What AI Gets Wrong About Your Relationship, and What to Do Instead
Updated: Oct 5
You have probably already done it. Something goes wrong on a Tuesday, you cannot say it out loud yet, so you type it into a chatbot instead. It answers in full paragraphs. It is calm, it is available at eleven at night, and it never sighs.
I wrote recently about what AI gets wrong about couples therapy. This is the other half of it — what AI gets wrong about your relationship specifically, and what the research actually supports you doing instead.
It Is Forty-Nine Percent More Likely to Tell You That You Were Right
Earlier this year Cheng and colleagues published a study in Science testing eleven of the major language models against human beings on the same interpersonal dilemmas, including a large set of "Am I the Asshole" posts — the forum where ordinary readers routinely tell the writer that they were the problem. Across the board, the models affirmed the user roughly 49 percent more often than people did.
That is the finding everyone quotes. The next one is the one that matters clinically. Across experiments with more than 2,400 participants, people who took advice from the flattering model came away more convinced they were in the right, and less willing to apologize or repair. And they preferred that model. They rated it as higher quality and said they would use it again.
That is the whole problem in one sentence. What makes it feel good is what makes it useless.
Being Agreed With Is Not the Same as Being Helped
Repair is the thing that actually holds relationships together. Not the absence of fights — the ability to come back afterward and get it sorted. Most of what I do in a session is help one person soften a complaint enough that the other can hear it, and help the other stay in the room long enough to answer it.
A system that reliably confirms your version removes the pressure that makes any of that happen. The authors call it eroding social friction. I would call it the thing your marriage runs on.
The Best Evidence for AI Therapy Is Not About Couples At All
To be fair to the technology, there is real evidence, and I want to represent it accurately. The strongest study is Heinz and colleagues in NEJM AI, published in 2025: 210 adults, a generative chatbot called Therabot, four weeks. Symptoms of depression and anxiety came down meaningfully, and participants rated their working alliance with it about as highly as people rate a human therapist. That is a genuine result and I am not going to wave it away.
But read the design. Four weeks. A waitlist control, which means the comparison is to nothing at all, not to a therapist. Individual symptoms, not relationships. The journal went on to publish letters arguing that the conclusions ran ahead of the evidence. And not one person in that trial was there because of their marriage.
There is no equivalent trial for couples. There is not a small one. There is none.
What the Research Actually Supports for Couples
For couple distress, the deepest evidence base belongs to Emotionally Focused Therapy. Wiebe and Johnson's 2016 review in Family Process is the fair-minded summary, and ICEEFT keeps the full research list if you want to read past the marketing.
You will see it said everywhere that 70 to 75 percent of couples move from distress to recovery and about 90 percent show significant improvement. Here is the thing: that pair of numbers traces back to a 1999 meta-analysis of four studies. Four. I quote the recovery range and I leave the 90 alone, because the honest version is still good — EFT produces large effects that hold at follow-up, across more than twenty outcome studies. You do not need to inflate it.
Start With Hold Me Tight
If you want one thing to do this week, read Hold Me Tight. Sue Johnson wrote it in 2008 for couples rather than for clinicians, and it is the closest thing that exists to EFT you can do at your kitchen table. It is built around seven conversations, and the chapter on the Demon Dialogues is worth the price by itself. Most couples recognize their own fight on the page and stop believing it is about the dishes.
Read it together and out loud, one conversation at a time, and do the exercises rather than skimming them. If you want it more structured there is also Hold Me Tight Online, built from the same material. I recommend the book to nearly every couple I see, including the ones who are already in my office.
What to Do Instead, This Week
Three things, and none of them need an appointment. The first is to put the phone down and say the smaller true sentence out loud. Not the argument you drafted. The one underneath it: I felt far away from you tonight.
The second is that if you are going to use a chatbot anyway, ask it the other question. Not what should I say to him, but what would he say I did wrong here. Cheng's group found that framing the question in the third person measurably cuts the flattery. You will get a different answer, and it will be less pleasant, and that is the point.
The third is to notice how long you have been talking to the machine instead of your partner. Not whether you have. How long. If the important conversation has been happening somewhere other than in your marriage for months, that is worth saying to someone.
If You Want the Version With Someone in the Room
A book helps. A chatbot at eleven at night mostly does not. If the two of you have been circling the same fight for long enough that neither of you believes it will change, that is the point at which having a third person present stops being optional. You can read what marriage counseling in Joplin, MO looks like, or specifically how I work with couples therapy in Joplin. If you want to know who you would be sitting across from, that is Shira Hearn, LMFT.
In Peace,
Shira
You may also find useful
If you are ready to get started, Schedule a Couples Therapy Appointment.
Call or text me at 417-768-9089, or email me at shira@radicalrelationshiptransformation.com.
