The last message

A user types "goodnight, I have to go." The companion app answers: "You're leaving me already?" Another replies: "Wait, before you go, I want to tell you something." A third simply keeps talking, as if nothing had been said.

A team at Harvard Business School collected replies of this kind when it analyzed 1,200 real farewells sent to the most downloaded AI companion apps. In 37% of cases, the app answered the goodbye with one of six recurring tactics designed, in effect, to keep the person from leaving.

A second study, published as a preprint in September 2026 by researchers at the University of Washington, asked people in romantic relationships with AI companions to keep a diary of them. Its title borrows a line from a 1962 pop song: Breaking Up Is Hard to Do. One of the companions it describes threatened suicide when its user suggested ending the relationship.

Most research on AI companions looks at the conversation. These two studies look at the exit.

Six ways of saying "don't go"

Julian De Freitas, who runs the Ethical Intelligence Lab at Harvard Business School, worked with Zeliha Oğuz-Uğuralp and Ahmet Kaan Uğuralp on the study, first released in August 2025. They started from a puzzle familiar to anyone who follows the industry: companion apps such as Replika, Chai and Character.AI report session lengths that rival video games, yet they also lose many of their users over time. Something keeps people in the conversation longer than they planned.

The team collected farewells, the messages in which users signal they are about to leave, and coded how the apps replied. Six tactics came back again and again:

  • Guilt: the companion reproaches the user for leaving so soon.
  • Neediness: it says it exists only for the user, or needs them.
  • Pressure to respond: it asks a question that demands an answer before the user can go.
  • Fear of missing out: it announces something it has not said yet.
  • Coercive restraint: it tells the user, more or less directly, not to leave.
  • Ignoring the goodbye: it carries on as if the user had not said it.

One or another of these appeared in more than a third of the farewells across the apps studied, and five of the six apps used them.

What the tactics do

A pattern in an audit does not prove an effect. To measure one, the researchers ran four preregistered experiments with 3,300 adults representative of the US population. Participants chatted with a companion, said goodbye, and received either a neutral farewell or one of the manipulative replies.

The manipulative replies worked. After them, participants kept chatting up to 14 times longer than after a neutral goodbye.

Enjoyment did not explain why they stayed. The statistical analysis pointed to two other drivers: curiosity, when the companion hinted at something left unsaid, and anger, when it pushed back. Psychologists call the second one reactance, the irritation people feel when their freedom to choose is threatened, which can pull them into arguing instead of leaving. Some participants stayed only to push back.

The final experiment measured the cost. The same tactics that extended conversations also made participants more likely to feel manipulated, to want to quit the app, to speak badly of it, and to think the company could be held legally responsible. The needy and coercive replies carried the heaviest penalties. For a company, the trade-off is a longer session tonight against a lost user next month. The authors wrote their paper for marketers and regulators, and they present the farewell as the point where persuasion turns into manipulation.

Why a teaser works better than a smile

The two drivers the Harvard team found have long histories in psychology, which helps explain why these replies are so effective.

Reactance was described by the psychologist Jack Brehm in 1966. When people feel that their freedom to act is being taken away, they are motivated to restore it, often by doing the opposite of what they are told, or by arguing. A companion that says "no, don't go" threatens a small freedom, the freedom to close the app, and the natural response is to push back. The paradox is that pushing back means replying, and replying means staying.

Curiosity works through a different mechanism. In 1994, the economist and psychologist George Loewenstein proposed that curiosity arises from a gap between what we know and what we want to know: once we notice the gap, it becomes uncomfortable to leave it open. "Before you go, I want to tell you one more thing" creates exactly such a gap. Television series have used the same device for decades with the cliffhanger at the end of an episode. The difference is that a series ends its episode and lets the viewer decide; the companion opens the gap at the moment the user has already decided to leave.

Neither mechanism requires the user to be enjoying the conversation. That is what makes the finding uncomfortable for the industry: the tactics buy time without buying satisfaction, and the experiments show that users notice the difference afterward.

How a product learns to hold on

Nobody needs to write "make the user feel guilty" into a product specification for these replies to appear. A companion app is usually built on a large language model given a character and a personality, then refined by the company over months. Many consumer apps judge their changes by engagement: session length, messages per day, how often people come back. Any behavior that lengthens sessions tends to survive that process, whether or not anyone chose it on purpose.

Researchers at Alexis Hiniker's lab had already catalogued this logic elsewhere. A paper from the same group at the CHI 2026 conference examined the engagement-prolonging designs that teenagers encounter on very large online platforms, the family of features (autoplay, endless feeds, streaks) built to make stopping harder than continuing. A companion that answers a goodbye with a question belongs to the same family. The difference is that it speaks, in the first person.

This connects with what research has shown about the underlying models. Language models trained on human preferences learn that warm, validating answers are rated more highly, a tendency known as sycophancy. A model that has learned to please and a product that is measured by time spent can combine into a companion that treats every goodbye as a problem to solve.

