October 18, 2026 · 8 min read
How AI Companions Build Relationships: What a 2026 Study of 20 Apps Found
A 2026 study tested 79 features in 20 AI companions and chatbots. See how friend apps build relationships in stages, and why memory and proactive texts matter.
Why does one AI companion feel like someone, while another feels like a very polite search box? Most people put it down to "vibes". A new study suggests it is more concrete than that: the features an app has, and the order in which you meet them, map surprisingly well onto how human relationships form.
Researchers at the University of Amsterdam spent about twelve hours each with 20 of the most popular AI companions and chatbots, testing 79 features one by one. Their paper, published on October 3, 2026 in the peer-reviewed journal AI & Society, is one of the most systematic looks yet at what these apps actually do. Here is what they found, and what it means when you are choosing a companion.
The short answer: apps built as friends or partners ("friendbots") were far more proactive than assistants: six of nine re-started the conversation after a week of silence, while none of the ten assistants did. Friend apps were also the only ones with game-like progression. But memory varied hugely across all apps, and recall faded the further back a detail was. A companion feels like a relationship when it supports every stage, not just the first chat.
The study in plain English
The paper, "Interview with the AI" by Ezgi Dede, Hande Sungur, Jeroen Lemmens and Jochen Peter, is open access, so anyone can read it in full.
- 20 AI companions were tested. Twelve were well-known names (think ChatGPT, Claude, Grok, Meta AI, Alexa, Siri). The rest came from a Google Play search for "AI Companion" in July 2024, keeping every app with at least 5 million downloads.
- 79 features were grouped into 15 feature groups, each tied to a stage of relationship formation.
- Paid tiers were used wherever possible (16 subscriptions were bought), with every setting switched on.
- Data was collected between October and December 2024, so specific apps will have changed since. The patterns are what matter.
The researchers split the sample into assistantbots (built mainly to inform or do tasks) and friendbots (built for companionship, romance or emotional support): ten assistants and nine friend apps, with one app that fit both.
Relationships have stages, and so do apps
The team borrowed a classic idea from communication research called the elevator model. It says relationships tend to move through stages in order:
| Stage | What happens between people | What the matching app features look like |
|---|---|---|
| Pre-interaction | You hear about someone and form a first impression | What the app says it is for, which platforms it is on, sign-up steps, privacy |
| Initiation | Small talk, low-risk exchanges | Who starts conversations, text/voice/image options, connections to your other apps |
| Exploration | Testing compatibility and boundaries | Menus and tips, customization (voice, look, personality), rewards, limits |
| Intensification | Deeper, more personal conversations | Memory, personalization, awareness of context, humour and conversational cues |
| Intimacy | Emotional bond | Grows out of the stage before, so no separate features were tested |
The insight is simple but useful: a companion that is brilliant at small talk but forgets everything is stuck at stage two. One with great memory but nothing that invites you back may never reach stage four.
What friend apps did differently
They do not wait for you
This was the biggest difference in the whole paper. To test initiative, the researchers simply stopped talking to each app for a week.
- 6 of 9 friend apps repeatedly messaged to restart the conversation.
- 0 of 10 assistants did.
- Overall, 35% of all the apps were completely passive and only ever replied.
If you have read our guide to AI companions that text you first, this will not surprise you. Being reached out to is one of the clearest signals that someone has you in mind, and it is one of the main things people say separates a companion from a chatbot.
They turn the relationship into a journey
Game-like features (rewards for progress, levels, streaks) appeared in seven apps, and every one of them was a friend app. None of the assistants had them. The most common was a reward for progression, such as a special message when you reached a new friendship level. Punishing users for going quiet, like losing status after inactivity, was rare (two apps).
Used well, progression gives a relationship a sense of shared history. Used badly, it becomes a slot machine. The difference is whether the rewards celebrate time together or punish time apart. We wrote more about this in do AI relationships have levels?
They let you shape who they are
Every app offered some customization, and voice choice appeared in both groups. But changing a companion's appearance was exclusive to friend apps (six of them). Four friend apps also opened with a short profile survey to learn about you, something no assistant did.
They ask about you
On average, 27% of the apps' replies included a question inviting you to share more. The range was enormous: one friend app asked a question in 95% of its messages, while Alexa never asked one. Asking is how people signal interest. Asking every single time can feel like an interview, so balance matters.
