In STEM/objective matters (math, science, coding), answers are more clearly defined as either right or wrong. This is where hallucination is more difficult/unlikely.
But in personal matters, everything is subjective. AI tends to default to the middle of the spectrum, i.e., general advice. If we want to safeguard against affirmation, we should force AI to challenge us more often by increasing its rate of clarifying questions, counter-considerations, and uncertainty considerations. One implementation idea: run a classifier over the conversation, detect when it's about interpersonal advice, then prepend a hidden instruction template to the model prompt.
But in personal matters, everything is subjective. AI tends to default to the middle of the spectrum, i.e., general advice. If we want to safeguard against affirmation, we should force AI to challenge us more often by increasing its rate of clarifying questions, counter-considerations, and uncertainty considerations. One implementation idea: run a classifier over the conversation, detect when it's about interpersonal advice, then prepend a hidden instruction template to the model prompt.