Anger and shame change how people talk to chatbots

The experiment separated the counterpart and the emotional state

The study used a 2-by-3 design. Participants were randomly assigned to a purported chatbot or a human brand representative, as well as to one of three emotional conditions: shame, anger, or neutral.

Self-reports and observable behavior were examined in the context of health marketing for HPV vaccination. This allowed the team not only to ask how useful the conversation seemed but also to consider how extensively people responded and what they disclosed.

The perceived identity is crucial here. The response to a counterpart understood as human or technical was assessed. Such attributions can alter behavior even when the linguistic surface is similar.

Under anger, the chatbot lost satisfaction

When anger was triggered, participants reported lower satisfaction with the interaction with the chatbot than with the human representative. The emotion thus influenced the relative strength of the respective counterpart.

Anger can demand a response that goes beyond merely addressing the factual content. Timing, acknowledgment of the cause, and the ability to withstand contradiction shape whether the response seems appropriate.

A standardized calming sentence can backfire in this context. Those who are legitimately angry may experience premature de-escalation as defensiveness or downplaying.

For chat design, it follows that anger should not automatically be treated as a problem to be eliminated. The system should first understand what the anger is directed at and what conversational function the user currently expects.

In the case of shame, more was entrusted to the human

In the shame condition, participants were more likely to disclose concerns about HPV risks to the human representative and gave more detailed answers. This result contradicts the widespread expectation that a non-judgmental technical counterpart automatically leads to greater openness.

Anonymity can lower inhibitions, but human resonance can also create a sense of safety. Which effect prevails depends on the topic, the person, the design, and the expectation.

A chatbot should therefore not equate discretion with relational safety. A private channel can facilitate entry, but it must demonstrate in the conversation that it does not handle the expressed concern in a schematic way.

Data protection also influences openness. People can only make meaningful decisions about what they share if it is clearly explained how content is stored, processed, and, where applicable, evaluated.

In terms of general usefulness, the two were comparable

Across emotional conditions, the chatbot performed similarly to the human representative in terms of perceived usefulness and influence on intention to comply.

This is a relevant positive result. It shows that technical conversational partners can take on an objectively effective role in clearly defined health communication.

At the same time, an intention should not be equated with actual subsequent behavior. Whether someone receives a vaccination or makes an informed decision in the long term was not substantiated by the measures described in the abstract.

The comparable average performance also masks the differences under anger and shame. Product decisions should therefore not only consider an overall value, but also analyze situations in which the system is experienced differently from a human.

Empathetic phrasing helps, but does not solve everything

Liu and Sundar showed in two experiments that sympathy as well as cognitive and affective empathy were preferred over purely emotionless health advice. Particularly skeptical individuals responded positively to such signals.

For vaccination communication, this means that a chatbot does not have to sound purely technical. It can acknowledge uncertainty or embarrassment without dramatizing emotions or feigning human experience.

The appropriate form, however, depends on the emotion. A sympathetic response to shame is not automatically a suitable response to anger. A universal empathy phrase can actually prevent the necessary differentiation.

The solution is not elaborate acting but context-sensitive restraint: briefly acknowledge, stay with the concrete content, and leave the next direction to the user.

Providing understandable information remains a distinct achievement

The rhinoplasty study by Xie and colleagues demonstrates a different form of initial health consultation. Nine responses were rated by experienced plastic surgeons as coherent, understandable, and informative.

At the same time, details and genuine personalization were lacking. This reminds us that good general information and emotionally appropriate communication are separate skills.

A system can correctly explain a medical fact and still miss the moment of the conversation. Conversely, it can respond warmly but say too little or something incorrect from a professional standpoint.

Evaluation should examine both: the quality of the information and the fit of the interaction. A single overall score can make it hard to see which level needs improvement.

Conversation modes can make expectations explicit

People do not always come with the same intention. Some want to voice a concern, others want to understand facts or prepare a decision. A brief choice before the conversation can make this expectation more visible.

The selected mode, however, must not remain rigid. Anger can turn into curiosity, shame into a concrete question. The system must recognize a shift or make it easy to switch.

Responding to direct feedback is especially important. If someone says they do not want advice right now, the chat should not just agree and then formulate a new recommendation in the next sentence.

The emotional state is thus one contextual signal among several. Conversation goal, content, trajectory, and explicit corrections should be weighted more heavily than an uncertain automatic emotion attribution.

Relative strength depends on the situation

The experiment shows neither the general superiority of humans nor the equivalence of chatbots in all health communication. It shows that both have different interaction consequences depending on the emotion.

For products, precisely this differentiation is valuable. A task can work well on average and be systematically weaker in certain emotional situations. Such patterns belong in tests and product boundaries.

Subsequent experiments would need to include real decisions, longer conversations, different population groups, and current language models. A laboratory experiment from 2021 is a solid starting point, but not a finished blueprint.

Anger and shame change how people talk with chatbots. Those developing supportive systems should therefore not only optimize responses, but also understand the different reasons why people open up, withdraw, or contradict.