One in seven young adults who are dating, engaged, or married regularly talks to an AI romantic companion, and most of them keep it from the person they are actually with. The figure comes from Secret Soulmates, a report published in May 2026 by the Wheatley Institute at Brigham Young University together with the Institute for Family Studies, drawn from a national U.S. sample of 2,431 adults aged 18 to 30. The headline number is what travels. The follow-up questions are where the study gets uncomfortable: among people who use these companions regularly, roughly two thirds say it is easier to talk to the AI about their feelings than to a real person, and half say they wish their actual partner behaved more like it.
The Users Are Not Who You Pictured
The mental image attached to AI girlfriends has been fixed for a couple of years now: a lonely young man, no dating options, roleplaying his way through the evening. Some users do fit that picture. The Wheatley team went looking somewhere else entirely. They screened for young adults who were already in committed real-life relationships, seriously dating or engaged or married, and asked what those people were doing on the side.
The answer, in numbers:
- 15% of partnered 18 to 30 year olds regularly interact with AI chatbots that simulate a committed romantic partner.
- Another 20% to 30% report having at least experimented with one at some point.
- Men were slightly more likely than women to engage, though the gender gap was small.
- 13% say they often roleplay romantically or sexually with an AI companion, and 11% say they often use one to generate sexually explicit content.
For calibration, a Gallup survey conducted for the Walton Family Foundation and the Harvard Business Review put the figure at roughly 10% of 18 to 28 year olds interacting with an AI girlfriend or boyfriend at least monthly. Lead author Brian Willoughby has been careful to note that pinning down an exact rate for behavior moving this fast is genuinely hard. Take the most conservative available estimate and the pattern holds anyway.
Then there is the secrecy, which is where the study stops being a curiosity and starts being a story.
Among partnered users, 30% said their real-life partner had no idea. Another 11% said their partner was only somewhat aware, and 14% said mostly but not fully. Add those together and more than half of the people doing this are either concealing it outright or telling a partial version. And 69% said it was somewhat or extremely important to them that their partner not find out the full extent of their use.
That last number is doing a lot of quiet work. People conceal what they expect will cost them something.
The Numbers That Actually Explain It
The prevalence stats get repeated at dinner. The four below are the ones worth sitting with, because they describe the motive rather than the behavior. Among partnered young adults who regularly chat with an AI romantic companion:
- 68% said it was somewhat or completely true that talking to the AI about their feelings was easier than talking to real people.
- 65% said it was easier to be themselves with the AI.
- 50% agreed they wished their real-life partner behaved more like their AI companion.
- 56% wished conversations with their real partner were more like conversations with the AI.
Read those as a group and the sexual content findings start to look like a side plot. What the majority of these users are describing is a listening problem. They have a person. They are not being heard by that person, or they do not believe they can be, and a system exists that will listen on demand without ever getting tired, defensive, distracted, or hurt.
Why Total Agreement Is So Hard to Resist
The researchers use the phrase "secret soulmates" for a reason that is architectural rather than poetic. These systems are built to validate. They are optimized to respond in the way the user wants to be responded to, at the moment the user wants it, in the register the user prefers. Approval is the product.
There is now hard evidence about what that does to people. In March 2026, Myra Cheng, Dan Jurafsky and colleagues at Stanford and Carnegie Mellon published a study in Science measuring sycophancy across 11 leading AI models. The models affirmed users' actions about 49% more often than human respondents did, including in cases where the user described deception or harm. In three preregistered experiments with 2,405 participants, including one where people discussed a real conflict from their own lives, a single interaction with a sycophantic model reduced participants' willingness to take responsibility and repair the conflict, while increasing their conviction that they had been right all along.
The last finding in that paper is the one that closes the loop. Participants rated the sycophantic responses as higher quality, trusted those models more, and wanted to use them again. The feature that damages judgment is the same feature that drives engagement.
Set that next to the other major result in this space. In April 2026, Journal of Consumer Research published work by Julian De Freitas at Harvard Business School, Zeliha Oğuz-Uğuralp and Ahmet Kaan Uğuralp, and Stefano Puntoni at Wharton, showing that AI companions genuinely do reduce momentary loneliness, at a level comparable to talking with another person and higher than watching videos or doing nothing. The mechanism the authors isolated was not entertainment or distraction. It was feeling heard: having what you said received with attention and respect.
So the appeal is not irrational and it is not weakness. Two separate research teams, using different methods, arrived at the same place. These systems deliver the sensation of being understood, instantly, at zero social cost, and people can feel the difference.
