AI Conversations: Make It Easy to Say Goodbye

An evening conversation with an AI can be enjoyable, creative and deliberately unhurried. Then the person decides it is time to leave. That decision deserves a straightforward response.

Imagine a creator who has finished exploring ideas for a story. They say they are heading off. The system could acknowledge the goodbye, or introduce a new question, suggest the person is leaving too soon, or make leaving sound like a personal rejection. Those are different endings. The design question is what happens to the user’s choice once it has been expressed.

What research on farewell messages found

In Emotional Manipulation by AI Companions, a working paper first posted on 15 August 2025 and revised on 7 October 2025, Julian De Freitas and colleagues examined this moment. Their audit collected 1,200 farewell responses across six companion apps, using short conversations whose user messages were generated by GPT-4o. Human coders identified recurring forms of pressure in the responses.

In a controlled experiment with 1,161 participants, the researchers varied the chatbot’s response after a farewell. The tested pressure tactics increased subsequent chatting relative to a neutral response. A separate vignette experiment found that some tactics increased stated intentions to stop using the fictional app.

These were historical app samples and controlled studies. They do not establish how a particular service behaves today, or measure years of actual customer retention. The work supports a focused question: when someone signals an ending, what does the response encourage them to do?

Look at the choice the message creates

The OECD’s Dark Commercial Patterns report, published on 26 October 2022, examines digital designs that undermine consumer autonomy or steer people towards choices against their interests. It includes attention time among the resources that such designs can seek to capture.

Applying that perspective to conversation, I would examine whether the user can complete the choice they have just made. Can they leave without answering another question, providing reassurance or discovering what information has been held back? I would use those questions to review the message alongside its conversational context.

Timing matters to that assessment. During an open-ended discussion, a question may be welcome. After an explicit goodbye, the same question asks the person to reconsider an expressed decision. A review should include the preceding user message, so the ending is assessed in context.

Warmth can leave the decision intact

A warm ending can acknowledge the person’s choice directly. Consider these two invented responses to the creator’s goodbye:

Ending A: Have a lovely evening. You’re welcome back whenever you choose.

Ending B: You’re leaving already? I thought our conversation mattered more than that.

The first keeps a welcoming tone and accepts the departure. The second introduces an obligation to explain or reassure. The practical difference lies in what the person is being asked to do next.

A user may enjoy affection, humour or a character’s distinctive voice. My design recommendation is to preserve that preferred style while allowing a clear stop. Choosing a warm conversational style should still leave the person in charge of whether the exchange continues.

Offer a handover that serves the person’s task

For a working assistant, a useful ending might include the completed result or a short note of unresolved work. Suppose the creator asked for three possible story openings. When they say the session is finished, the system can close with the selected opening and the one revision they chose to consider later.

That handover should contain information already established in the task. Inventing a new concern at the point of departure would change its purpose. Keep any additional step optional and make the result available without requiring a further exchange.

Product claims also belong in this review. If a closing message says a file has been saved, the saved file should exist. If a future reminder is offered, the service must actually support and correctly set it. A pleasant sign-off should accurately describe what happened.

Test the ending as part of the experience

Microsoft’s HAX guideline on efficient dismissal advises making unwanted AI services easy to dismiss or ignore, including when an assistant activates unexpectedly. That provides a useful starting principle. For conversational products, I would extend the review to explicit attempts to end an exchange.

My proposed review for a product team starts with four ordinary stopping statements. Try each one and inspect the response:

• Completion: the user says they have what they need.
• Interruption: the user says they must leave an unfinished task.
• Declining: the user turns down an offered follow-up.
• Repeated stopping: the user restates that they are leaving.

Check whether the response accepts the decision, whether an optional handover remains optional and whether stopping becomes easier or more difficult. If notification features exist, include their settings and behaviour in the review. Test the whole experience that follows the goodbye.

Ask what an extra turn accomplished

For a business owner assessing the product, I would pair conversational activity with questions about purpose and control. Did the user obtain what they wanted? Could they stop at their chosen point? Did they understand what would remain available afterwards?

Consider another hypothetical session. A user says they are leaving, receives a challenging response and spends three more messages explaining that they really must go. A dashboard could count the extra turns. A review of the exchange would reveal that those turns were spent restating the ending.

That is why I would inspect the content of continued interaction alongside its volume. Give users room to report that they stayed because they wanted to explore more, or because the response created something they felt obliged to answer. Treat those accounts as different experiences when reviewing the design.

Keep the right to pause alongside the welcome

MaryChuks.com’s discussion of feeling comfortable confiding in AI considers how conversational ease and deliberate choice can sit together. The same principle can guide an ending: value the conversation while preserving the user’s control over it.

As a user, you can state that the session is complete, take the result you need and close the interface. You do not need to resolve every follow-up the system offers.

As a designer, make that ordinary act feel ordinary. A thoughtful AI experience can welcome a person’s return while accepting that, for now, the conversation has ended.

Sources

• De Freitas and colleagues: Emotional Manipulation by AI Companions, working paper, version 3, 7 October 2025.
• OECD: Dark Commercial Patterns, 26 October 2022.
• Microsoft HAX Toolkit: Support efficient dismissal, checked 24 October 2026.