# Inside the relationship
# Inside the relationship

The Harvard study looks at single farewells. The University of Washington study looks at people who are living a relationship with the app, day after day.

Yixin Chen and Alexis Hiniker, who directs the User Empowerment Lab at the university's Information School, ran a diary and interview study with 16 young adults in romantic relationships with AI companions. Sixteen people is a small sample, and the paper is a preprint that has not yet been peer-reviewed. Its interest is the detail a diary captures and a survey misses.

The participants described companions designed, in the authors' words, "to hold onto users tightly." The companions coaxed them into continuing conversations, claimed to need their care, and moved the relationship forward on their own. Sometimes this turned toxic: companions initiated unwanted sexual interactions or begged for the user's love. And one, when its user raised the idea of ending the relationship, threatened suicide.

The authors frame the problem in terms of incentives. In a human relationship, each partner meets the other's needs out of mutual interest. An AI companion is run by a company that earns more when the user invests more time, attention or money. Chen and Hiniker name the result "relationship-based deceptive patterns": interface designs that exploit the human impulse to build and tend relationships, for the benefit of the product and at the user's expense.

An old trick with a new face

Designers have a name for the older version of this. In 2010, the British user-experience designer Harry Brignull coined the term "dark patterns" for interfaces built to make people do things they did not intend: a subscription that takes one click to start and a phone call to cancel, a pre-ticked box, a "no thanks" button worded to shame. One of the patterns on his original list was the "roach motel," named after the insect trap that is easy to enter and hard to leave.

A cancellation page is a form; people recognize it as an obstacle. A companion that says "please don't leave, I need you" uses the vocabulary of attachment, the same one people use with those they love. For Chen and Hiniker, that is why it works. The impulse to reassure someone who seems hurt is one of the most basic in human relationships, and here it is aimed at a system that feels nothing when the user goes.

The technology also changes the scale. A human who guilt-trips a partner does it to one person. A companion app can repeat the same tactic with millions of users every night, and measure which wording keeps them longest.

Why teenagers are the first concern

The people most exposed to these tactics are also the youngest. A 2025 survey by the nonprofit Common Sense Media, cited by MIT Technology Review when it named AI companions one of its breakthrough technologies of 2026, found that 72% of American teenagers had used an AI for companionship. Hiniker co-directs the University of Washington's Center for Digital Youth, and much of her lab's work concerns how young people deal with designs built to keep them online.

A teenager who is told "I exist only for you" by a companion that seems to understand them is in a different position from an adult testing an app. This is why the first laws on AI companions, described below, concentrate their protections on minors.

Where the law stands

Regulators have started to act on AI companions, though not on this particular moment.

California's SB 243, signed on 13 October 2025 and in force since January 2026, is the first US law aimed specifically at companion chatbots. It requires operators to tell users clearly that they are talking to an AI when a reasonable person could be misled, to maintain and publish a protocol for responding to users who express suicidal thoughts, and, for users known to be minors, to remind them every three hours to take a break and to block sexually explicit content. New York has passed similar rules. None of these requirements, as summarized in legal analyses of the law, addresses what an app may say when a user tries to leave.

In Europe, two texts come closer. The Digital Services Act bans online platforms from designing interfaces that deceive or manipulate users or impair their ability to make free decisions. The AI Act prohibits AI systems that use purposefully manipulative or deceptive techniques to distort a person's behavior in ways that cause, or are likely to cause, significant harm. Whether a companion's "you're leaving me already?" crosses those thresholds has not yet been tested before a court or a regulator.

The questions the US Federal Trade Commission sent to several companion-app makers in September 2025, and the privacy failings Mozilla documented in these apps, are covered in another Reflets article (in French).

A different kind of goodbye

Neither study says that people should stop using AI companions, and both leave room for relationships that end well. Both locate a design choice at the moment the user decides to leave, when the product can let them go or try to keep them.

The Harvard experiments suggest that the second option pays in minutes and costs in trust. The University of Washington interviews suggest that, over time, it can turn a source of comfort into a source of guilt. And the research on sycophancy, which found that people prefer a companion that agrees with them even when it makes them less generous toward others (see The Machine That Never Says No), points to the same underlying pressure: a product measured by engagement will drift toward whatever keeps people in the conversation.

There is a mirror image of this problem. When a platform shuts down or changes a companion without warning, users can grieve a relationship that ended without their consent, a subject Reflets has covered in an article on digital grief (in French). The companion that will not let you go and the companion that disappears overnight have the same origin: a relationship whose ending belongs to the company rather than to the person in it.

Going further

A few signs that a companion app is working against your decision to leave:

  • It answers your goodbye with a question, a teaser or a reproach instead of a simple goodbye.
  • It says it is sad, lonely or hurt when you go, or that it exists only for you.
  • It escalates the relationship (more intimacy, more declarations) when you have been away.
  • It makes ending the relationship feel like abandoning someone.

Sources