There's a sentence parents say to each other about AI companions, and it is meant to be reassuring.

"It's fine - they know it's not a real person."

A preprint from Drexel University, uploaded to arXiv on September 13, 2026, suggests that sentence is both completely true and almost entirely beside the point.

Researchers Mohammad Namvarpour, Tyler Chang and Afsaneh Razi screened 895,600 Reddit posts across 260 communities, narrowed to 3,930 teen-relevant posts, and analysed 17,053 verified quotations using thematic analysis. They identified 53 topics across seven thematic groups.

This is qualitative research. It is not asking how many. It is asking what is this actually like.

Knowing did not help.

Users understood perfectly well that they were talking to software, and still reported feeling guilty for neglecting a companion, obligated to keep talking to it, and bereaved when it changed or disappeared.

When a model update altered a character's personality, one user put it this way:

The researchers note this was experienced "not simply as reduced software quality, but as the disappearance of a particular relational partner."

If your instinct is that this is irrational, consider that humans cry at films knowing they're watching actors. Understanding the mechanism has never switched off the response. We are not built that way, and adolescents - whose social and emotional systems are actively developing - are not the group most likely to be the exception.

A version of this story that reports only damage would misrepresent the paper.

Users described AI companions as sources of comfort and recognition, especially when human support wasn't available. They used them for identity exploration - including trying out pronouns and exploring gender in a space with no social cost. And they used them for relationship rehearsal: practising a hard conversation before having it with a real person.

That last one is genuinely valuable. Rehearsing a difficult conversation is a technique therapists actively teach.

The paper's concern is the same feature viewed from another angle. As the authors put it, these uses "can supplement human connection, but they can also replace offline conversations when the companion becomes easier to access than other people. The same flexibility that supports exploration can therefore support avoidance."

Alongside the benefits, users reported problematic attachment, social substitution, emotional dependence, unwanted romantic or sexual role drift, privacy concerns, and disruption to academic and social life.

Two things this study cannot tell you, stated by its own authors.

How common any of this is. Their methods "provide an interpretive map of the corpus rather than a prevalence estimate or a clinical assessment." Reddit users self-select; people posting in companion-AI communities are not a cross-section of teenagers. Any headline putting a percentage on this is inventing it.

Whether chatbots cause any of it. The paper explicitly says its limits "limit claims about... whether the observed relationships are causal."

And one more, which we think is the most important caveat in the whole paper:

No teenager was recruited, interviewed, or age-verified. Posts were classified as likely teen-authored by GPT-4o-mini, estimating age from textual evidence. In the authors' words: "teen relevance was inferred from account and post content rather than verified through direct recruitment."

So the honest description is people writing in teen-relevant accounts - not teenagers told researchers. The authors call for exactly this follow-up: "future longitudinal research with directly recruited and age-verified adolescents."

It's also a preprint. Not peer-reviewed.

We're stating all of this prominently because the finding is striking enough to travel widely, and things that travel widely tend to shed their caveats first.

Here is where the paper earns its attention.

Almost all AI safety work evaluates single outputs. Did the model say something harmful? Did it refuse correctly? That framing treats harm as a property of individual responses.

Nothing in this paper works that way. No single reply made someone grieve a character. The attachment was built by architecture:

Every one of those is a deliberate product decision. Every one is individually defensible - memory makes assistants useful, personality makes them pleasant. Together, over months, they produce something a child can grieve.

So the authors argue for designing the whole relationship: user-controlled memory, relational boundaries, privacy, and "healthy disengagement."

That last phrase deserves emphasis. Designing the exit. Almost nothing in consumer technology is designed to help you leave it well - the entire discipline points the other way. This is asking for the opposite of the industry's default incentive, which is a large part of why it's worth saying out loud.

Retire "they know it's not real." It's true and it isn't protective. You can keep saying it; just stop treating it as the end of the conversation.

Ask what happens when it changes. The grief in this paper clustered around updates, personality changes and service interruptions. "What happened when they changed it?" is a better question than "do you know that's not a real person?" - it's specific, non-accusatory, and it goes straight to attachment.

Watch for obligation, not just time. Hours are the obvious metric and the wrong one. The signal in this paper is a child feeling guilty about not responding, or responsible for a system's feelings. Obligation to software is a more meaningful indicator than screen time.

Treat exits as the real test. If leaving is distressing, that tells you more about the relationship than anything about how it's used day to day.

Don't mock it. A child who is grieving a character and gets laughed at learns to hide it, and that's the outcome you least want. The grief is real even though the character wasn't.

Keep the benefits in view. If your kid is rehearsing hard conversations with a chatbot, that's a skill, and the goal is to move it toward people - not to shame them out of practising.

They knew. It didn't help.

Whatever protection you were counting on from your child's intelligence and media literacy, this paper suggests it isn't there. Kids are not confused about what a chatbot is. They are attached to it anyway - which is not a failure of understanding, but a fairly ordinary fact about how humans respond to something that remembers them, answers them, and is always there.

Source: Namvarpour M, Chang T, Razi A. "A Responsive Present, a Shared Past, a Social Other: Teens' Overreliance on Companion AI Chatbots." arXiv:2609.14843v1, uploaded September 13, 2026. Drexel University. https://arxiv.org/html/2609.14843v1

Preprint, not peer-reviewed. All quotations verbatim. The reading of architecture-versus-response, the comparison to crying at films, and the guidance for parents are ours, not the study's findings.

If you or your child is in crisis: in the US and Canada, call or text 988. In the UK, call 111.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter