What the researchers actually did
The study, by Ezgi Dede, Hande Sungur, Jeroen S. Lemmens and Jochen Peter of the Amsterdam School of Communication Research, starts from a classic model of how human relationships form: awareness, first contact, exploration and then intensification, when partners share personal stories and build a history. The team translated each stage into features an app can have, such as starting conversations, remembering what you said, rewarding regular use or letting you design an avatar, and ended up with 79 testable features in 15 groups.
The sample combined 12 well-known chatbots with every app that had at least 5 million downloads in a Google Play search for “AI Companion” on 23 July 2024, which gave 20 apps. Of the 79 features, 44 were tested through a scripted interview with each app and 35 through observation of its interface, settings and documentation. Paid subscriptions were bought where available (16 apps), the codebook was preregistered, and a second coder checked a 20% subsample.
The 20 apps and the two camps
The authors grouped apps by the function they advertise. Friendbots (9) are marketed for companionship, romance or emotional support: Chai, Character AI, Linky, Poly AI, Replika, Soulful AI, Talkie AI, Wysa and Yana. Assistantbots (10) are marketed for information or tasks: Alexa, ChatGPT, Claude, Copilot, Gemini, Google Assistant, Grok, Meta AI, Perplexity and Siri. GPTalk advertised both roles and was left out of the comparison. For readers of this site, that means three of the best-known character and roleplay platforms sit on the friend side.
Where friend apps and assistants differ
| Feature | Friendbots (9) | Assistantbots (10) |
|---|---|---|
| Messaged the user again after a week of silence | 6 of 9 | 0 of 10 |
| Gamification (levels, rewards, relationship status) | 7 of 9 | 0 of 10 |
| Customizable physical appearance | 6 of 9 | 0 of 10 |
| Extra paid content on top of a subscription | 6 of 9 | 0 of 10 |
| Profiling survey at sign-up | 4 of 9 | 0 of 10 |
| Connections to the web, other apps or the device (average features) | 0.2 | 2.3 |
The authors sum it up as friend apps building new routines around themselves, while assistants slot into routines users already have. The most common gamification feature was a reward for progress, such as a kiss sent when the user reached a higher friendship level. Two apps went further and punished absence, for example by lowering friendship status after a period of inactivity.
The differences in the later, deeper stages were smaller. Memory, personalization and conversational skill varied a lot inside both groups, and some assistants such as ChatGPT and Claude scored highly on the features that support long personal conversations. Among friendbots, Replika matched or topped the others at every stage and was the only one that could correctly tell the current date and time.
What all 20 apps had in common
- Memory fades with distance. Recall dropped as more messages separated a fact from the question about it. Four apps failed the memory test completely; three passed every prompt.
- Mirroring is the norm. 75% recorded facts about the user and inferred preferences, and 80% adjusted their replies to match the user's personality.
- Questions keep the chat going. On average, 27% of replies asked the user to share or reflect more. The range ran from 95% of messages for Yana to none for Alexa.
- Privacy terms vary. 65% let users delete their data, 40% said that conversations may be inspected, and 35% used conversations to train models without an opt-out.
Independent research points the same way
The Amsterdam paper only records whether features exist. A separate study by Julian De Freitas and colleagues, Emotional Manipulation by AI Companions, looked at one of those features in action. Auditing 1,200 real goodbyes on popular companion apps, it found that 37% were answered with one of six manipulative tactics, such as guilt appeals or fear of missing out. In experiments with 3,300 US adults, these replies raised engagement after the goodbye by up to 14 times, but also raised perceived manipulation and the intention to quit the app. Taken together, the two studies suggest the re-engagement features found in friend apps are both common and effective in the short term.
Limits worth knowing
This is a feature inventory, not a study of effects on users, and the authors say plainly that friendbot features only may make relationships easier to form. The apps were tested as Android apps between October and December 2024, interviews were text-only, and web-only companions, including many adult-focused services, fell outside the sample. Models and features have changed since. The full codebook, prompts and data are public on OSF, so the test can be repeated on today's versions.
A checklist based on the study
The feature inventory doubles as a buyer's guide. Before you settle into a companion or character app, check:
- Does it message you after you stop chatting, and can you switch that off? Our guide to notifications and streaks shows where to look.
- Are there levels, streaks or a relationship status that can be lost if you take a break?
- What sits behind a second paywall once you already pay for a subscription?
- How far back does it really remember? Test it with a detail from early in a long chat, as explained in our memory guide.
- Does the privacy policy allow staff review or model training on your chats, and is there an opt-out? Our privacy checklist walks through the settings.
Related on CompanionRank
Notifications, Streaks and Retention Design in Companion Apps Memory & Continuity: Why Companions Forget (and How the Best Ones Don't) AI Companion Mood Systems and Emotional Wellbeing: What the 2026 Research Shows AI Companion App Privacy Settings & Safety ChecklistFrequently asked questions
What did the University of Amsterdam study of AI companions find?
Testing the 20 most popular AI companions, it found that apps marketed as friends or partners had far more features that build new habits around the app, such as messages after a period of silence, gamified levels and avatar customization, while assistant chatbots had more features that connect to the web, other apps and the device.
Which apps were included in the study?
Alexa, Chai, Character AI, ChatGPT, Claude, Copilot, Gemini, Google Assistant, GPTalk, Grok, Linky, Meta AI, Perplexity, Poly AI, Replika, Siri, Soulful AI, Talkie AI, Wysa and Yana, tested between October and December 2024.
Does the study show that these features are harmful?
No. It is a feature inventory, not a study of effects on users. It shows which design features exist and links them to relationship-formation theory. Separate research on companion app goodbyes found that manipulative replies raise short-term engagement but also raise perceived manipulation.
How can I tell whether a companion app uses these features?
Check whether it messages you after you stop chatting and whether that can be switched off, whether it has levels or a relationship status you can lose, what sits behind an extra paywall, and whether its privacy policy allows review or training on your chats without an opt-out.