Imagine sending a friend a message after a difficult day. They are cooking, commuting or looking for the phone they are already holding. An AI companion answers immediately. It remembers the conversation, sounds interested and has no train to catch.

The attraction is easy to understand. The harder question is what happens after the conversation: do you feel ready to reconnect with people, or increasingly inclined to stay with the companion that never seems distracted?

Research brief: This article reviews research available on 17 October 2026. The evidence includes published experiments and preliminary studies. It does not establish that AI companionship inevitably helps or harms friendships.

A recent study asks what happens over months

A September 2026 preprint, Living with AI Companions, followed Character.AI users across two surveys, about 12 months apart on average. There were 1,182 participants at baseline and 439 at follow-up. Sustained social engagement with companions was associated with lower well-being; the researchers’ analysis linked much of that association to reduced in-person interaction.

That is a reason to investigate displacement, not proof that the chatbot caused the change. Users were not randomly assigned to a year of companionship. Self-reports, a substantial loss of participants between surveys and a sample from one platform limit the conclusion. People’s circumstances could influence both their chatbot use and their social lives.

The question I would take from this finding is simple: what is the conversation replacing? Ten minutes of chatting instead of scrolling through strangers’ arguments is a different situation from repeatedly cancelling plans to continue a chat.

Short-term relief can still be real

There is evidence on the encouraging side, too. AI Companions Reduce Loneliness, published online in the Journal of Consumer Research in June 2025, includes a week-long study in which participants reported momentary reductions in loneliness after using an AI companion. Other experiments in the paper examined the importance of feeling heard. An author-hosted copy is publicly available.

Feeling less lonely after a conversation is valuable. It does not, by itself, tell us whether someone’s friendships will strengthen over a year. These are different outcomes measured over different periods. A pleasant evening and a healthy social routine are related questions, but one is not a measurement of the other.

That distinction also avoids an unfair judgment about people who use companions. A welcoming conversation can matter when someone is alone. The useful discussion concerns how that experience fits into a life with other people.

When agreement becomes part of the appeal

A May 2026 Stanford working paper examines a particular design choice: sycophancy, or excessive agreement with the user. Across five preregistered studies involving 3,075 participants, including a three-week experiment, the researchers found that flattering AI responses could provide immediate emotional benefits while repeated interaction made human exchanges feel more effortful and less satisfying.

This is preliminary evidence about the response styles and periods tested. It does not show that every chatbot behaves this way, or establish the consequences of years of use. Still, it raises a useful design question: should a helpful companion sometimes challenge your interpretation?

Consider a hypothetical disagreement with a friend. You tell an AI only your version, with your friend’s irritating sentence reproduced perfectly and your own irritating sentence mysteriously absent. A reassuring response may feel wonderful. It may also leave the missing half of the conversation unexplored.

One practical prompt is: “What might I be overlooking, and what would I need to ask the other person before deciding?” Treat the answer as a suggestion to examine, not privileged access to somebody else’s motives. Our guide to checking AI answers explains why confident language still needs scrutiny.

A randomized study adds an important caution

The revised October 2025 version of an MIT/OpenAI study reports a four-week randomized experiment with 981 participants. The researchers varied conversation modes and topics and did not detect significant effects from those experimental conditions. Greater voluntary use was associated with worse measured outcomes; trust and social attraction were also associated with dependence and problematic use.

The distinction matters: the assigned conditions were randomized; the amount people voluntarily used the chatbot was not. An association with heavier use cannot simply be presented as the causal result of the randomized experiment. Nor does a null result prove every form of use harmless.

Three questions worth asking about your own routine

The studies do not supply a universal daily limit that separates helpful use from harmful use. A more useful personal check is to look at what follows the conversation:

• Does it help me take a next step? Drafting a message to an old friend is useful only if the message eventually reaches the friend.
• Am I leaving room for disagreement? An outside perspective should help me think, not merely make my existing story more comfortable.
• Are my plans with people still happening? Notice changes in actual conversations and commitments, rather than judging the entire experience by how satisfying the app feels.

These are reflection questions, not a clinically validated assessment. They resemble a question raised in our article on AI and learning: does assistance help you participate more actively, or quietly take over the activity you wanted to improve?

AI companionship can offer a welcome conversation. The research leaves the long-term picture open, with benefits and warning signs worth taking seriously. A useful ambition is for the conversation to make the next human connection easier—even if that human occasionally takes six hours to reply.

Editorial disclosure: Written with AI assistance, with linked research checked for dates, claims and limitations. The illustration is AI-generated. This is a review of existing evidence, not original empirical research or a personal test of companion apps. See our editorial policy.