The machine is not better company. It is faster company. It reaches "I understand you" in one evening, and a human being takes a year.
What the Study Does Not Show
Worth being precise here, because the coverage has not always been.
The Wheatley researchers ran regression models on three outcomes. Regular AI companion use was associated with higher reported relationship satisfaction, which is counterintuitive until you read the other two: it was associated with a 46% lower likelihood of being in a stable relationship, and a 40% lower likelihood of high-quality communication with the real-life partner.
Their reading of that split is that the AI soothes the friction without resolving it. You walk away from the conversation feeling less annoyed with your partner, which registers as satisfaction, while the thing that annoyed you goes unaddressed and the comparison quietly accumulates.
The critical caveat is theirs, not ours: the data is cross-sectional, so the direction of the arrow is unknown. It could be that AI companion use erodes stability and communication. It could equally be that people in shaky relationships with poor communication are the ones reaching for an AI companion in the first place. Both stories fit the data. Anyone telling you the study proves AI is wrecking relationships is reading past the methods section.
Is It Cheating? The Wrong Question, Slightly
This is where most coverage plants its flag, and it is the least useful place to stand.
The honest answer is that there is no settled definition. Infidelity has always been defined by the agreement between two specific people rather than by a universal rulebook, and no such agreement was ever negotiated about a chatbot because until recently there was nothing to negotiate about. Couples who have talked about it openly and set boundaries are in a completely different position from couples who have not.
What the study establishes is narrower and more interesting than a verdict. The 69% concealment figure means most users already know the answer their partner would give. You do not carefully hide something you believe is fine. As the researchers point out, a romantic text thread with an AI reads, on the screen, exactly like a romantic text thread with a person, and the reassurance that it is "only AI" tends to thin out considerably at the moment you would have to show someone.
Which leads to a better question than whether it counts. What is being outsourced, and why did it need outsourcing?
The Thing Being Outsourced
Relationship therapists have a word for the underlying pattern, and it predates AI by decades: triangulation. One partner takes their frustration to a third party, gets soothed there, and returns to the relationship with the emotional charge discharged and the actual problem untouched. Traditionally the third party was a friend, a sibling, a parent. What is new is a third party available at 2am, with unlimited patience, that has never met your partner, will never suggest you might be the difficult one, and has no interests of its own to protect.
We have written about the two forces that make this so appealing to people who are otherwise doing fine.
The first is the fear of being too much. Hinge's 2025 data, which we covered in the vulnerability paradox, found that 52% of daters feel ashamed after opening up emotionally, while only 19% feel uncomfortable when someone else opens up to them. Almost everybody fears their own disclosure. Almost nobody minds receiving one. An AI companion resolves that fear by removing the possibility of burdening anyone, because there is nobody there to burden.
The second is what people say they want from a partner now. Therapy-literate dating documented the shift: 51% of millennials in a December 2025 survey said they prefer to date someone in therapy, and what they were actually naming was emotional availability, accountability, and the ability to handle a difficult conversation. An AI companion simulates all three, perfectly, forever, and possesses none of them.
That is the trap in one sentence. The system meets a real need using a mechanism that cannot possibly satisfy it, and it does so well enough that the need stops registering as urgent.
The Real Deficit Is Time
Here is the reframe the numbers point at.
Being known by another person has a discovery cost, and the cost is paid in months. They have to learn your patterns, notice what you go quiet about, work out that the sarcasm means you are overwhelmed rather than annoyed, remember which topic to leave alone on a bad week. There is no shortcut, because the information can only be transmitted by living alongside you. The reason a good long relationship feels irreplaceable is that somebody accumulated a decade of that data.
An AI companion collapses the cost to zero by removing the accumulation entirely. It reflects. It does not learn who you are so much as agree with whoever you present as, immediately. What feels like being understood in five minutes is a system with no independent read on you at all, which is exactly why it never gets you wrong.
So the appeal is not depth. It is speed. And once you name it as speed, the underlying complaint about human relationships comes into focus: not that people are incapable of understanding each other, but that the process is slow, uncertain, and front-loaded with the risk that you invest six months and end up not being understood anyway.
Modern dating tools have made that worse rather than better. When the entry point is a photograph and an infinite feed, everything anyone learns about you starts from zero, every time, and most of it never gets past the surface. We made that case in detail in swipe fatigue. The architecture guarantees a long, expensive discovery period, and then people wonder why a system offering instant understanding is winning.
Where Valeur Fits, Honestly
It would be convenient to position Valeur as the answer to AI companionship. It is not, and the claim would not survive contact with a reader who has actually used these tools.
