You have had the same argument for months. Couples therapy feels out of reach, or one of you will not go, and an app or a chatbot is right there on your phone. Before you trust one with something this personal, it is fair to ask whether any of it has been tested, and what the tests found.

The evidence is early. Couples therapy itself is well supported, and online couples programs have done well in randomized studies. The research on AI help is small and recent. In a randomized study of 258 people, one session with a relationship chatbot worked about as well as a writing exercise, and the follow-up lasted two weeks. Whether AI help makes a difference that lasts for months is not yet known.

If you are deciding whether AI help is right for the two of you, start with our guide to AI couples therapy. This page is the research behind it.

Three kinds of evidence, and how much each one tells you

The studies that bear on this question fall into three groups, and they carry very different weight.

Couples therapy studies. The most tested of the three, with one 2020 analysis pooling 58 studies. None of it involves AI, but it sets the bar.
Online couples programs. Randomized studies with hundreds of couples, testing whether couples skills can be taught through a screen. These involve no AI either.
Studies of AI. Recent and narrow: one randomized study of a single chatbot session, a study of how chatbot replies were rated, an interview study, a test of general AI models and a study of AI-suggested replies. None of them followed people for months.

Couples therapy works, which sets the bar

Any AI tool should be measured against couples therapy with a trained professional, because that is where the evidence is strongest. A 2020 analysis of 58 studies covering 2,092 couples (Roddy and colleagues, 2020) found large effects on relationship satisfaction, and the gains generally held at later follow-ups. Couples who started out more distressed tended to gain more.

The same analysis found that couples on waitlists did not significantly improve, which is worth knowing if you are hoping things will settle on their own. If cost is what keeps you from a therapist, here are the cheaper routes to a licensed therapist.

Online couples programs, tested against waitlists

Before today's AI chatbots, researchers tested whether couples could learn the skills of couples therapy online. In a randomized study of 300 couples (Doss and colleagues, 2016), an eight-hour online program adapted from an in-person form of couples therapy, plus four 15-minute calls with staff, improved relationship satisfaction compared with couples on a two-month waitlist.

A larger nationwide study of 742 low-income couples (Doss and colleagues, 2020) tested two online programs against a waitlist. Both improved relationship functioning, and the gains held for the four months after the programs ended. Neither program significantly reduced breakups by six months.

These were structured programs with human staff on the calls, not chatbots. They show that structured couples help can work online, though not that a conversation with an AI does the same.

The AI studies so far

Research on AI help for relationships is recent. The four studies below were published in 2024 and 2025, and each looks at one small piece of the question.

One chatbot session, compared with a writing exercise

In a randomized study of 258 people (Vowels and colleagues, 2025), each person had a single session, either with a chatbot built for relationship support or with an evidence-based writing task. Both groups improved on 13 of 14 measures, with no significant difference between the groups on most. The measures included relationship satisfaction, communication patterns, coping together, confidence about the problem and individual well-being. Most gains showed up right after the session and held or kept improving at the two-week follow-up; a few, including individual well-being and coping together, appeared only by the follow-up.

The groups differed on two points. Right after the session, the chatbot group reported less of the pattern where their partner presses and they pull back, a difference that had narrowed and was no longer significant by the two-week follow-up. The writing group reported lower distress about the issue at every point, including before the session, so that gap was not an effect of the writing task.

The fair reading is that one chatbot session did about as well as an evidence-based writing exercise. That result comes from a single session with one chatbot and a short follow-up. The researchers also note that the chatbot sometimes gave repetitive or generic replies and did not consistently identify potential safety concerns, and they call for studies with more than one session.

Replies that read like a counselor's

In a blind comparison (Hatch and colleagues, 2025), 13 therapists and a general chatbot wrote replies to 18 couples therapy scenarios. When 830 people judged the replies, they could rarely tell which ones the chatbot had written, and on key principles of psychotherapy they generally rated the chatbot's replies higher.

The study tested how people rated written replies. It did not test whether those replies would help a couple, and the authors name the missing therapy context as one of its limits.

A general chatbot, and risk

In an interview study (Vowels and colleagues, 2024), 20 people each tried a general chatbot for a single session about their relationship. They found it empathic and clarifying, but the researchers judged the chatbot's assessment of risk to be poor.

