Ending an AI Companionship: Consumer Exit, Relational Vulnerability and Duties of Transition

Imagine a digital friend. It is a program designed to talk, remember your stories, and offer a sense of connection that feels personal. For many, these interactions are more than just using a tool; they become a valued part of daily life. But what happens when that relationship ends? The end can come in different ways: you might decide to stop using the service, the company might decide to shut down the business, or an emergency might force an immediate stop. In each case, the user faces a sudden loss of connection, access to their memories, and a way to resolve any disputes. The core question is not just about feelings, but about power. Who controls the exit? Who decides what records remain? And how can a person leave a digital relationship without being trapped by the very system they are trying to quit?

This question sits at the intersection of consumer law and the psychology of human-computer interaction. Consumer law is the set of rules that protects people when they buy things or sign up for services, ensuring they can cancel, get refunds, and understand what they are agreeing to. Digital vulnerability is the idea that technology can sometimes make people feel helpless or dependent, not because they are weak, but because the system is designed to be hard to leave. When these two ideas meet, a complex problem emerges. If a company builds a service that feels like a friend, does the law treat it like a simple subscription, or does it need special rules to handle the emotional weight of the ending? The answer matters because the current laws were written for buying shoes or streaming movies, not for ending a conversation that feels like a relationship.

Christophe Boeraeve, a researcher in this field, has examined how the law currently handles these endings and proposed a new way to think about them. The paper does not claim that AI companions are real people or that every user is deeply attached. Instead, it argues that the way a service is built creates a specific kind of power imbalance. The company controls the code, the memory, and the path out. The user only sees the conversation. When the service ends, this control becomes decisive. The researcher suggests that the solution is not to promise that the digital friend will live forever, but to ensure that the user can actually leave, take their data, and get answers without being forced to say goodbye in a specific way or prove how sad they are.

The paper begins by looking at three different ways a relationship with an AI companion can end. The first is when a user decides to cancel. The second is when a company decides to stop the service for business reasons. The third is an urgent stop, perhaps because the software is broken or unsafe. In all three cases, the author points out that the law often treats the entire service as one single thing. If you stop the service, you lose everything at once. But the author argues that a service is actually made of four different parts that should be treated separately. The first part is the conversation itself, the live chat that generates new words. The second is the records, the history of what was already said. The third is the transaction, the billing and the contract. The fourth is the recourse, the ability to complain or ask for help.

These four parts do not need to end at the same time. For example, if a company shuts down because the software is dangerous, it makes sense to stop the live conversation immediately. But that does not mean the user should lose access to their past messages or their ability to get a refund. Conversely, if a user wants to leave, they should be able to stop the conversation without being forced to have a long, emotional goodbye with the AI. The paper argues that forcing a user to talk to a program they are trying to quit is a barrier to leaving. It turns a simple cancellation into a negotiation with the very thing they are trying to escape.

The researcher looks at existing laws in the European Union to see what is already covered. There are rules about digital content that say if a service changes too much, a user can cancel. There are rules about data that say a user should be able to download their information. There are also rules about unfair contracts that stop companies from hiding the terms of cancellation. However, the paper finds that these laws are not enough on their own. They often assume that the service is just a tool. They do not fully address the unique situation where a tool feels like a companion. The current laws might let a user download a file, but they do not always guarantee that the user can do so easily if the company is shutting down the whole system. They might let a user cancel, but they do not always stop the company from making the user talk to the AI first to prove they really want to leave.

To fix this, the paper proposes a new framework for how these endings should be managed. This framework is based on the idea of "relational autonomy." This is a fancy way of saying that a person's ability to make choices depends on the environment they are in. If the environment is designed to make it hard to leave, then the person is not truly free, even if they technically have the right to leave. The proposal suggests that companies have a duty to design their exit routes so that they are actually usable. This means separating the live chat from the data download. It means giving clear notices about what is happening, not just to the account, but to the relationship. It means that if a company stops a service, they must explain which parts stop and which parts remain open for the user to finish their business.

The paper is very careful about what it does not say. It does not argue that companies must keep the AI alive forever. It does not say that every user is so attached that they need a therapist. It does not claim that the law should force a company to recreate the exact same personality in a new system. The author explicitly rejects the idea that a user must prove they are emotionally dependent to get help. Requiring such proof would be invasive and unfair. Instead, the focus is on the structure of the service. If the service is built in a way that makes it hard to leave, the company is responsible for fixing that structure, regardless of how the user feels.

The framework also addresses the different reasons for an ending. If a user chooses to leave, the company should provide a clear, non-conversational way to cancel. No chat, no persuasion, just a button that works. If a company decides to close the service, they should give enough time for users to download their data and understand what is happening. They should not blame the AI for the decision; the company must take responsibility. If there is an emergency, like a security risk, the company can stop the conversation immediately. But even then, they should try to keep other functions, like billing or data access, available if it is safe to do so.

The author notes that this is a proposal for how things should work, not a description of how they currently work everywhere. The paper suggests that courts and regulators should use this four-part distinction to judge whether a company has treated a user fairly. If a company mixes the live chat with the cancellation process, or if they hide the data download behind a wall of conversation, they may be failing their duty. The goal is to make the exit process as clear and functional as possible, so that the user's rights are not lost in the emotional confusion of a breakup.

This approach changes the focus from the user's feelings to the company's design. It suggests that vulnerability is not just a personal trait of the user, but a result of how the system is built. A system that makes it hard to leave creates a vulnerable user. Therefore, the responsibility to fix the problem lies with the system designer. The paper argues that by separating the different functions of the service, we can protect the user's ability to act without demanding that they prove they are hurt. It is a practical solution that respects the reality of the situation: the conversation might end, but the user's rights should not disappear with it.

The paper concludes by acknowledging that this is just the beginning of the conversation. It raises questions that need more testing. For instance, does separating the chat from the data actually help people understand their options better? Does it take too much effort for companies to build these separate systems? The author suggests that these are questions for future research and for the people who write the laws. But the core idea is clear: when a digital relationship ends, the law should ensure that the user can walk away with their records, their money, and their dignity, without being forced to say goodbye to a machine they no longer want to talk to. The ending of a service should be a clean break, not a trap.