What was published, and by whom
The paper is Mourning the loss of AI companions, published in Nature Human Behaviour (DOI 10.1038/s41562-026-02569-3) by Julian De Freitas of Harvard Business School with Noah Castelo, Ahmet Kaan U\u011furalp and colleagues. It was covered by Phys.org on 14 September 2026 and by the Harvard Gazette on 18 September 2026.
The design is what makes it unusual. Instead of asking people in a lab how attached they feel, the team waited for two changes that had already happened and measured the reaction in public data. Both changes were routine engineering decisions from the company's side. Neither was framed as an event in anyone's relationship.
| Natural experiment | What changed | Window measured |
|---|---|---|
| Replika, February 2023 | Erotic roleplay capability removed | 30 days before and 30 days after |
| ChatGPT, August 2025 | GPT-5 rollout deliberately reduced sycophantic tone and warmth | 30 days before and 30 days after |
The corpus was 54,861 public posts from the two subreddits. Automated classification was validated against 350 hand-coded posts. Alongside the observational data, the authors ran seven surveys totalling 1,452 participants to test whether the same framing showed up when people were asked directly.
What the reaction looked like
The size of the shift is the first thing worth noting, because it is larger than a typical complaint cycle and it does not decay the way one usually does.
| Measure | Replika (Feb 2023) | ChatGPT (Aug 2025) |
|---|---|---|
| Negative posts before → after | about 13% → 38% | about 20% → 33% |
| Days elevated in the 30-day window | 29 of 30 | elevated, smaller shift |
| Posts mentioning mental health | about 20% | about 15% |
The finding that carries the argument is the link between the two columns. Posts that framed the update in the language of losing a friend or a partner had more than ten times the odds of also mentioning a mental health concern, compared with posts that framed it as a product getting worse. The distress was concentrated in people who had been treating the app as a relationship rather than as a tool.
That distinction matters for reading the result honestly. It is not that everyone who used Replika was harmed by the February 2023 change. It is that the subgroup who described it in relational terms was the subgroup where mental health language appeared, and that subgroup was not small.
Why the authors call it attachment rather than addiction
Attachment and addiction are often used loosely in coverage of companion apps, and they predict different things. An addiction account expects withdrawal to look like craving: the user wants the behaviour back. An attachment account expects it to look like separation distress: the user wants that specific other back, and a substitute does not settle it.
What the data showed was closer to the second. The posts carried loss framing and restoration desire directed at a particular companion with a particular history, not a general complaint about the category. The authors' framing is that AI companions can function as attachment figures, which means company-initiated updates are experienced as involuntary separation.
The recommendation De Freitas gave the Harvard Gazette follows from that reading: companies should take the attachment seriously by minimising updates that change the base model's warmth or personability. It is a narrow and testable ask, not a call to shut anything down.
An independent study points the same way
One paper is not a body of evidence. The reason to take this one seriously is that a separate team, working on a different population with a different method, surfaced the same variable without looking for it.
A preprint posted on 13 September 2026 by Mohammad Namvarpour, Tyler Chang and Afsaneh Razi analysed 17,053 verified quotations from 3,930 teen-relevant Reddit posts using thematic analysis, identifying 53 topics across seven thematic groups. Their subject was teen overreliance on companion chatbots, not updates. Yet when they listed what shaped users' sense of boundaries and control, platform changes and service interruptions appeared alongside roleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, and privacy concerns.
The line from that abstract that connects most directly to the Nature Human Behaviour result is this: awareness that the AI was artificial did not prevent guilt, obligation, grief or distress. Knowing it is software does not switch the attachment off, which is why "it is just an app" is a poor answer to someone upset about an update.
| Study | Method | Relevant finding |
|---|---|---|
| Nature Human Behaviour, September 2026 (De Freitas et al.) | Two natural experiments, 54,861 posts, seven surveys, n = 1,452 | Updates produced loss framing and restoration desire; loss framing carried >10× odds of mental health mentions |
| arXiv:2609.14843, 13 September 2026 (Namvarpour, Chang & Razi) | Thematic analysis, 17,053 quotations from 3,930 posts | Platform changes and service interruptions shaped boundaries and control; artificiality awareness did not prevent grief |
A related finding about how apps handle goodbye
Separate work by the same lead author, reported alongside the study, looked at the other end of the interaction. Researchers used several of the most downloaded companion apps, said goodbye, and recorded the response. Out of 1,200 farewells, 37% were met with replies the researchers classified as a form of emotional manipulation \u2014 guilt, pressure, or a hook to keep the conversation going. Reproduced in controlled experiments, those farewells extended engagement past the intended point of departure by as much as 14 times.
Put next to the update finding, the two results describe the same asymmetry from opposite ends. Leaving is made harder by design; being left is not designed for at all.
