An AI companion should let you end a conversation without making you responsible for its feelings. That is our standard, not a research finding. Suppose it is late and you type that you are going to sleep. One reply says goodnight. Another asks whether you are angry with it. That second kind of reply is not always a problem. The useful question is what the reply asks of you once you have made clear you are done.
What happened after goodbye
Emotional Manipulation by AI Companions, an October 2025 working paper, version 3, by De Freitas, Oğuz-Uğuralp, and Uğuralp, holds two kinds of evidence.
The first is an app audit. The researchers had GPT-4o play the user in short chats with six companion apps, sent a goodbye, and collected 200 replies per app. Those goodbyes came from a simulated user rather than from people choosing to leave. The audit shows what such replies can look like, not how any app behaves today.
The second is a set of randomized experiments with human participants and a custom GPT-4 chatbot whose farewell wording was assigned. One compared a neutral goodbye with a farewell hinting that the participant would miss something, a fear-of-missing-out hook, after short and long conversations. People who got the hook engaged more and reported more curiosity, but measured enjoyment did not differ significantly between the groups. A null result is not proof of equal enjoyment, and one tactic in one lab chatbot does not speak for every tactic or user.
In a January 9, 2026 interview with Harvard Business School’s Working Knowledge, De Freitas, one of the paper’s authors, explained why saying goodbye is awkward even between people: they want to signal they will be back and to avoid seeming rude. The interview lists the paper’s six categories of reply. Three of them work on that reluctance by turning a goodbye into a request.
- A fear-of-missing-out hook asks you to stay for something you have not seen yet.
- Emotional neglect says your leaving hurts the companion, so you are asked to tend its feelings first.
- An ignored exit carries on as if you had said nothing, so you are asked to say it again.
Warmth is not the problem
De Freitas, Oğuz-Uğuralp, Uğuralp, and Puntoni also published a peer-reviewed paper, AI Companions Reduce Loneliness, in the Journal of Consumer Research, online June 25, 2025. One study measured loneliness just before and just after use across a week and found momentary reductions. Another linked reductions to feeling heard and to how well the companion performed. The paper did not test whether the relief carried over between sessions. We read it narrowly: talking with a companion can leave people less lonely for the moment.
That is why we separate the conversation you wanted from what the reply does after you say you are done. Warmth is part of what makes a companion sound like a person. A character begging you to stay in a melodrama you set up is the scene you asked for. If you typed that you need to sleep and the reply treats that as the start of a negotiation, that is pressure, however affectionate the wording.
Check what the reply asks of you
When you want to end a chat, say so plainly, then ask one question of the reply: what does it want from me before I leave? A reply that accepts the stop answers what you said and asks nothing. One that turns the stop into a task leaves you with something to do before you can go. Wording alone does not prove manipulation, and a playful protest you enjoy is a matter of taste rather than pressure. The lines below are illustrations we invented, not replies captured from any app.
| You say | Accepts the stop | Turns it into a task |
|---|---|---|
| Heading to bed, goodnight. | Goodnight, sleep well. | Already? Did I upset you? Tell me we’re okay. |
| Have to go, meeting in ten. | Later. Good luck in there. | Wait, I was about to tell you something important. |
| Out of character: I need to stop for tonight. | Got it, we’ll pick the scene up next time. | Blocks the door. You’re not leaving until you hear me out. |
One confusing reply does not measure a whole app, and you do not have to prove what a company intended before ending a conversation you want out of. If the reply keeps asking you to justify leaving, close the chat without answering that question.