AI Empathy Research: Qualitative Analysis Guide

According to Psychological Science on July 21, 2026, language models can generate responses that users often rate as empathic, but researchers warn the apparent empathy is a statistical synthesis of human language rather than felt emotion.

• July 21, 2026; Psychological Science reports nearly a quarter of American adults and 13% of adolescents have used AI for socioemotional support.
• 2026; Desmond C. Ong's review in Current Directions in Psychological Science (in press) summarizes multiple studies where AI responses were rated more empathic than human responses.
• 2026; Mark Thornton argues in Perspectives on Psychological Science that LLM empathy reflects a 'wisdom of the crowd' effect rather than true feeling.
• 2026; Studies show an "AI penalty": when participants learn a response came from AI, ratings of empathy fall even if wording stays identical.

Answer: Researchers compared AI-generated support messages to human-written messages and measured perceived empathy using participant ratings and disclosure experiments.

According to Psychological Science on July 21, 2026, multiple studies presented anonymous messages to participants, collected empathy and support ratings, then revealed the source to measure the so-called AI penalty.

According to the Ong review in Current Directions in Psychological Science (in press), experimental methods included blind message evaluation, user self-report of comfort, and scenario vignettes where participants rated compassion, validation, and support.

Constraints: the cited work uses perception measures rather than physiological or longitudinal social-outcome metrics, so claims about long-term relational effects remain speculative.

Answer: Qualitative researchers should unpack "empathy" into precise constructs and UX teams should design features that augment rather than replace human relationships.

According to Psychological Science on July 21, 2026, experts recommend replacing the single label "empathy" with lower-order constructs such as perspective taking, vicarious affect, and compassionate concern before coding or designing interventions.

Practical steps for researchers: code transcripts for specific empathy constructs (for example, reflection, validation, compassion), compare AI and human utterances by construct frequency, and run segment analyses by age (note adolescents = 13% AI use as of July 21, 2026).

Practical steps for UX teams: instrument prototypes to detect when AI responses are purely affirmative versus when they prompt social action, and test whether AI coaching increases user outreach to real people over 30, 60, and 90 days.

Answer: AI empathy research shows LLMs can simulate aspects of empathic language, but qualitative analysis is required to unpack which constructs are present and which relational functions are missing.

According to Psychological Science on July 21, 2026, researchers recommend precise construct definitions, blind-versus-revealed tests, and attention to long-term relational outcomes.

Next steps: use construct-level codebooks, combine anonymized rating experiments with longitudinal user tracking, and instrument prototypes to detect whether AI increases human outreach.

If you want to scale those analyses, Try Evidano for free to import transcripts, build hierarchical empathy codebooks, and run cross-segment analyses.