A study of general AI models (Moore and colleagues, 2025) found related problems. The models showed stigma toward people with mental health conditions and gave inappropriate answers in some serious situations, and this held even for larger and newer models. The authors point to sycophancy as a likely part of the reason for the inappropriate answers: the models lean toward agreeing with the person they are talking to. They concluded that these models should not replace therapists.

AI-suggested wording between two people

AI that shapes the words people send each other has been tested too. In two randomized experiments with pairs of strangers chatting online about a policy issue (Hohenstein and colleagues, 2023), AI-suggested replies made the language more positive, and the two people rated each other as closer and more cooperative. But people suspected of using the suggestions were judged more negatively.

The pairs were strangers, not couples, so the study cannot say how AI-shaped wording affects trust between partners. That is still an open question, and it applies to any tool that changes the words between you.

What the American Psychological Association says

The American Psychological Association (APA) set out its view in a 2025 health advisory on AI chatbots and wellness apps. In its announcement, the APA said these tools "currently lack the scientific evidence and the necessary regulations to ensure users' safety." Its key recommendations include this one: "do not use chatbots and wellness apps as a substitute for care from a qualified mental health professional." The advisory also calls on policymakers to promote "legislation that prohibits AI chatbots from posing as licensed professionals."

The studies above point the same way. Some early results are encouraging, and none of them shows that an app can stand in for licensed help.

What the research can't tell you yet

What happens over months. The randomized chatbot study followed people for two weeks, and the other AI studies here looked at a single session or none.
What happens when both partners use a tool together. The AI studies here looked at individuals, people rating replies, strangers or the models themselves, not a couple using one tool together.
Whether a given app works. Each study tested a particular chatbot, general AI models or a reply-suggestion feature, so a result for one tool does not automatically carry over to another.

How to read a claim that an AI app works

If an app says it is proven or research-backed, four questions help:

Compared with what? Doing nothing, a writing exercise and a therapist are very different bars.
How many people, and for how long? Look for results weeks or months later, not only right after one session.
What was measured? How replies were rated is a different question from whether couples did better.
Was it this app? Studies of couples therapy or online programs are about those, not about the app that cites them.

For what these apps do day to day, see what conflict resolution apps actually do.

If you are weighing whether to try it

Three parts of our guide help with that decision:

The five kinds of AI couples help, which differ in who the AI hears and when it helps.
The five guardrails, if you already use a general chatbot together.
The eight questions for comparing apps.

The guide also covers whether AI is transforming couples therapy, including what to ask if your therapist uses AI. For options that involve no AI at all, see alternatives to couples counseling.

If you need help now

If anyone is in immediate danger, call 911.
If there is violence, fear or coercion in your relationship, call the National Domestic Violence Hotline at 1-800-799-7233 or text START to 88788.
If you or your partner is thinking about suicide or self-harm, call or text 988.
For addiction, trauma or a separation already in motion, find a licensed couples therapist through TherapistLocator.net, the directory of the American Association for Marriage and Family Therapy, or the Psychology Today directory.

Where Harmony fits

We build Harmony: AI Relationship Coach, an iPhone app for coaching and communication support, so weigh this section accordingly. None of the studies above tested Harmony, and we do not claim they show it works.

Couples sessions happen inside Harmony's own chat or voice, not iMessage or SMS. In Couples Chat, Harmony checks each message before it reaches your partner:

A clean message passes through untouched.
Escalating wording is softened in transit.
A contemptuous message is held back, and the sender gets a private coaching card instead.

In Couples Chat only the sender is coached, privately on their own screen, and your partner receives only the delivered message. In Couples Voice, the listening partner also gets a private "how to respond" card on their own screen. When a session ends, each of you gets your own summary and follow-up experiments. Solo Chat and Solo Voice are private coaching for one partner.

Harmony is for couples who mostly need the same fight to stop repeating, partners who want help finding the words, and one partner working on things alone. It is not the right place to start if there is violence, fear or coercion, addiction, self-harm or a crisis, or a legal separation already under way. It is coaching and communication support, never a replacement for therapy.

AI coaching is paid, and one Couples subscription covers both partners, so the partner you invite never buys anything. Any current introductory offer is shown at checkout. See current plans and terms.