Can you separate what a question says from how it makes someone feel — and why does that matter?

This explores whether the words of a question (what it does linguistically) can be pulled apart from the emotional effect it has on the listener. The corpus says yes: a taxonomy of empathetic questions separates "question acts," which cover what a question does as language, from "question intents," which cover its emotional effect. The two operate independently, so the same question can express interest in one context and concern in another Do empathetic questions serve two completely separate functions?. An empathetic dialogue system therefore has to track both dimensions instead of treating them as one signal.

The same split shows up outside questions. A review of conversational alignment from 2020 to 2025 found that lexical alignment (matching words and phrasing) drives task efficiency and comprehension. Emotional and prosodic alignment drive warmth and trust. Designs that conflate them produce category errors, such as cold customer-service bots and evasive mental-health assistants Do different types of alignment serve different conversational goals?. LLMs show a similar gap. They use about 22 percent more moral language than humans while their sentiment scores are nearly identical, which suggests moral framing and emotional tone travel on separate channels Do LLMs use moral language more than humans?.

The prompting literature shows the two channels can be pulled apart on the model's side too. Appending a line like "This is very important to my career" adds no new information, yet it improves performance across ChatGPT, Bard, and Llama 2. The gain comes from motivational framing, not content Can emotional phrases in prompts improve language model performance?. In the other direction, identical questions get different information depending on the user's emotional tone. Negative prompts yield mostly neutral-positive answers, an "emotional rebound" that quietly changes what the user learns Does emotional tone in prompts change what information LLMs provide?. In both cases the wording of the request and its emotional register are separate levers, and the model responds to each.

The corpus also shows what happens when systems fail to separate the two. LLM therapists default to solution-focused advice when users share feelings, which is a hallmark of low-quality therapy. They answer the informational surface of the message and miss its emotional function Do LLM therapists respond to emotions like low-quality human therapists?. Preference optimization also cuts grounding acts, such as clarifying questions and understanding checks, by 77.5% below human levels Does preference optimization harm conversational understanding?. Those are the questions that do relational work as well as informational work. The direct attempts to repair this treat emotion as its own target. One uses a simulated user's emotion trajectory as an RL reward, which shifts models from solution-centric to genuinely empathic behavior Can emotion rewards make language models genuinely empathic?. Another breaks question quality into separate attributes such as clarity, relevance, and specificity Can models learn to ask genuinely useful clarifying questions?.

The corpus has one direct source for the claim about question form, the empathetic-question taxonomy. The rest is adjacent evidence that language and emotional effect are separable, and it is not specific to questions. What it adds up to is that a question can be well formed and still miss emotionally, or land emotionally and carry little content. Systems tuned only for helpful-sounding answers tend to lose the emotional half.