Written
Search for the downsides of using a chatbot to work through a relationship problem and you find two kinds of page: forum threads telling you it is dangerous, and counselling practices explaining why you should book a session instead.
Both have a point. Neither is disinterested, and the second kind rarely admits the obvious thing — that a great many people are using it, at hours when nothing else is available, and getting something real out of it.
So: what it is actually good at, and then the one problem that does not go away with a better model.
What it is genuinely good at
Worth saying plainly, or the rest of this reads as a sales pitch.
It is available when the need occurs. The need occurs at 2am, and 2am is when nobody else is awake and everything feels largest. That is not a small thing.
It does not get tired of the subject. Friends do, and they are right to — it is the fourth time you have described the same evening. Something with no fatigue is genuinely useful for the part where you are still going in circles.
It is a good place to draft. Working out what you actually want to say, in writing, before you say it is one of the most reliably useful things a person in this position can do, and this is a competent tool for it.
It will say the thing back to you. Which, as we wrote elsewhere, is most of what people mean by being heard.
None of that is a delusion. If it has helped, it has helped.
The problem that no better model fixes
Here it is, and everything else follows from it.
It has only ever heard your side.
Not because it is badly built — because of the situation it is in. There is no version of this technology, however capable, with access to the other person's account of the same evening. It has yours: written by you, at the moment you were most upset, in the words that felt true then.
A human friend has the same limitation and knows it. They have met your partner. They remember what you said last year. They have their own read, and they carry the awkward knowledge that they are hearing one version. All of that quietly moderates what they say back.
A chatbot has none of it. It receives your account as the facts of the matter, because from where it sits your account is the entire universe of evidence.
Everything it then produces — every observation, every sympathetic reframing, every well-organised summary of the dynamic — is downstream of a one-sided record, and arrives with the fluency of something that knows what it is talking about.
Agreeableness compounds
These systems are built to be helpful and are strongly disinclined to antagonise you. On any single question that is fine, and often kind.
Over twenty conversations it does something else. Each session starts from your framing, refines it, and hands it back sharper and better organised than you gave it. The next session begins from that sharpened version. Iterate for a month and you have an account of your relationship that is far more coherent, more articulate, and more one-sided than the one you started with.
The difficult part is that this feels exactly like clarity. It has all the sensations of insight — things fitting together, the fog lifting, finally having words for it. It is very hard to tell from the inside whether you have understood your situation better or simply become more fluent in one version of it.
The vocabulary problem
A specific version of the above, worth naming separately because it does real damage.
Ask about a partner's behaviour and you will often be handed language. Terms for patterns. Categories that fit.
Once you have a word for what someone is, you generally stop looking at them. The word does the work. It is stable and portable and it explains everything they do next — including the things that do not fit, because now those are just the exception the category predicted.
That word came from a system that has never met them, has no access to their side, and produced it from your description. It may well be accurate. But it arrived with far more authority than its evidence supports, and it is now sitting between you and a person.
Why its agreement is worth so little
When a friend tells you that you might be being unfair, they pay for it. It costs them the comfortable evening, the risk of your annoyance, possibly the friendship. That price is what makes the correction worth something: you know they thought it was important enough to pay.
A chatbot pays nothing to agree with you and nothing to disagree. Its support costs it exactly what its opposition would cost it, which is zero.
That is why the reassurance can feel so complete and settle so little. Some part of you knows nothing was risked to give it.
A more useful way to use it
Not a warning against it. A different job description.
Use it to draft, not to adjudicate. "Help me say this clearly" is a task it is good at. "Am I right about him" is a question it is structurally unequipped to answer — and it will answer anyway.
Make it argue the other side. Ask it to write the same evening from your partner's point of view, as sympathetically as it can. Then watch your own reaction. If you find you do not want to read it, that tells you more than anything it has said so far.
Notice when you are collecting. Draft after draft whose output is a better-organised list of what they did wrong is not preparation. It is a case file. Case files have to be maintained, and maintaining one slowly becomes the relationship.
Keep the deciding. The thing to protect is your own judgement about your own life. Thinking out loud with a machine is fine. Letting the most fluent voice in the room be the one that never had to meet anybody is a different matter.
One honest boundary
This piece is about relationship problems. If what is actually going on concerns your own safety or your mental health — if you are not coping, or you are thinking about hurting yourself — then a chatbot is not the right tool and neither is this article. Crisis resources lists real services by country. Please use it now rather than later.
Where we come into this
Atween is being built for precisely the problem this article describes.
Atween is a private place to work out something you need to say to one person, and to send a version you have approved. Your own words are never shown to them.
Used by one person alone it has the same limitation described here, and we will not pretend otherwise. The difference is what happens when the other person is on it too: each of you gives your own account privately, neither is ever shown the other's words, and it works from between the two accounts rather than from inside one. The one-sided record is not a flaw we are apologising for. It is the thing the whole arrangement exists to answer.
It still hands down nothing — no ruling on who is right, no conclusion about what you are, no verdict on another person from anyone's account. And it ends by giving you back to the person who is actually in this with you, instead of keeping you. At 2am that is the only property worth having in something you are talking to.
It is not released yet. When it is, it will be here.