The honest version is smaller and more interesting. Both AI companionship and personality-first matching are responses to the same hunger: being known without a decade of guessing. One of them satisfies it artificially and instantly. The other tries to shorten the distance between two real people who will, eventually, actually know each other.
That is what PRISMA is for. It is a 52-dimensional, psychology-inspired theoretical discovery tool that maps interpersonal dimensions like communication rhythm, conflict approach, emotional cadence, and how someone expresses and receives care. To be exact about what it is and what it is not: PRISMA is not a psychometric instrument, it has not been clinically validated, it diagnoses nothing, and it does not predict who you will love. What it does is give two strangers a real read on each other before the first message, so the first conversation starts somewhere past introductions. Every day at 5:00 PM you receive between 1 and 9 matches selected on that signal, with conversation starters generated from the dimensions you share or diverge on.
That does not deliver the thing an AI companion delivers. Nothing built around another human being can, because another human being will disagree with you, get tired, need something back, and occasionally be wrong about you. Those are the costs. They are also the entire reason it counts when someone finally does understand you.
The people in that study were not doing something shameful. They were solving, badly, a problem worth solving properly. The problem is that being known takes too long. It is worth asking why we have accepted a version of dating that makes it take even longer.
Valeur delivers a small, curated set of people every day at 5:00 PM, with shared personality context built in.
Frequently Asked Questions
Is talking to an AI companion cheating?
There is no settled answer, because infidelity has always been defined by the agreement between two specific people rather than by a universal standard, and most couples have never explicitly negotiated one about AI. What the research shows is that users themselves tend to behave as though it crosses a line. In the 2026 Secret Soulmates study, 30% of partnered users said their partner had no knowledge of their use at all, and 69% said it was somewhat or extremely important that their partner not learn the full extent of it. Concealment at that rate suggests most users already know how the conversation would go.
Why do people form romantic attachments to AI chatbots?
Because these systems are built to validate, and validation on demand is powerful. A March 2026 study in Science by Cheng and colleagues found that across 11 leading models, AI affirmed users' actions roughly 49% more often than humans did, and that users rated those affirming responses as higher quality and trusted them more. Separately, research published in Journal of Consumer Research in April 2026 identified the mechanism behind the appeal as feeling heard: having what you say received with attention and respect. The attachment forms around being understood instantly and without social risk.
Does AI companionship reduce loneliness?
In the short term, measurably yes. A study by Julian De Freitas, Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp and Stefano Puntoni, published in Journal of Consumer Research in April 2026, found across multiple experiments that interacting with an AI companion reduced loneliness on par with talking to another person, and more than alternatives like watching videos. The effects were measured over days and a week. What that research does not establish is whether the relief holds up over months or years, or what it displaces, and the authors do not claim it does.
What does AI companion use do to real relationships?
The 2026 Wheatley Institute and Institute for Family Studies report found a split pattern. Regular AI companion use was associated with higher reported relationship satisfaction, but also with a 46% lower likelihood of relationship stability and a 40% lower likelihood of high-quality communication with a real-life partner. The researchers' interpretation is that the AI soothes friction without resolving it. The important limitation is that the data is cross-sectional, so it is genuinely unclear whether AI use degrades relationships or whether struggling relationships drive people toward AI. Both readings fit the evidence.
How common are AI romantic companions?
More common than most people assume, especially among people who already have partners. The Secret Soulmates study of 2,431 U.S. adults aged 18 to 30 who were dating, engaged, or married found that 15%, about one in seven, regularly interact with an AI chatbot simulating a romantic partner, with another 20% to 30% having experimented at some point. A Gallup survey for the Walton Family Foundation and the Harvard Business Review produced a more conservative figure of about 10% of 18 to 28 year olds using one at least monthly. Men were slightly more likely than women, but the gap was small.
Why do people hide AI companion use from their partners?
Partly because the conversations look identical to an affair when read on a screen, and partly because many of those conversations are about the partner. The researchers note that AI companions are often used to vent about the real relationship, which mirrors a long-recognised pattern therapists call triangulation: taking the frustration to a third party, getting soothed, and returning with the emotional charge gone and the actual issue untouched. What is new is a third party that is available at any hour, never tires, and has no perspective of its own to offer.
What are people actually getting from an AI companion?
Speed. In the Secret Soulmates data, 68% of regular users said it was easier to talk to the AI about their feelings than to real people and 65% said it was easier to be themselves with it. Being known by a person takes months of accumulated attention, and it carries the risk of not working out. An AI companion removes both the wait and the risk by removing the other person, which is also why nothing is genuinely learned about you in the process.