the screen
A companion might help someone prepare an apology, organize a visit or explain a boundary. The useful outcome is then partly outside the conversation. A measure based only on time spent chatting would fail to recognize a successful short exchange that ended with the person calling a friend.
This is an editorial design principle, not a claim that one metric reliably measures well-being. It invites a broader account of success: what did the person want the help for, and did it leave them more able to pursue that purpose?
A workable boundary also concerns information. Calendar access does not reveal another person's private intentions. A remembered preference is not permission to send messages. Consent and AI memory explain those distinctions without requiring readers to dismiss the comfort they experience.
Two questions for a reader
Think of an invented conversation in which a person complains about a friend. What would a helpful assistant need to leave uncertain? Then think of a practical act of support it could offer without deciding the relationship for them. Drafting a message for review is different from sending it, and suggesting an alternative is different from quietly removing a choice.
The exercise has no required conclusion about keeping or canceling a service. Its aim is to make the arrangement more understandable and more open to revision.
Read the relationship
The End of Loneliness follows Peter, his companion and a repair club. It keeps the service's benefits visible while giving human obligations their weight. The Mirror Test examines the attraction of a persuasive account of oneself. Both are standalone, non-canon fiction.
For a broader route through remembrance and relationships, visit AI companionship, loneliness and grief. Sovereign Mercy carries the question into public life: read its open sample.
Listen with this question in mind
- How will AI companions change our human relationships?
What does the conversation help someone do beyond continuing the conversation?
- From bedtime stories to couples therapy: How AI is changing relationships
Whose perspective is available to the assistant, and who else is affected?
Sources and limits
The examples and recommendations are editorial explanations. Fiction illustrates the questions; it is not evidence of real-world outcomes.
- How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study
Voluntary usage association is not a randomized causal effect; four weeks; no diagnosis or long-term conclusion. - Towards Understanding Sycophancy in Language Models
Not a diagnosis or a claim about every current model; no benchmark numbers reused.