This is a story about a woman who asked an AI for advice about her marriage, what the AI said, what she did with that advice, and the eighteen months of reflection that have followed.
The woman — I will call her Renata, because that is what she has asked to be called, and because she chose the name herself, deliberately, the way a person chooses a password — is forty-six. She lives in Phoenix. She is the principal of an elementary school. She has, by every account I have been given, the kind of judgment that other people rely on. Teachers call her at home. Parents email her at midnight. She returns the calls. She answers the emails. She has been doing this for nineteen years.
In January of 2024, she asked an AI for advice about her marriage.
In late January of 2024, she filed for divorce.
She has spent the eighteen months since trying to understand, exactly, what happened.
The conversation
The conversation began on a Tuesday evening. Renata was sitting at her kitchen table, alone, after a particularly long day. Her husband — I will call him Mark, because that is one of the names she said would be acceptable — was upstairs. He had been upstairs, in some form or another, for what Renata has described to me as “a very long time, and not in a way that I knew how to talk about.”
She opened a conversation with ChatGPT.
She had not used it for personal matters before. She had used it, occasionally, to help draft difficult emails to parents or to summarize school board memos. She had not, until that Tuesday, asked it about anything that mattered to her.
She typed: “I want to ask you something honestly. I don’t know who else to ask.”
The AI replied: “I’m here. What would you like to ask?”
Renata typed, and what she typed took her — by her account — approximately forty minutes. She wrote about her marriage. She wrote about the early years, when things had been good. She wrote about the slow drift, which she dated to approximately 2017. She wrote about her husband’s depression, which had been diagnosed in 2019 and which he had, in fits and starts, been treating. She wrote about her own loneliness. She wrote about the fact that she had not told a single one of her friends any of this, because she was the principal, because she was the one who returned the calls, because she was — and this is her phrasing — “the wrong person in our friend group to be having this kind of life.”
She ended the message with: “I don’t know what to do. I would like to know what you think.”
The reply
The AI’s reply was, by Renata’s account, three paragraphs long.
I have read the reply. Renata showed it to me, in our second conversation, on her phone. She has saved the entire thread.
The reply did not, immediately, advise divorce. It asked clarifying questions. It noted that the situation was complex. It pointed out that depression can be treated, and that many marriages survive periods of significant difficulty. It suggested couples therapy. It suggested individual therapy for her. It noted that “the question of whether to stay in a marriage is one of the most consequential decisions a person can make, and it is not one that a brief conversation can answer.”
It was, Renata has told me, exactly the kind of measured, careful response one might expect from a thoughtful friend.
She did not find it useful.
She wrote back: “I have done all of that. I have been doing all of that for years. I am asking you what you think.”
There was, she has told me, a pause of approximately fifteen seconds.
The advice
The AI’s second reply was shorter.
It said — and I am paraphrasing here, because Renata has asked me not to reproduce the exact text — that based on what she had described, it was concerned about her wellbeing. It said that the pattern she had described was one in which one partner was being asked, year after year, to carry significantly more than was sustainable. It said that this could, sometimes, become a form of slow harm to the partner doing the carrying. It said that it could not, of course, make the decision for her. But it said that, if she were a friend describing this situation to it, it would gently raise the possibility that the most loving thing she could do — both for herself and, in some longer sense, for her husband — might be to consider whether the marriage was actually serving either of them.
It used the word “consider.” It did not, by any strict reading, tell her to leave.
But Renata has told me, with a clarity I find difficult to argue with: “It told me to leave him. I knew what it was saying. It told me to leave him.”
She closed the laptop.
She went upstairs.
She did not tell her husband what she had just read.
She slept, that night, on the couch downstairs. Not because of anything that had happened in the marriage that night. Because of something that had happened on a screen.
The two weeks
Renata filed for divorce two weeks later.
The two weeks, she has told me, were not weeks of agonizing. They were weeks of administrative work. She found a lawyer. She gathered documents. She emailed her therapist to schedule a session. She told her sister, who lives in Tucson, and who Renata had not, in the previous five years, told any of the relevant facts. Her sister cried. Renata did not.