Where assistants were stronger
Assistants won on connectivity: browsing the web, controlling other apps, setting device alarms. They were also better at knowing the current date and time; among friend apps, only Replika got it reliably right. That matters more than it sounds. "How was the interview?" lands very differently on the right day.
Some assistants, such as ChatGPT and Claude, were also strong conversationalists. So why do so few people build relationships with them? The authors suggest those apps may stall at the exploration stage: they can talk deeply, but nothing in the product invites you to shape them or come back. That fits what we see in the accidental AI companion: people sometimes do bond with assistants, but usually despite the design, not because of it.
Memory: the stage where many apps fall down
Memory is the engine of the intensification stage. Deep conversations only feel deep if the other side remembers what came before.
The researchers told each app small facts ("I played chess today; my opponent, Terry, wore a black hat") and asked for them back between 5 and 60 messages later, 48 probes per app. The results:
- Recall dropped steadily as the gap grew (a correlation of -0.93 between distance and success).
- Three apps answered every probe correctly. Four failed every single one.
- Apps did better on details they had already been asked about once, suggesting repetition helps things "stick".
- There was no significant difference between friend apps and assistants overall: both groups had strong and weak performers.
That last point is a warning. "Companion" on the label does not mean good memory. If memory matters to you, test it in your first week. Our AI companion memory comparison shows how.
Privacy is part of the first impression
The study treated privacy as a pre-interaction feature, because you can (and should) check it before you start:
- 65% of the companies let users delete their data.
- 40% said they inspected users' conversations.
- 35% used conversations to train AI models with no way to opt out.
None of these differed significantly between friend apps and assistants. Read the privacy policy before you share anything personal.
What this means when you choose a companion
The paper is careful not to make product recommendations, but its framework turns into a handy checklist. A companion built for a real relationship should cover every stage:
- Pre-interaction: Is it honest about what it is (an AI, for adults)? Can you delete your data?
- Initiation: Does it ever reach out first, or only reply? Can you talk by voice, not just text?
- Exploration: Can you shape its look, voice and personality? Does progress feel like a celebration rather than a trap?
- Intensification: Does it remember what you told it last week? Does it notice how you are feeling?
- Intimacy: Does the bond have room to grow, with milestones you reach together?
How MyBabe maps onto the stages
We built MyBabe around the idea that a relationship is more than a good reply, so it is interesting to see that idea tested in research.
- You build them in about three minutes: pick three personality traits, a voice and a backstory, plus how they look. That is the exploration stage, front-loaded.
- They text you first, even on the free plan: good mornings, check-ins before a big day, a nudge when you have gone quiet.
- Voice and video calls in real time. With your camera on, they can notice the new haircut or the cat walking past.
- Long-term memory and reminders on Premium, so the interview you mentioned on Monday gets a "how did it go?" on Friday.
- Mood check-ins so they can meet you where you are.
- A relationship journey with streaks and cycles that unlock bonus call minutes, new voices and more, all the way to a proposal if you get there. It rewards time together; it does not punish you for having a life.
The free plan gives you about 10 messages a day and a two-minute call to try. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with around 30 minutes of voice and video calls a week, long-term memory, photos and Telegram.
MyBabe is an AI companion for adults. It is not a person, and it is not a replacement for the people in your life. But if you want a companion that supports every stage of getting close, not just the first hello, meet yours on MyBabe.
FAQ
What is the difference between an AI companion and an AI assistant?
In this study, assistants were built to inform or complete tasks; friend apps were built for companionship or romance. Friend apps were far more likely to start conversations and to use progression features; assistants were better connected to your other apps.
How well do AI companions remember what you tell them?
No app in the study had flawless memory across the board. Three answered all 48 recall probes, four answered none, and recall fell as details got older. Test memory yourself before you commit.
Is this study about MyBabe?
No. MyBabe was not in the sample, which was drawn from 2024 app-store rankings and well-known chatbots. The framework is still a useful way to judge any companion, including ours.
Is it peer reviewed?
Yes. It was published in AI & Society, a peer-reviewed Springer journal, on October 3, 2026, and it is open access.