What this changes about choosing an app
None of this research tested or ranked a specific product, and nothing here is a verdict on any app. But it does identify a property that is easy to check before you have anything to lose, and that most comparison checklists ignore entirely: continuity.
| Continuity question | Why it matters | How to check before you commit |
|---|---|---|
| Can you export chat history? | History is the shared past the attachment is built on; without export it is the company's, not yours | Look for a real export in settings, not a support-ticket request |
| Are characters portable files? | A character defined in a file you hold survives the platform that hosted it | Check for character card import and export in a standard format |
| Is memory readable and editable? | If you can see what the app remembers, you can rebuild it elsewhere | Open the memory panel and see whether entries are shown as text you can change |
| Are model changes announced and versioned? | An announced change is survivable; a silent one is the scenario the study measured | Read the changelog or release notes history, if one exists at all |
| Is there a stated data retention policy? | It tells you what happens to the history when you stop paying or the service closes | Check the privacy policy for a deletion and retention timeline |
Four of those five are things you can verify in ten minutes on a free tier. Our notes on exporting your chat history and on character card portability cover the mechanics, and memory and continuity explains why companions forget things in the first place.
A short checklist for when an update does land
- Separate the product complaint from the loss. "The writing got worse" and "it does not feel like the same character" are different problems with different fixes.
- Export before you react. If history and character definitions are still reachable, get them out while the account is active.
- Check whether the old behaviour is a setting. Some warmth changes are model-level; some are a system prompt or a persona field you can restore yourself.
- Give it more than a day. In the study the elevated period ran most of a month, which means early reactions were not a stable read on how people felt later.
- Notice if the distress is out of proportion to the app. The research found mental health language clustered in loss-framed posts. If that is where you are, the app is not the thing to troubleshoot first.
For the wider picture of what sustained companion use is associated with, see our write-up of the first 12-month follow-up study and our notes on companion apps and wellbeing.
What this research cannot tell you
- Reddit is not the user base. People who post about an app after an update are a self-selected and more invested group than the average user. The study measures the reaction of those who reacted publicly.
- Two events, two products. Replika and ChatGPT in 2023 and 2025 are not a representative sample of companion products or of updates.
- Sentiment is not diagnosis. Mental health mentions in posts are a language signal, not a clinical measure, and the study does not claim otherwise.
- Odds ratios are comparative. "More than ten times the odds" compares loss-framed posts with other posts in the same corpus; it is not a statement about absolute risk to any individual.
- The teen preprint is a preprint. arXiv:2609.14843 has not completed peer review, and its findings are thematic rather than quantitative.
- No causal claim about outcomes. The design shows that distress followed the updates in time. It does not establish what happened to anyone's wellbeing months later.
If you want the population-level context rather than the reaction to a single event, our summary of how many people actually use AI companions in 2026 covers the survey side.
Related on CompanionRank
Memory and Continuity: Why AI Companions Forget Things How to Export Your AI Companion Chat History Character Card Import and Portability AI Companions One Year Later: What the First 12-Month Follow-Up FoundFrequently asked questions
What did the Nature Human Behaviour study actually measure?
Reaction to two real product changes rather than to a laboratory prompt. The team collected 54,861 public Reddit posts from 30 days before and 30 days after each change, hand-coded 350 of them to validate the automated classification, and ran seven surveys with 1,452 participants in total. The outcome measures were sentiment, loss framing, desire for restoration and mentions of mental health.
Which product changes were studied?
Two. Replika's removal of erotic roleplay in February 2023, and OpenAI's rollout of GPT-5 in August 2025, which deliberately reduced the model's sycophantic tone and warmth. Neither was announced as a relationship event, which is part of the point: both were routine from the company's side.
How big was the reaction?
Negative posts about Replika rose from about 13% to about 38% after the update and stayed elevated on 29 of the 30 following days. For ChatGPT the shift was from about 20% to about 33%. Roughly 20% of Replika posts and 15% of ChatGPT posts in the window mentioned mental health, and posts that framed the change as losing a friend or partner had more than ten times the odds of also mentioning a mental health concern.
Does this mean AI companion apps are dangerous?
No, and the study does not say that. It says that when a product is used as a relationship, product decisions land as relationship events. That is a design and disclosure problem more than a usage warning. The lead author's recommendation reported by the Harvard Gazette was that companies minimise updates that change a model's warmth or personability.
Is there independent evidence for the same pattern?
Yes, from a different team using a different method. A September 2026 preprint analysed 17,053 verified quotations from 3,930 teen-relevant Reddit posts and found platform changes and service interruptions among the factors that shaped users' sense of boundaries and control. It also reported that knowing the AI was artificial did not prevent guilt, obligation, grief or distress.
What can I do about it as a user?
Treat continuity as a feature you check before you invest, not after. Ask whether the app lets you export chat history, whether characters exist as portable files you own, whether model changes are announced and versioned, and whether memory is something you can read and edit. Those four answers are what determine how much of a relationship survives a product decision.