AI guidance
What it does, and what it will not do
What it is good at
- Context. It has read what you told it before, so you are not starting from scratch every time you are upset.
- Untangling. Separating what happened from what you took it to mean — the two get fused when you are hurt, and pulling them apart is most of the work.
- Finding the words. Turning “you never listen” into something specific enough to respond to.
- Slowing things down. One question at a time, especially when things are escalating.
- Noticing repetition. The same argument in different clothes, three weeks apart.
- Preparing. Rehearsing a conversation you have been avoiding.
What it will not do
- Tell you that you are right. That is not a service, it is flattery.
- Diagnose anyone. Not you, and definitely not someone it has never heard from.
- Tell you what they are thinking. It does not know, and guessing does real damage.
- Promise an outcome. No software can tell you whether to stay.
- Replace a therapist. If you need clinical support, we will say so.
- Carry a secret into a shared room. Your private notes are not loaded into a joint session at all.
Where it does take a side
Neutrality is the default, and it has a hard limit. OnBetterTerms will not treat violence, threats, sexual coercion, stalking, or controlling someone's money, movements or contacts as one side of a reasonable disagreement. It will not suggest compromise or patience as a response to being harmed. In those situations your safety stops being one factor among several.
You are in charge of how it talks to you
Rate any response — helpful, not helpful, or with a reason like “too harsh”, “too long”, or “misunderstood me”. It adjusts: how direct, how detailed, how many questions, whether you want examples or a plan or just to be heard first. That preference is private. The other person is never told how you like to be spoken to.
Before
You rated this: too long
After
Ratings, then outcomes
Ratings say what you liked. Outcomes say what helped.
Rate a reply and the next one changes — length, directness, whether it asks questions or hands you options. That is the easy half.
The half that matters comes later. After something has played out you answer a few short questions on your own: did it feel resolved, did you feel heard, what do you want to remember. An approach can be gratefully received and still leave the argument exactly where it was, and only the debrief knows the difference.
It adapts to you rather than being retrained by you — nothing you write changes the underlying model. Your preferences and outcomes are yours, and invisible to anyone you share a space with. What carries between sessions is how you like to be spoken to and how things actually went — not a record of the conversations themselves, which are not kept.
What you brought, and what you have written since.
Suggestions shaped to how you take them in.
The conversation you actually had.
Whether it resolved, and whether you felt heard.
More of what worked for you, less of what did not.
Honesty about the model
Responses are generated by a large language model. It can be wrong, and it can be confidently wrong. Longitudinal observations are correlations in your own self-reported data, not causes, and we word them that way deliberately. Where an observation is based on thin data, it says so instead of reaching.
The formal version is the AI disclosure, and the companies that run the models are named on the subprocessors page.