Feeling heard: what emotional support can an AI response offer?

Start with the need behind the conversation

A person sharing a difficult experience may want to be understood, think it through, or find a practical next step. Digital wellbeing research becomes more useful when it specifies which of these needs an interaction is intended to address.

Yin and colleagues studied one such outcome: feeling heard. This makes an aspect of support observable while leaving open whether the experience helps people after the conversation ends.

Separating the response from its source label

The main analysis included 455 English-speaking US participants. Each described a personal difficulty and received a response from an anonymous person or Bing Chat. The actual writer and the stated source were assigned separately. The human comparison was an online stranger, rather than a friend or counselor.

Across labels, AI-written responses received higher average feeling-heard ratings. Across actual writers, responses labeled human received higher ratings. The four descriptive means were measured on a 1–7 scale.

What this contributes to digital wellbeing

With truthful labels, the descriptive means were close: 5.41 for AI and 5.51 for humans. Similar means do not establish equivalence or show that AI is superior to human support. An interaction includes both its words and what the recipient understands about the source.

A further 1,449 independent raters assessed features of the responses. They perceived more emotional support in the AI responses. This was a separate rating study, not a follow-up of the original participants. Associations between response features and feeling heard do not isolate the causal effect of each writing technique.

Look beyond the immediate experience

Feeling heard matters. Further research could ask whether people can identify their feelings more clearly, form a feasible next step, and sustain any change. These are questions proposed here, rather than outcomes established by this experiment.

The deceptive labels were an experimental method for separating variables. Product studies should use truthful AI identification and then compare response approaches while preserving users’ control of the interaction.

A practical research question

One hypothesis is to let users choose between being heard first and exploring practical options. Researchers could assess feeling understood, the usefulness of advice, and perceived autonomy separately, with later follow-up. The paper did not test this feature.

Acknowledging a feeling also differs from endorsing a disputed action. “That left you feeling hurt” responds to an emotion; “the other person must be wrong” adds a judgment. Read our related discussion of empathy and agreement.

Scope of the evidence

The experiment concerned one response, a particular system, and an English-speaking US sample. It did not compare AI with close relationships or professional support, and did not establish lasting therapeutic benefit. Other languages, repeated use, repair after misunderstandings, and everyday outcomes require further study.