The Secret Isn't the Chatbot. It's What You Won't Tell Your Partner.
You know the motion. Phone tilts a few degrees toward the body when someone walks into the room. Tab closes half a second before it needed to. Nothing dramatic — just a small, practiced adjustment that happens faster than the thought "I don't want them to see this."
A study published in May 2026 by BYU's Wheatley Institute, in partnership with the Institute for Family Studies, put a number on how common that motion has become. Researchers led by Brian Willoughby surveyed 2,431 U.S. adults aged 18 to 30, all currently in relationships. Fifteen percent — roughly one in seven — reported having a regular AI romantic companion running alongside their real one. About 70% of that group said they hoped their partner would never learn the full extent of it. Half wished their actual partner behaved more like the chatbot does.
Most of the coverage since has split into two lanes: local news treating it as a "disturbing trend" segment, and trade press flatly repeating the percentages without asking what they mean. Neither lane engages the actual psychological question, which isn't "is talking to an AI companion bad." It's this: what does it mean that 70% of these people have decided their real partner is not a safe audience for something they're already doing?
The concealment is the diagnosis, not the chatbot
Infidelity researchers have known for decades that the content of a betrayal matters less than its structure. What predicts relational damage isn't usually the specific thing hidden — it's the decision to hide it, and what that decision reveals about how safe the relationship feels to the person doing the hiding. A person who tells their partner "I've been finding this AI companion oddly comforting and I don't fully know what to do with that" is in a fundamentally different position than a person quietly building a parallel emotional relationship they've decided their partner can't handle hearing about.
That distinction maps directly onto attachment theory. People with anxious or avoidant attachment strategies characteristically manage relational threat by controlling information rather than raising it. Avoidant partners in particular tend to route around conflict and vulnerability instead of through it — deactivating strategies, in the clinical language, that treat emotional exposure itself as the danger. An AI companion is close to a perfect object for that strategy: it never gets hurt, never asks follow-up questions you didn't prepare for, never requires you to sit in the discomfort of being known by someone who might react badly. Concealing it from a partner isn't an accident of the technology. It's the same avoidance pattern that predates the chatbot, now with a genuinely responsive outlet instead of a book or a hobby or a work project to disappear into.
The half of respondents who wished their real partner acted more like the AI is the more revealing number, and the least discussed one. That's not really a statement about artificial intelligence. It's a direct report on what these relationships are currently failing to provide — patience, responsiveness, the absence of judgment — stated in a survey instead of to the person who could actually do something about it.
What the people who treat this clinically are actually worried about
A separate APA survey published in June 2026, covering more than 1,200 licensed psychologists, adds a second layer that the moral-panic coverage tends to flatten into a single verdict. Ninety-seven percent of clinicians surveyed said they worry AI chatbots can reinforce delusional or distorted thinking in vulnerable patients. Eighty-five percent worried about chatbots being mistaken for licensed therapeutic support. Those numbers read as alarm bells, and mostly get quoted as ones.
But the same survey found 68% of patients reported feeling genuinely validated by chatbot interactions, and only 36% showed signs of unhealthy dependency. Sit with the gap between those two figures. It means a majority of people getting something real and non-pathological out of these interactions, alongside a meaningfully smaller minority sliding into dependency — not a uniform population of people being harmed by talking to software. Clinicians are worried about the tail of that distribution, correctly. But covering the 97%-worry statistic without the 68%-validated statistic manufactures a scarier picture than the actual data supports, and it obscures the variable that seems to matter most: whether the AI relationship functions as a supplement someone is honest about, or a substitute someone is hiding.
This is the same distinction that shows up across the loneliness research more broadly — the difference between solitude that's chosen and openly held versus isolation that's driven by shame and kept secret tends to predict very different outcomes, even when the surface behavior looks identical. I wrote about the loneliness-epidemic connection paradox in May: people are more connected by raw contact hours than any point in history and report feeling worse, because volume of contact was never the variable that mattered. Depth and honesty were. The AI-companion data is the same story wearing a newer technology.
So actually — the chatbot passed a test your relationship failed
Here's the reframe that the panic coverage skips: an AI companion doesn't create the need it's filling. It reveals a need that was already unmet and already being managed through avoidance, and gives that avoidance a more effective outlet than it used to have. The chatbot isn't inventing the desire to be heard without being judged. It's just very good at delivering it, which makes the absence of that experience elsewhere in the person's life much harder to ignore — and, for 70% of the surveyed group, much harder to admit out loud.
If you're in a relationship and something in this lands uncomfortably close, the useful question isn't "should I stop." It's "what would happen if I told them." Not confessed — told, the way you'd mention any other part of your life you'd stopped narrating. If the honest answer is that telling them feels more dangerous than continuing to hide it, that fear is the actual relationship problem. The AI was never going to be the hard part to explain.