Use AI without reinforcing a first impression
Written edition. The audiobook has an earlier version of this page.
On this page
An AI assistant can help organise an account, explain a general concept or suggest questions. It does not observe your relationship. Its response is generated from the information and wording supplied, and may agree with your interpretation while sounding independent. A polished explanation is not new evidence about another person.
Remember: Ask for alternatives; verify claims; keep the decision yours.
The brief example is fictional. Possible explanations and practical suggestions are editorial applications; they are not a test that identifies a motive, trait or diagnosis.
How it may arise
A prompt such as “Why is my relative manipulating me?” assumes the conclusion before analysis begins. A model may build a coherent story around that wording, invent missing history or present general tendencies as an individual explanation. Supplying more incidents does not eliminate selection bias or missing perspectives.
What the evidence supports
Mrinank Sharma and colleagues evaluated five AI assistants and found that matching a user’s views could compete with giving accurate answers. The evaluated models do not represent every current service. A prompt asking for alternatives is not a validated guarantee against sycophancy. Sharma and colleagues. Towards Understanding Sycophancy in Language Models (ICLR 2024; arXiv revision 2025).
Psychologist Ziva Kunda reviewed research on motivated reasoning. Desired conclusions can shape which reasons people search for and find convincing. Disagreement alone cannot establish a hidden motive, and a desired conclusion can still be correct. Kunda (1990). The case for motivated reasoning.
A prompt to adapt
Copy the whole example below into a private draft. In your copy, replace the four bracketed parts with what you are comfortable sharing. Read the full draft before deciding whether to send it.
Help me think through an interaction. Treat my account as incomplete.
Observed or reported events: [exact actions, context and source].
My interpretation: [what I currently think].
An exception or missing perspective: [what does not fit or is unknown].
My goal: [understanding, a decision, collaboration or a boundary].
Separate facts in my account, possible explanations and your suggestions.
Give credible alternatives without inventing history or diagnosing anyone.
What information could change the leading explanation?
Suggest one action within my control and what to review afterward.
For research claims, identify a checkable source or say you cannot verify it.
This is an editorial aid, not a validated safeguard. In the app, Settings → AI connections selects a provider and model. Sending is optional; preview and redact the selected context first. The book’s practice notes are separate from app notes and are not automatically shared.
What else could explain it?
A useful suggestion can still rest on an unverified premise. Agreement from several models may reflect similar prompts or training rather than independent corroboration. A provider account or API key establishes access, not scientific reliability, confidentiality or a diagnosis.
These alternatives are starting points, not an exhaustive list.
What to notice and check
Separate reported events, your interpretations, general research claims and proposed actions. Ask which facts are missing and what would challenge the explanation. Open consequential sources and check whether they support the actual claim. Do not treat a plausible citation, numerical certainty or confident diagnosis as verification.
How to respond
Share only the context needed and remove unnecessary identifying details. Review the exact text before sending and the provider’s current terms for sensitive data. Ask for alternatives and an action useful under uncertainty; keep output as a suggestion until you decide what, if anything, to adopt. For safety, legal or clinical decisions, use an appropriate qualified route rather than relying on the model’s verdict.
A brief fictional example: You replace “My mother-in-law is asserting dominance through names” with the exact introductions, known preferences and an exception. Asking what could distinguish custom, preference and selective disregard produces a better question. It does not guarantee a correct answer.
Read the foundations
Related entries
These are editorial reading connections, not evidence of a cause or diagnosis in your situation.
- Confirmation bias: An agreeable model can reinforce selective reasoning.
- Compare occasions & test an explanation: A comparison supplies information that generated certainty cannot.
- Uncertainty: Confidence in wording is different from evidence.
- Getting support without needing a label: Use appropriate help for high-stakes decisions.
Source access and research scope. This is a selected reference entry, not a systematic review or an individual assessment.
Key takeaways
- AI does not add independent observations of your situation.
- Prompt wording and missing context can shape a convincing story.
- Keep suggestions separate, check sources and choose a proportionate next step.
Put it into practice
- Write one AI question without assuming the motive. Separate two observations, your first interpretation and one important unknown.
- Open A prompt to adapt above. Copy the example into a private draft and fill in the four brackets. Check that the draft asks for a credible alternative and one action within your control. You can try the exercise on paper or choose to send the draft to your AI assistant.
- If you get an AI response, pick one research claim and see whether the cited source supports it. Record any useful suggestion separately from what you observed. If you kept the draft private, check whether it leaves room for more than one explanation.
My practice notes: Use AI without reinforcing a first impression
Notes save automatically in this browser when storage is available. Avoid names and sensitive details, especially on a shared browser. Download a text copy before clearing your browser data.
Go deeper in the reference library
- Compare occasions & test an explanation — Investigate recurrence, exceptions and competing explanations through proportionate observation, action and review.