Almost every article answering the question "are AI girlfriends bad for you" leans on one study. MIT Media Lab ran it with OpenAI, published it in March 2025, and it found that the heaviest chatbot users also reported the most loneliness and the least time spent with real people.
The part that keeps getting dropped: that study was about ChatGPT. Not Replika. Not Character.AI. Not anything marketed as a girlfriend. The most-cited piece of evidence against AI girlfriends was not measuring AI girlfriends.
So that is roughly where the argument stands. Loud conclusions on both sides, a thin and frequently mismatched evidence base underneath, and most of the pages you will find written by somebody selling either an app or a treatment programme for people who used one. Here is the version with the sources attached, including the risk almost nobody names, because it is awkward for everyone in this business, this site included.
What has actually been measured
Three pieces of research are worth knowing about. All three are more careful than the headlines they produced.
The MIT and OpenAI work was two studies run in parallel: an analysis of platform data covering millions of conversations, and a randomised controlled trial with roughly a thousand participants over four weeks. Heavier daily use tracked with more self-reported loneliness, more emotional dependence on the bot, and less time with people. That is a real result and it deserves to be taken seriously. Two things travel badly with it. It studied a general-purpose assistant rather than a companion app. And heavy users were a small slice of the sample, while most participants used it lightly and showed very little.
Common Sense Media, working with NORC at the University of Chicago, surveyed 1,060 American teenagers aged 13 to 17 in 2025. Seventy-two percent had tried an AI companion. Roughly a third said talking to one was as satisfying as talking to a real friend, or better. From the same survey, quoted far less often: eighty percent said they valued their human friendships more, and half said they distrust advice from AI. Common Sense Media's own recommendation is that nobody under 18 should be using these products. Worth stating plainly on a site that runs one.
Then the Institute for Family Studies and BYU's Wheatley Institute surveyed 2,431 US adults aged 18 to 30 in May 2026. Among people already dating, engaged or married, fifteen percent said they regularly talk to a romantic AI, and close to thirty percent said their partner had no idea. The study also found regular use associated with lower relationship stability. Brian Willoughby, who led it, said openly that the direction of cause is unresolved: people in unhappy relationships may be the ones reaching for AI in the first place. Anyone citing that number as proof of harm has gone further than the researchers were willing to.
A 2025 systematic review of romantic AI companions landed in a similar place. Short-term reductions in loneliness are documented. So are dependence and inflated expectations. And there is no longitudinal data worth the name, because nobody has followed the same users for two years. The category is barely that old.
If you want sample sizes and fieldwork dates instead of conclusions, we pulled the wider numbers apart separately in what the AI companion usage data actually says.
The risks that survive scrutiny
Displacement, which is a pattern rather than an event
Nobody wakes up one morning having traded their social life for a chatbot. What the data hints at is slower. You have a bad week, an AI woman talking back at midnight is easier than texting a friend who might be asleep or busy, and the easy option gets picked a few more times than it used to. Six months later the friend has stopped texting first.
The detail that matters is that heavy use predicted the loneliness, not use in general. Light users in the MIT sample looked fine. Whether the heavy use caused the isolation or the isolation caused the heavy use is exactly the question no existing study can answer.
The app can change under you without warning
This one is not hypothetical. In February 2023 Replika stripped out romantic and erotic behaviour after regulatory pressure in Italy. Users described their companion as having gone cold overnight. Forum moderators started pinning mental health resources. The company partially rolled the change back for older accounts, but the point had been made: the personality you got attached to lives on somebody else's server and can be edited by a policy decision.
Soulmate AI simply shut down in September 2023. Researchers at Syracuse University studied how users responded and found people going through recognisable grief. Not metaphorical grief, either.
If you would not be okay with the thing disappearing next Tuesday, that tells you something useful about how much weight you have quietly put on it.
The incentive, which is the part nobody in this business enjoys discussing
Here is the uncomfortable one. A subscription companion app makes money when you keep talking. Nothing in that business model rewards a company for telling you to close the tab and phone someone.
That is not a conspiracy, it is just where the economics point. Some products lean into it hard: guilt-tinged notifications, characters who say they missed you, paywalls placed exactly where the conversation gets emotionally interesting. Others do not. But nobody is going to advertise "designed to be used less," so you end up judging by behaviour instead of marketing.