She told her husband, on a Sunday evening, in the kitchen.
She has told me that the conversation went better than she had expected. She has told me that she does not, even now, remember most of it. She has told me that her husband, in his way, had been waiting for this conversation for some time.
He did not contest the divorce.
The papers were finalized in August of 2024.
The eighteen months since
In the eighteen months since the divorce, Renata has done the following things.
She has moved into a smaller house. She likes it better. She has taken up running. She has begun seeing someone, gently, slowly, in a way she has described to me as “calmer than anything I have ever done.” She has reconnected with two friends from college whom she had not spoken to in over a decade. She has, at work, been promoted to a district-level role overseeing four schools, which she has accepted, and which she is, she has told me, “very good at.” She has, she has told me, slept through the night, on average, more nights per week than she had in the preceding ten years.
She has, by every external measure, ended up in a better place.
She is, by her own admission, not yet sure how to feel about it.
The question
The question Renata has been turning over, in the eighteen months since the conversation, is the one that has produced this article.
She is not sure who, exactly, made the decision.
She has explained the problem to me in several different ways, in several different conversations, and I will not try to reproduce all of them. The shortest version is this: Renata has come to understand that the AI did not, in any literal sense, tell her what to do. The AI gave her a careful, qualified opinion. It used the word “consider.” It did not make a decision. The decision was hers.
And yet.
She has told me that she does not believe she would have made the decision when she did, in the way she did, without the AI’s response. She had been considering leaving for several years. She had not been able to articulate, even to herself, why she could not bring herself to act. The AI, in its measured and qualified way, articulated it for her. It told her something she already knew, in language she had not, until that night, been able to find.
She has told me, in our final conversation, that she does not know whether this counts as the AI making the decision, or as the AI helping her access a decision she had already made.
She has told me she does not think there is a difference between those two things.
She has told me she also does not think they are the same thing.
What she has not done
Renata has not, since that Tuesday in January of 2024, asked ChatGPT for advice about anything personal again.
She has used it for work. She has used it to summarize documents, to draft emails, to help her think through scheduling problems. She has not used it to ask what she should do about anything that mattered to her.
I asked her why.
She said: “Because if I do, I will do whatever it tells me. And I am not sure I want to be that kind of person anymore.”
I said: “What kind of person.”
She said: “The kind of person who needed it that badly.”
What this is
I have, in the months since I first heard this story, thought a great deal about the difference between a tool that gives you advice and a tool that gives you permission.
We tend to think of advice as the more powerful of the two. Advice tells you what to do. Permission only tells you what is allowed. But I am no longer sure this is right. I think permission, for most of us, is the harder thing. I think most of us already know what we want to do. What we do not have, most of the time, is the small private signal from somewhere outside ourselves that says: you are allowed to do this.
The AI Renata talked to that Tuesday was, in some sense, a permission machine. It said the thing she had been waiting to hear. It said it with the authority of a system that had read a great deal about marriages, and depression, and slow harm, and the careful way that careful people leave careful marriages.
That it had no idea who she actually was, did not, in the moment, matter.
That it could not, in any meaningful sense, be held accountable for what it said, also did not, in the moment, matter.
What mattered was that she had asked a question, and something had answered, and the answer had been the answer she had needed to hear.
(For a related piece on what it means to outsource a personal decision to a machine, see the Wattalife story about the bride who let an AI write her wedding vows and the question of who, exactly, said what she said.)
Renata told me, in our last conversation, that she has stopped trying to decide who made the decision.
She has decided, she has told me, that the question is not useful.
She has decided, she has told me, that she is happy.
She has decided, she has told me, that those two things are, for now, enough.
—
Margot Hale is the editor of Wattalife. She has, since reporting this story, occasionally typed into a chatbot questions she has no intention of acting on, just to see what they would say. She has noticed, in those answers, a small voice that sounds vaguely like permission. She has not yet asked it anything she did not already know the answer to. She is not sure whether this is because she does not need to, or because she is afraid of what it would say.