Cross-checking and cultural bias when there are three of us in the relationship

After an argument, Petr puts the last six messages from his partner into the chat. He asks whether she is manipulating him. His partner does the same but selects a different part of the conversation. Both then feel an impartial analysis has confirmed their view. They need not have any personal attachment to the chatbot. Giving it the role of referee and each supplying different material is enough.

A message may contain a threat or another specific problem that needs taking seriously. At the same time, its meaning is often shaped by circumstances, earlier agreements and what happened outside the text. Automated analysis does not replace knowledge of this context. ‘AI says you’re behaving toxically’ also lets someone hide their own stance behind supposedly objective authority. ‘This behaviour bothers me and I need to discuss it’ must be owned by the person saying it.

The second chat should check the conclusion, not condemn the partner again

The sycophancy described in the previous chapter here also affects someone absent from the conversation. The assistant accepts unexplained behaviour as long-term neglect, and further responses build on this interpretation. In work published in Science, involving eleven models and three preregistered experiments with 2,405 participants, sycophantic responses strengthened people’s conviction that they were right and reduced their stated willingness to repair a conflict.^[1] What mattered was how the response supported the participant’s own interpretation; the finding cannot be applied to every instance of agreement in an ordinary conversation.

If you want to check such an analysis, take a continuous section of your conversation and ask a new chat to review the original assistant’s work. You need to find where it moved from your words to its own conclusion. Another verdict about your partner would merely move you to a second, similarly limited assessment. The full prompt is in the practical appendix.

A new chat does not become an independent witness to what happened at home. You are checking the relationship between supplied evidence and conclusion. For a sensitive transcript, remove unnecessary names and other identifiers, avoid a public sharing link and keep enough context from both sides. Renaming everyone alone does not guarantee anonymity.

A useful finding might say: ‘The transcript contains one broken agreement, but no evidence of an intention to punish me.’ You can address the broken agreement even without knowing why the other person broke it.

Whose idea of a relationship are you reading?

Ordinary words can also hide a value judgement. ‘Healthy boundary’, ‘authenticity’ or ‘the right partner’ sound clear but do not say whose customs are used to assess the situation. Model responses may carry particular cultural assumptions, even when presenting them as obvious. Research on several GPT versions found their value responses more similar to those from some cultural environments than others.^[2] When reading particular advice, we can therefore ask: does the response present one idea of family, privacy or a good relationship as the only acceptable one?

Return such a sentence to your actual agreement. What did you promise each other? What bothers whom? Who can decide what? A useful response should help untangle these questions, not assign you someone else’s ideal life. A specific threat or coercion does not thereby lose seriousness; protection from violence cannot be bypassed by invoking different customs.

An apology you truly mean

AI can help you word a sincere message. But read carefully for anything added that you do not feel or cannot promise. ‘I understand everything you’ve been through’ sounds good. If you do not yet understand, ‘I want to understand better what was difficult for you’ is more accurate.

Before sending a sensitive message, read it without the surrounding chat. Does it reflect your stance? Could you explain the promise if the other person asked about it? Change what you would not say yourself. You can specify this in the request:

Help me say this: [my own content]. Preserve the meaning and suggest more natural wording. Do not add feelings, apologies, accusations or promises I did not state. Do not send the message to anyone.

For an assistant that acts, the prohibition on automatic sending must also match actual permissions. Check the account, recipient including copies, content, attachments and any further parameters. General trust in the assistant does not confirm a specific send.

When you have already sent a message based on faulty analysis

Give yourself time to establish what you actually communicated. There is no need to have AI write a second, even more forceful defence. If contacting the other person is safe, you can correct the unsupported conclusion: ‘I attributed an intention to you that I don’t know. That wasn’t fair. I want to talk about the specific situation that bothered me.’

Using AI does not remove your responsibility for the text you sent. At the same time, an error can be repaired step by step, without requiring immediate forgiveness from the other person. In a violent or controlling relationship, do not address safety with a universal apology or confrontation. Seek support outside it.

Someone else’s privacy is not just more context

A partner’s message may matter in understanding the situation. That does not mean you need to upload the entire archive, their health records and photographs. Your own summary often suffices to formulate a response. Passing a sensitive transcript to another service creates another processing location. Keeping your own copy on a device and uploading it to someone else’s application are different decisions.

Particularly under coercion, it may not be safe to begin a discussion with your partner about all your conversations. Phone monitoring, demands for passwords or threats of punishment for seeking support need a different framework from an ordinary agreement about screen time. Seek safe support outside the controlling relationship. Address immediate danger according to the first page.