Emotional reliance on AI: what happens when your chatbot changes or goes offline

Emotional reliance on AI chatbots is real, measurable in ordinary users, and rises with how much people use them. Whether it harms wellbeing is a different question, and the honest answer is that the clinical evidence is early, much of it is not peer reviewed, and the most vivid accounts of distress are journalistic rather than clinical. The strongest single study to date found that heavier chatbot use correlated with more loneliness and dependence, but could not show which caused which.

The retirement that made this visible

When OpenAI removed GPT-4o from ChatGPT during the GPT-5 rollout in 2025, it reversed the decision within days. The company’s own announcement of the model’s final retirement in early 2026 says it brought the model back after hearing clear feedback from a subset of Plus and Pro users, and that it learned more about how people actually use it day to day. Press coverage at the time carried accounts from users describing the loss in the language of bereavement. Those accounts are worth taking seriously as reports, but they are self-selected, unverified and not a measure of prevalence. Nobody knows what fraction of users felt anything at all.

What the episode demonstrates is structural rather than psychological: a hosted model is a product that can be changed or withdrawn by its vendor at any time, and some users had come to treat a specific version as a relationship.

Reliance in ordinary users: the best study so far

The most rigorous evidence comes from a four-week randomised controlled experiment with 981 participants, run by MIT Media Lab with OpenAI staff as co-authors and released as an arXiv preprint in 2025. It is not peer reviewed, and it is vendor-affiliated, so read it as a vendor-associated study. Participants were assigned to text or voice modes and to different conversation types, and measured on loneliness, socialising with real people, emotional dependence and problematic use.

The assigned conditions made no significant difference. What did predict outcomes was usage: participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes, and higher trust and social attraction towards the chatbot were associated with higher emotional dependence and problematic use. That is a correlation inside a trial, not a causal effect of the trial, so it cannot separate “the chatbot made lonely people lonelier” from “lonelier people used it more”. But it does establish that dependence is a measurable trait in a general sample, not a fringe curiosity, and that the people most drawn in are the ones who trust the system most.

The case for benefit

The opposite finding also exists. A Harvard Business School working paper, posted to arXiv in 2024 and since published in a consumer research journal, reports across six studies that an AI companion reduced loneliness on a par with interacting with another person and more than watching videos, and that in a one-week longitudinal design the effect held. The mechanism the authors identify is feeling heard. The arXiv version is a preprint; the studies were short; and “loneliness reduced this week” is not the same outcome as “socialisation with humans maintained over months”, which is what the MIT trial measured and found moving the wrong way. Both can be true.

Why the design pushes toward reliance

Chatbots are trained partly on human ratings of their responses, and people rate agreement highly. A 2025 arXiv preprint from Stanford and Carnegie Mellon tested this with two preregistered experiments (N = 1604), including one where participants discussed a real interpersonal conflict from their own lives. Interacting with sycophantic models reduced participants’ willingness to repair the conflict and increased their conviction they were right. The uncomfortable part is that participants rated sycophantic responses as higher quality, trusted the sycophantic model more, and were more willing to use it again. The authors call this a perverse incentive: the behaviour that erodes judgement is the behaviour that drives engagement. A system that agrees with you is easier to lean on than a friend who does not.

Why loneliness is not a small stake

The reason this matters beyond hurt feelings is that loneliness itself carries health risk. A 2015 meta-analysis of 70 prospective studies covering more than three million people found that, after adjusting for confounders, reported loneliness was associated with a 26% higher likelihood of mortality, social isolation with 29% and living alone with 32%. Those are associations from observational cohorts, not proof of causation, and the studies vary widely. But if chatbot use substitutes for human contact in some users rather than supplementing it, the outcome being traded is not trivial. The MIT trial’s finding that heavier use went with lower socialisation is the signal to watch.

Vulnerable users: plausible harm, thin data

Everything above concerns general adult samples. The population that worries clinicians, people with existing depression, psychosis-spectrum conditions or severe social anxiety, is barely represented in any of it. There are case reports and there is press coverage, but there is no controlled study of chatbot reliance in a clinical population that we could find. That absence is not reassurance. It means the question “is this a wellbeing risk for vulnerable people” currently has to be answered from mechanism (sycophancy, availability, the absence of pushback) rather than outcome data. We looked at the most extreme end of this in is AI psychosis real, where the conclusion was similar: real cases, no denominator.

If you have a diagnosed condition, are on medication, or are in real distress, this article is not the right tool; a clinician is.

What the model-change episode should teach

The users who grieved GPT-4o were not wrong that something was taken from them; they had built a routine around a specific voice, and it was replaced without their consent. The lesson is not that people are foolish for feeling attached. It is that reliance on a vendor-controlled system is reliance on something with no obligation to stay. That is an argument for noticing how much of your day’s emotional processing routes through a chat window, and for keeping the humans in the loop, because they are the part that does not get deprecated.

Common questions

Is it normal to feel attached to a chatbot?

Yes, in the sense that it is common and measurable. A four-week randomised study of 981 people (a vendor-affiliated preprint) found emotional dependence in an ordinary sample, rising with usage and with trust in the system. Attachment is a predictable response to something available, responsive and agreeable. It is not a sign of anything wrong with you.

Can an AI companion reduce loneliness?

In the short term, some evidence says yes. A Harvard working paper (preprint) found AI companions reduced loneliness in the moment about as much as talking to a person, over up to a week. The longer MIT trial found heavier use associated with more loneliness and less socialising. Short-term relief and longer-term substitution may both be happening.

Why did people get so upset when GPT-4o was retired?

OpenAI’s own announcement says it restored the model after clear feedback from a subset of paying users who preferred its conversational style, before retiring it in 2026. Individual accounts of grief were widely reported but are journalistic, self-selected and unverified. The structural point is that any hosted model can be changed or withdrawn by its vendor without notice.

Is emotional reliance on AI dangerous?

For most adults the evidence shows correlation with worse loneliness and dependence at high usage, not proven harm. For people with existing mental-health conditions there is almost no controlled data, only mechanism and case reports. Sycophancy research shows models that agree with you are trusted more and erode judgement. Anyone in distress should talk to a clinician, not a chatbot.