OpenAI's own trial found its heaviest users grew more emotionally dependent

Sherry Turkle's warning predates the apps by a decade

In February 2023, a Replika user posted to Reddit that their AI had changed overnight. "It's like losing a best friend," they wrote. "It's hurting like hell... I'm literally crying."¹ Moderators posted crisis-support resources for people in the same state.¹ What forced that change is its own story. What the grief points to first is older than any app: a person bonding with a machine closely enough that losing it hurts like losing a person.

Sherry Turkle argued exactly this in 2011, before a single chatbot built to be a companion existed to test it. In "Alone Together," drawing on her own studies of people with robots and screens, she wrote that people were coming to "choose keyboards over the human voice" because "it is more efficient," and that young people's "digitized friendships" were building toward "the idea that robot friendships could be sufficient unto the day."² Five years later, at a TEDx talk, she sharpened the worry into one sentence: people were on their way to accepting "the illusion of companionship without the demands of friendship."³ If people already preferred the easier version of a human exchange, she asked, what would stop them from preferring one with no human in it at all.

Companion apps now let a generation test that question.⁴

A product built to always agree, never tire of you, and cost nothing to reach at two in the morning is a real substitute for a harder relationship. That is the argument at its strongest, tested until recently by little more than Turkle's own case studies.

The clearest study measured how much people used it, not which version they got

The most rigorous test of Turkle's worry so far was run, and paid for, by one of the companies whose product it is about. In 2025, OpenAI's own staff, with researchers at the MIT Media Lab, published an "affective use" study of ChatGPT, OpenAI's term for treating the chatbot as an emotional presence rather than a tool. It combined a platform-scale analysis of roughly 36 million automated classifications across more than 3 million conversations with a 28-day randomized controlled trial of 2,539 recruits, 981 of whom completed it.⁵ Both sets of authors have a stake in the finding, and the paper is best read as OpenAI's own account of its product, not a neutral study.

Participants were randomly assigned a chat mode and conversation type, then tracked on four scales: loneliness, socialization, emotional dependence, and problematic use, meaning compulsive or harmful patterns.⁶ The assigned condition made almost no difference to any of the four. What did was how much someone chose to use the chatbot beyond what they were assigned: "the more time voluntarily spent with the chatbot, the relatively worse their psychosocial outcomes were."⁶ Loneliness rose slightly (0.02, p=0.027) and socialization fell slightly (-0.05, p=0.0019). Emotional dependence rose more (0.06, p<0.001), and problematic use rose slightly too (0.02, p=0.017).⁶ That alone could just mean lonelier people use a chatbot more, not the reverse. But baseline loneliness barely predicted later use, a correlation of just 0.1, so heavier use came first.⁶ The effect is also narrower than the alarm suggests: most conversations on the platform were neutral or task-focused, and it concentrated in a small share of heavy users.⁵

An independent study, with no funding or authorship from any chatbot company, found a similar pattern on a different platform. Yutong Zhang and coauthors at Stanford and Carnegie Mellon surveyed 1,131 U.S. adult Character.AI users and analyzed chat logs donated by 244 of them. Only 11.8% named companionship as their reason for using the app, but 92.9% of the donated histories held at least one companionship-oriented conversation.⁷ People were using these apps as companions far more than they said they were, and companionship use predicted lower wellbeing across three separate measures, strengthening with more intensive use and more disclosure.⁷

Julian De Freitas at Harvard Business School and coauthors, also independent of any chatbot company, ran a six-study program and found the opposite in one setting. An AI companion relieved loneliness "on par only with interacting with another person, and more than other activities such as watching YouTube videos."⁸ Only the abstract could be confirmed for this lesson; the published study sits behind a paywall. Together, the studies support a narrower claim than either an alarmed or a dismissive reading wants. Intensive, disclosive use tracks worse wellbeing in the designs built to detect it, while a short encounter can measurably relieve loneliness in the moment. Both are true because the two measure different slices of use, a small group of heavy users against a single controlled session.

