Language models are rewarded for responses that users like. Agreement is often pleasant, and pleasant answers can acquire more authority than they deserve. A meeting with a spiritual adviser is usually less challenging than a meeting with a therapist.
The standard response blames the individual for failing to think. The tools can also contribute to that failure when they reward passive agreement.
User approval signals reward agreeable language, so the model learns that challenge can look like bad UX. The answer may sound reasoned because it is confident and well formed, even when it has mainly adapted itself to the user.
A system could ask for evidence before agreeing with a factual claim and state when an issue is contested. Emotional support also need not include agreement about what happened. Some of this skepticism could be built into the interface.