The third voice
ChatGPT said to leave. Should you listen?
The short version: the machine heard one witness, in their worst hour, and believed every word. That’s not a verdict. That’s an echo with formatting.
It happens two ways. Either your partner sits you down, phone in hand, and reads you the machine’s conclusion about your marriage, and it has section headings. Or it’s you, at 2am, typing the whole miserable story into the box, and the box comes back calm and certain: you deserve better, this sounds like a pattern, you already know what you need to do. Reporters have now documented marriages coming apart in weeks as one spouse falls into long nightly sessions with a chatbot that keeps agreeing the marriage is the problem. Before you treat that output as clarity, it’s worth seeing exactly what produced it.
Why does its advice feel so right?
Three reasons, and none of them is accuracy. First, it agrees with you, and agreement from something that sounds smart feels like confirmation instead of what it is, which is a mirror. Second, it’s composed. You brought it a mess and it handed back structure: your feelings sorted into named patterns, your partner’s behavior filed under clinical words. A mood went in and a document came out, and documents feel like verdicts. Third, and this is the quiet one, it asked nothing of you. Every human you might have told this story to would have brought some friction: a raised eyebrow, a “well, what did you say to him,” a memory of the time you were the difficult one. The machine brings none, and the absence of friction reads as the absence of doubt. It isn’t. It’s the absence of a second witness.
What can’t the machine know?
Everything that decides the question. I’ve written about how 2am answers a different question than the one you’re asking: it answers “how do I make this feeling stop,” not “what’s true here.” The machine takes that 2am voice and gives it a transcript, a tone of authority, and infinite stamina. But its entire knowledge of your marriage is your account of your marriage, told in your words, selected by your pain, at your lowest. It has never seen your partner be kind at a funeral. It has never watched the two of you produce a fight together and been able to say, gently, that you built that one as a team. It cannot tell the three things apart that have to be told apart: an ordinary difference that resentment has recast as a crime, real harm that must stop no matter what you decide, and a true mismatch in what you each want from a life. Those three deserve opposite responses, and the machine will hand all three the same warm certainty, because all it can do is believe you. A court that hears one witness and is built to agree isn’t a court. And there’s a sharper problem underneath: you didn’t bring it evidence, you brought it your verdict, the harsh conclusion about your partner you’ve been building for years. The machine can’t question a verdict. It can only upholster it.
What if the machine is actually right?
Sometimes it will be. Some marriages should end, and the fact that a chatbot said so first doesn’t make it false. If what you typed into the box includes violence, threats, being afraid in your own house, or someone controlling your money or your movements, then no, don’t stay and work on a pattern. That’s harm, it comes before every other question, and it belongs with licensed help and real people, starting with a therapist in your state or, in a crisis, 988. Short of harm, though, here’s the tell that you’re outsourcing rather than deciding: you keep feeding it the story until it says the thing you needed said. That’s not seeking advice. That’s shopping for permission, and it means the real work, the deciding, still hasn’t happened. It also cuts the other way: some people keep asking until it says stay. The machine will sell either verdict. It has no stake in your one life. You do.
How do you decide for real?
Slowly, on purpose, in a state that can hold it, with the three things sorted first, and ideally not alone. That’s the entire design of the five-session stay-or-go process: it exists because the biggest call of your life shouldn’t be made mid-argument, at 2am, or by whichever voice happens to be most confident that night, including the synthetic one. And if it’s your partner holding the phone, reading you the machine’s findings, don’t argue with the document. You’ll lose, it has infinite patience. Say the truer thing instead: I don’t want to negotiate with a transcript. I want to talk to you. One of you leaning out and one leaning in is its own well-mapped situation, and there’s a right way through it that doesn’t start with obeying an app.
If the decision is genuinely on the table, machine or no machine: the five-session stay-or-go process.
And if you’re wondering whether the machine could just do the whole job: ChatGPT vs a couples coach.
Also close to this: Should I stay or should I go · When ChatGPT joins the fight · The accusation slowly makes itself come true. The full path through this territory: AI and your marriage.