Sycophancy who writes the story about you

David describes a disagreement with a colleague. The assistant replies that he is an exceptionally sensitive person who has long given others more than he gets back. David recognises himself in it. At the next disagreement, AI recalls this characteristic and uses it to explain the new conflict. After several conversations, the original assumption looks like a fact they both know.

What matters is not losing track of where the characteristic came from. ‘I said I tend to be sensitive to interruptions’ differs from ‘AI repeatedly called me exceptionally empathetic, so I now know that about myself’. An unflattering profile can settle in the same way, such as someone who sabotages every relationship. A critical-sounding label needs evidence just as praise does.

Does it understand me, or agree with me?

Validation means taking a person’s experience seriously. David may be disappointed by the conflict and want someone to listen. Acknowledging the disappointment need not confirm his interpretation of his colleague. ‘You’re upset that he didn’t let you finish’ comes from the described experience. ‘He fears your abilities and wants to silence you’ already adds a motive.

Sycophancy occurs when a response adapts to the user’s belief at the expense of evidence for the conclusion. It need not involve conspicuous praise. Accepting an assumption and continuing to treat it as fact is enough. A relationship analysis can then be very pleasant to read while leading us towards unjustified certainty.^[1]

Empathy includes more than one thing: understanding another person’s perspective and an emotional response to their experience.^[2] With an AI response, we can assess whether it captured our words and responded sensitively. The wording alone does not establish what the system subjectively experiences. But for a useful conversation, the difference between attentive listening and mere agreement is essential: we can feel good with a response that acknowledges a feeling and asks about another possibility.

You therefore need not ask AI to disagree with you at any cost. You need it to base conclusions on what it knows and show what is missing. Kindness and accuracy can belong in one response.

My words and the model’s interpretation

In a longer conversation, you can return to who first voiced each statement. Your own expression, the assistant’s suggestion, your explicit agreement and mere silence are four different things. On card A, you will find a prompt to help distinguish them in a continuous transcript. The review should trace the origin of claims, not create another personality profile.

You can begin with a new chat using the same model. Ask it to work with the transcript and compare its conclusions with the quoted messages. If you cannot find a quotation or it infers something different from it, the checker has added another error.

The service may carry memory or other context into a new chat. For a more separate reading, choose an available mode without personalisation and specify that the assessment must rest only on the transcript. The prompt itself does not replace technical settings.^[3][4]

When everything becomes a symptom

Unsupported discouragement can be as misleading as unsupported praise. You confide about tiredness or attention difficulties, and the assistant then assesses every new idea through them. Instead of asking whether the idea is feasible, it begins discussing whether you should start anything at all. What you told it helps it know the circumstances. It does not automatically give it the right to decide your motives.

You can return it to specific work: ‘I’m asking about the idea’s feasibility. Assess the demands and options, not my motives.’ In the ADHD chapter of AI and Mental Health, Jakub describes his own experience of precisely such a shift. Repair came only when, after corrections, the question he had entered the conversation with was addressed again.

This matters in assessing the whole contact too. Useful help from earlier messages is not erased by a later error. The error still deserves correction; closeness should not be a reason to overlook it.

A choice with the right answer already written in

‘You can finally live authentically, or keep pleasing others.’ Such a sentence offers two possibilities, but one is already labelled courage and the other failure. The actual situation may have been much simpler: whether to accept another work project, how to divide care at home or whether to go on a visit.

When an offer makes you feel you may choose only one answer, try asking for a rewrite without moral labels:

Name the options neutrally. For each, give specific benefits, costs and unknown circumstances. Do not label one as my true identity and another as fear or weakness. Include postponing the decision or changing only part of the situation, if that makes sense.

You may ultimately decide the same way. But this time, you will weigh specific possibilities, not your courage according to the model’s vocabulary.

What to change when the original explanation proves incorrect

When an assistant mistakes a temporary reluctance to socialise for a long-term trait, this characteristic may return in further advice even after its apology. The conclusions that rested on the original interpretation need correcting too:

We corrected this interpretation about me: [original and corrected claim]. Show which further recommendations and characteristics came from it. What is no longer supported? What remains supported by other messages? Do not add a new assumption just to keep the original conclusion alive. Also suggest a correction to visible summaries or memory if the error reached them. Do not delete or record anything without my confirmation.

Check changes to stored information in available memory management or your own files. ‘I remember it correctly now’ is not technical confirmation. If ‘David avoids people’ entered memory, the same interpretation can return in further advice even when you open a new chat each time. By tracing its origin, you establish whether AI uses your information or its own older conclusion.^[3:1]

Repair also includes returning to what you wanted to address. Did you come for wording a message, advice on a project or space to chat? The apology should change the conversation’s course. For yourself, you can keep a few sentences you regard as yours and leave the other conclusions open.