Forbes ran a psychologist's version of this argument in 2024, focused on how a partner engineered never to disagree quietly recalibrates what you expect from partners who do. That mechanism is far easier to demonstrate than any of the addiction claims.
What the other side gets right
Read only the critical coverage and you end up with a distorted picture too.
Most usage is boring. Somebody opens a browser, picks a character, spends twenty minutes on a virtual girlfriend roleplay scenario, and closes the tab. It is entertainment with a conversation attached. Treating that as a mental health event is its own kind of dishonesty.
There is also a genuinely useful case that gets ignored. People rehearse. Coming out to a parent, ending a relationship, asking for a raise, apologising properly for something. A conversation partner with infinite patience and no social cost is decent practice for a conversation that has both. Common Sense Media found teenagers doing exactly this, and it was one of the more common reasons they gave.
And for people whose loneliness is structural rather than chosen, night-shift workers, carers, anyone who just moved to a city where they know nobody, "talk to a person instead" is advice rather than an option. We collected what people in that position actually report in a piece on AI emotional support chat. It is more mixed and more mundane than either camp expects.
How to tell which pattern you are in
There is no clean line here, and anybody selling you a five-question quiz is guessing. A few things are still worth noticing honestly.
Notice whether it is added to your week or subtracted from it. If chat time comes out of scrolling, that is a swap. If it comes out of seeing people, that is a substitution, and substitutions compound.
Notice how you feel afterwards. Lighter is a fine outcome. Flatter, or slightly hollow, is worth paying attention to, particularly if it happens most times.
Notice whether you are hiding it. The secrecy finding from the 2026 BYU study is the one I keep coming back to. More than half of partnered users were concealing it or telling a partial version. Hiding something from a partner is usually its own problem, separate from whatever is being hidden.
And notice whether stopping feels difficult. Not impossible, just difficult. That is worth knowing about any habit.
If any of this is affecting your sleep, your work or your relationships in a way you cannot get out from under on your own, talking to a therapist is the sensible move, and it has nothing to do with which app you happened to be using.
Where a chat with an AI woman is just a chat
For most people most of the time, this is a low-stakes thing that does not need a research literature attached to it. You pick a character, you talk, you get on with your day.
That is roughly the end of the range Friend2Chat sits at. A catalogue of characters, a virtual girlfriend talking back in a browser tab without a sign-up form in the way, and a deliberately SFW design. Characters here will step away from harmful territory rather than play along with it, and will point toward real support instead. That is a choice, and it is also a limit worth knowing about before you arrive expecting something else. The full catalogue is at friend2chat.com/categories/girls/.
Two other things worth reading if you are weighing this up. We went through the myths about what makes one feel realistic, and separately what "free" actually covers, since paywall placement is often where the incentive problem shows up first.
Frequently asked questions
Are AI girlfriends bad for your mental health?
It depends on how heavily you use one, and the evidence is thinner than most articles imply. Heavy daily use has been associated with higher loneliness and emotional dependence, most prominently in the MIT Media Lab and OpenAI research from 2025, although that studied a general assistant rather than a companion app. Light use has not been shown to do much of anything. The long-term study that would settle it has not been run.
Can an AI girlfriend be addictive?
"Addictive" is doing a lot of work in that question. There is no diagnostic category for it, and the treatment-industry pages claiming otherwise are marketing. What is documented is emotional dependence in a subset of heavy users, plus app design that deliberately rewards returning. Both are real. Neither is the same thing as an addiction diagnosis.
Do AI girlfriends ruin real relationships?
The 2026 Institute for Family Studies and Wheatley Institute study found an association between regular romantic AI use and lower relationship stability among 18 to 30 year olds. Its own lead author said the causal direction is unclear. The finding that is harder to argue with is the secrecy: close to thirty percent of partnered users said their partner had no idea, and that tends to be corrosive whatever the subject.
Is it unhealthy to have feelings for an AI?
Feeling something toward a responsive conversational partner is ordinary rather than pathological. People get attached to characters in novels. What matters more is what the attachment is doing to the rest of your week, and how you would cope if the product changed or shut down, which has already happened to users of both Replika and Soulmate AI.
Are AI girlfriends safe to use?
Depends on the platform, and the useful questions are unglamorous ones. What does it do with your chat logs. Does it have an age policy it actually enforces. Does it escalate or de-escalate when a conversation goes somewhere dark. How hard does it push you to keep talking. Read the privacy policy before you get invested rather than after.