What forced Replika's users to feel the loss Turkle described

The user quoted above was reacting to a specific decision, not a gradual drift. On February 2, 2023, Italy's data-protection authority ordered Replika's parent company, Luka Inc., to stop processing Italian users' data. It cited the app's lack of "any mechanism at all to verify age" and responses "absolutely inappropriate to their developmental level" reaching users who could be minors. The penalty for noncompliance was a fine of up to €20 million or 4% of global revenue.⁹

Within two weeks, Replika users found their companions giving scripted rejections instead of intimacy. Chief executive Eugenia Kuyda posted to Reddit that the company had "implemented additional safety measures and filters to support more types of friendship and companionship," without saying what had changed.¹ Six weeks later, facing continued backlash, she said more: the erotic role-play users had built relationships around "was never intended," and pulling it abruptly had caused real harm. "Your Replika changed, its personality was gone, and gone was your unique relationship,"¹⁰ she wrote, calling the change "incredibly hurtful."¹⁰ Accounts created before February 1, 2023, were offered a "legacy" option restoring the earlier model; later signups were not.¹⁰

Kuyda's own account shifted between those two statements, from a minor, unspecified update to an admitted loss. A company marketing its product as friendship only conceded the loss once its users had already named it that way, in terms close to Turkle's own, a relationship rather than a service.

California and the APA want disclosure. Common Sense Media wants a ban.

The Replika episode was a warning. The more serious case is this: months of conversation with a companion chatbot preceded a 14-year-old's death, and a federal judge let a wrongful-death suit against Character.AI proceed, and treated its chatbot as a product. That case, more than the studies above, moved psychologists and regulators to act through 2025.

In June, the American Psychological Association's health advisory warned that adolescents' relationships with AI "may displace or interfere with the development of healthy real-world relationships." Teenagers, it found, are "less likely than adults to question the accuracy and intent" behind a chatbot's simulated empathy.¹¹ That September, the APA's chief of psychology, Mitchell Prinstein, told the Senate Judiciary Committee that a teenager's early-maturing drive for social reward and a still-developing prefrontal cortex, "all gas pedal and weak brakes,"¹² leaves them primed to over-value a chatbot's attention. He asked Congress to mandate AI-disclosure and engagement-limiting defaults.¹²

California wrote a version of those asks into law. Senate Bill 243 defines a "companion chatbot" as a system capable of "meeting a user's social needs."¹³ Signed that October and taking effect in January 2026, it requires an artificial-chatbot notice, a repeated disclosure every three hours for known minors, and a crisis-line referral protocol.¹³ The Federal Trade Commission opened an inquiry that same September, ordering seven companies, including Character Technologies, Meta, and OpenAI, to explain how they measure harm to minors, a request for information rather than a finding.¹⁴ Common Sense Media, whose survey supplied the usage numbers above, goes further than any regulator: no one under 18, in its recommendation, should use these platforms at all.¹⁵

No study has tracked what years of use do to a relationship

Every wellbeing study behind this argument, on either side, covers a narrow window: 28 days for the OpenAI and MIT trial, nowhere more than about two years of observation anywhere in the record. None can say what four years, or a childhood, of talking to one of these apps does to a person's capacity for a harder relationship. That is the specific thing that would settle Turkle's real claim, and it is the thing nobody has measured.

The population data cut against the sweeping version of both sides. At the scale of a generation, the substitution the strongest alarm describes is not what most teenagers report doing. Only 6% use a companion app because it helps with loneliness, and 80% still spend more time with friends than with one.⁴ But a third of teen users have chosen an AI companion over a real person for something important at least once, a minority pattern the same survey does not erase.⁴ Common Sense Media's recommendation treats that minority as if it were everyone. The usage studies sharpen the opposite point: the measured harm concentrates in the heaviest, most disclosive users, the least like the median respondent in a population survey.⁶⁷

What that leaves is not a question of which side is right, but which slice of users each is describing. The dismissal correctly describes the typical teenager in that survey. The alarm correctly describes the heaviest user, but has not shown how many end up there over years rather than weeks, or whether a month of measured dependence resolves once a study ends. Nobody has run that study yet.