Analysis, not certainty: Dialogatlas separates sources, observations, and editorial conclusions. New evidence may change this assessment.

Reading and real interaction were examined separately

Study 1 used a pre-scripted conversation. This allowed 158 participants to evaluate the same content. Study 2 let 88 people interact with a chatbot themselves, thus coming closer to actual use.

The combination is methodologically useful. A fixed script facilitates controlled comparison; an interaction additionally captures how people bring their own questions and reactions into the conversation.

The focus was advice on a sensitive personal problem. In such situations, expectations often go beyond correct information. Tone and perceived attitude influence whether a response is accepted at all.

The study measures perceptions of the service and the chatbot. It thus demonstrates communicative preferences, not automatically medical benefit or long-term psychological effects.

Three forms of emotional response were distinguished

Sympathy refers to a form of compassion or concern. Cognitive empathy is more focused on understanding and grasping the other person's perspective. Affective empathy emphasizes emotional resonance.

This distinction is important for chatbots because different phrasings create different expectations. A system can show that it understands the situation described without claiming to feel the same way itself.

Cognitive empathy can often be formulated closer to the content: "You want to start telling your story and don't want a solution yet." This response shows attentiveness to the user's intention.

A strongly affective formulation, on the other hand, can quickly simulate human inner experience. Whether it is appropriate depends on context, labeling, and expectations.

Emotionless counseling performed worse

The results support the assumption that people treat computers socially. Expressions of sympathy and empathy were preferred over purely emotionless counseling.

This does not mean that every response must begin with an acknowledgment of feelings. Rather, it shows that purely instrumental information on a sensitive topic can omit an essential layer of the conversation.

A brief acknowledgment creates a transition between personal disclosure and factual information. Without this transition, even correct content can come across as cold or dismissive.

The challenge is to keep the acknowledgment specific. Generic formulas such as “That must be incredibly hard for you” can become inappropriate if the person themselves do not assess their situation that way.

Skeptical people reacted particularly positively

Emotional expression was especially significant for individuals who initially doubted that machines possess social cognitive abilities. A well-designed dialogue can thus partially alter negative prior expectations.

This effect should not be bought with deception. The goal is not to make people forget that they are talking to an AI. Rather, the system can communicate respectfully and responsively despite clear labeling.

Transparency and social quality are not opposites. Visible AI labeling can be combined with natural language, appropriate pacing, and genuine responsiveness to corrections.

Trust should develop from observable behavior. If a system appears attentive but later repeatedly violates the same boundary, the positive first impression collapses.

Emotion changes the comparison with humans

Tsai, Lun, and Carcioppolo studied 142 people in communication about HPV vaccinations. Depending on the condition, shame, anger, or a neutral state was induced; the counterpart was presented as a chatbot or a human representative.

In the case of anger, satisfaction with the chatbot was lower than with the human. In the case of shame, participants were more likely to express concerns to the human representative and responded in greater detail.

In terms of perceived usefulness and intention to follow recommendations, the chatbot was overall comparable. The pattern shows that emotional communication cannot be described by a universal function.

A system can be sufficiently useful for one task and yet, in certain emotional states, leave less room for openness. These are precisely the situations that should be tested separately.

Expert information must still be independently verified

Xie and colleagues had ChatGPT answer nine common questions about rhinoplasty. Specialist physicians assessed accessibility, informational content, and accuracy.

The answers were coherent and understandable but offered only limited detail and personalization. The example again separates two quality dimensions: expert information and social fit.

Empathetic language must not mask a factual shortcoming. Conversely, a factually sound text should not be regarded as complete advice if it does not know the individual situation.

A responsible chatbot therefore requires separate checks. Sources and expert reviews substantiate claims; dialogue tests verify that the content is offered at the right moment and in the desired form.

Naturalness arises from responsiveness, not from slang

Product teams often try to fix the artificial tone with colloquial language, emojis, or short sentences. Such devices can fit the chosen style, but on their own they do not create social attentiveness.

What matters more is contingency: the response must follow precisely this message and show what has changed compared with the previous turn. Repeated stock phrases feel mechanical even with slang.

A natural chat may be imperfect and brief. It should not, however, invent details to seem more personal, nor add emotional intensity that the user has not expressed.

Settings such as normal, direct, casual, or calm can influence tone and pace. They must not override the core conversational function or the boundaries the user has stated.

The right dose is a question of the conversation trajectory

The experiments show that people may want more than emotionless information in sensitive health advice. They suggest deliberately shaping social signals rather than removing them entirely for fear of artificiality.

How these signals affect long conversations, however, must be examined separately. Repeated affirmation can be tiring, avoid contradiction, and press a conversation into a therapeutic template.

A good course of the conversation flexibly alternates between acknowledging, asking questions, informing, and remaining silent. The user should be able to help decide whether they want to narrate, understand, or hear a perspective at any given moment.

People do not want only factual information from chatbots. But they also do not want a machine that performs sympathy. The system becomes credible where emotional language is concrete, limited, and confirmed in subsequent behavior.

Sources & further reading